A system and method for adverse event analysis based on a sop system

Abstract

The application provides a kind of adverse event analysis system and method based on SOP system, related to production quality and safety control related technical field, the method presets standard operation procedure basic data matrix;Online analysis process monitoring data, solve adaptive dynamic boundary threshold and comprehensive degradation index execution trigger;On the report text executes mutual information condensation, generates unique event identification primary key;Cross heterogeneous bus call data and execute physical characteristics intelligent interpolation, assemble high-fidelity multi-modal data set;Map it to the manifold space to calculate the feature distance, recall SOP clauses in reverse and build a four-dimensional causal atlas to determine the root cause;After isolated sandbox backtracking verification safety convergence, update the benchmark.The application eliminates the slow change hidden danger caused by static threshold, breaks the subjective report and the multi-modal fault of physical time sequence, realizes the high safety closed loop self-evolution of control procedure.

Description

Technical Field

[0001] This invention relates to the technical field of production quality and safety control, and in particular to an adverse event analysis system and method based on a Standard Operating Procedure (SOP) system. Background Technology

[0002] In high-risk, long-process production environments such as process manufacturing, pharmaceuticals, or power, Standard Operating Procedures (SOPs) are control procedures that ensure consistency in process operations and reduce safety and quality risks. However, conventional adverse event analysis systems often face industry-specific pain points in practical engineering applications, such as delayed response to sudden hazards and underreporting of critical risks. When production anomalies deviate from SOP requirements on the production floor, if the system requires front-line operators to cross-system verification and manually fill in complex equipment fault codes, process section numbers, and structured information of all fields related to procedures, it usually leads to two limitations: First, on-site operators delay the initial reporting time of the system in order to prioritize safety or quality handling; second, due to the limitation of the system's full-field lockout, operators are prone to subjectively "filling in the numbers based on experience," directly causing serious distortion of the underlying data of the anomaly event in the database, which in turn misleads the direction of subsequent root cause investigation and procedure adjustment.

[0003] To improve the accuracy and automation level of production line anomaly identification, existing technologies have proposed a rolling adjustment scheme for production line simulation based on digital twins, such as the rolling optimization system and method for production line simulation disclosed in Chinese Patent Publication No. CN113361139B. This scheme periodically extracts recent production data and real-time event signals from the workshop manufacturing execution system (MES) and equipment monitoring and management system through a data integration interface. It then uses mathematical distribution models such as normal distribution, Ellang distribution, and binomial distribution to fit and reconstruct the feature values ​​of the process cycle analysis model, equipment fault analysis model, and production quality analysis model, thereby updating the digital twin model to correct simulation errors.

[0004] However, this type of system has the following shortcomings when solving the aforementioned industry problems: First, the risk triggering mechanism is simplistic and mechanical. Its iterative evolution triggering relies heavily on a single static deviation limit between the actual monitored value and the macroscopic simulation model, or on explicit sudden equipment shutdown anomalies, such as a deviation greater than 5% set in D1. Due to the lack of cumulative integral characteristics of the time-series deviation magnitude and weighted input of the rigidity of process topology constraints, the system cannot detect micro-critical risks that have not reached the static limit but have shown a continuous deterioration trend over multiple consecutive sliding statistical periods. As a result, these potential hazards are directly filtered out and missed by the system before they reach the limit deviation.

[0005] Second, there is a lack of mechanisms for cross-system integration and seamless completion verification of micro-transient data. Data integration in D1-type systems only performs distribution fitting and reconstruction based on statistical averages over macro-historical time periods, lacking an automatic extraction mechanism for micro-transient underlying physical time-series data streams that adaptively extend across subsystems and time axes for specific sudden event identifiers. This results in the system being unable to adapt to the progressive information processing requirements of first providing extremely simple responses at the second level and then asynchronously supplementing them with structured data. It also cannot automatically and intelligently interpolate and perform strong consistency static verification on manually entered fields using primary and foreign key error-proofing constraints in relational databases. Furthermore, it cannot directly map, feed back, and replace the root cause analysis results into specific SOP text control clauses in the basic database. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a defective event analysis system and method based on the SOP system, which can realize cross-system transient automatic and accurate mapping and cascaded error prevention consistency verification of heterogeneous production characteristic data, thereby constructing a convergent standard operating procedure control clause-level adaptive cyclic self-evolution driven by online real-sequence fact snapshots.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides an adverse event analysis system and method based on a SOP system, comprising the following steps: S1: Control clauses in accordance with standard operating procedure text Establishing an initial feature set for exogenous text modalities And construct a mapping table of equipment process thresholds. ; S2: Obtain physical sensor parameters to construct process monitoring data vectors Based on process monitoring data vector Latch-triggered transient event dataset ; S3: Based on the initial manual qualitative text Initial feature set of exogenous text modal Generate a unique event identifier primary key The level signal is locked by the interface hardware. Block and output a two-level structured reporting dataset ; S4: Use a unique event identifier for the primary key Retrieve heterogeneous time series raw data And reverse-inject a two-level structured reporting dataset. Generate a high-fidelity multimodal event dataset ; S5: Based on a high-fidelity multimodal event dataset Generating the temporal manifold feature matrix The temporal manifold feature matrix Initial feature set of exogenous text modal Mapping to a set of recall reverse semantic recall control clauses And generate a structured lesion mapping atlas. Determine the final root cause determination matrix. ; S6: Analyze the final root cause determination result matrix isolated sandbox memory area When the backtracking test converges safely, update the static initial baseline constraint vector. .

[0008] In the preferred embodiment, step S1 includes the following sub-steps: S11: Extract the standard operating procedures, quality control requirements, and equipment operation standards for each production process to establish a static initial benchmark constraint vector. For quality-related processes, a minimum product qualification rate is defined; for equipment-related processes, a baseline value for single equipment downtime is defined; and each static initial baseline is defined as a vector. Corresponding process code Perform unique memory address binding; S12: Compile a structured list of all production equipment required for each production process and create a list of related equipment interactions. In the list of interactions with associated devices Record the equipment model of each production machine , To which process node Equipment Management Identification and equipment operating parameter acquisition port information ; List of interactions with associated devices Each piece of production equipment is associated with a corresponding process code. The static initial reference constraint vector bound to the corresponding process. Generate a device process threshold mapping table by performing three-dimensional primary and foreign key association binding. Map the equipment process thresholds to the table. Persistently store the data in the business relationship database; S13: Quantify the impact weight of each production process on the final product quality and the constraint stiffness on the advancement of subsequent processes; set process priority coefficients. The process priority coefficient The process is divided into three discrete score levels, and the priority coefficient of each process is assigned accordingly. Corresponding process code Perform mapping; S14: Retrieve the historical sequence of long-cycle physical parameters under continuous, error-free operation conditions through the data integration interface of the workshop manufacturing execution system. Calculate the historical sequence of long-period physical parameters Long-period mean vector With long-period variance matrix Combining long-period mean vectors With long-period variance matrix Generate long-period data feature matrix This is to initialize the data distribution for the adaptive dynamic boundary triggering mechanism. S15: Read all standard operating procedure text control clauses stored in the business relationship database. Using a pre-trained natural language processing backbone network to control standard operating procedure text clauses Word segmentation, stop word removal, and hidden layer mapping are performed to extract semantic features and generate an initial feature set of exogenous text modalities stored in the manifold space. This completes the initialization of the exogenous semantic base for the cross-modal semantic mapping mechanism.

[0009] In the preferred embodiment, step S2 includes the following sub-steps: S21: The time-series parameters of the physical sensors of the production equipment are retrieved in real time through the workshop data acquisition layer. The multi-dimensional physical sensor parameters are then spatiotemporally aligned and denoised and normalized to assemble and generate the process monitoring data vector for the current time step. ; S22: Read the dynamic timing sliding window in the running memory and combine it with the long-period data feature matrix. An unsupervised prediction model is established, and the expected prediction vector at the current time step is calculated using the state evolution probability density function. ; S23: Calculate process monitoring data vector With the predicted expected vector The transient deviation residuals are used to construct the deviation residual matrix by performing an outer product operation on the transient deviation residuals and their transposes. ; S24: Extract the deviation residual matrix Logarithmic integration is performed on the diagonal elements within the dynamic time-series sliding window to generate a degradation trend negative feedback damping factor, which then constrains the static initial benchmark vector. Combined with the negative feedback damping factor of the degradation trend, a preset adaptive boundary adjustment coefficient is used. Execute Hadama The adaptive dynamic boundary threshold vector is calculated. ; S25: Extract the deviation residual matrix The trace term is used to calculate the adaptive dynamic boundary threshold vector. Relative static initial reference bound vector The L2 distance is used to apply the hyperbolic tangent function to the process priority coefficient. Perform a nonlinear continuous mapping; combine the trace term, the L2 distance, and the mapped process priority coefficients. Perform multidimensional mixed linear weighting to calculate and output a dimensionless comprehensive degradation index. ; S26: Overall Deterioration Index Perform a step-condition check control chain audit when the overall degradation index is... When the time span exceeding the safety threshold reaches the time limit of the hardware anti-jitter filter, a hardware interrupt command is sent to the physical bus actuator to toggle the hardware trigger level, blocking real-time bus data and vectorizing the process monitoring data. Sequence, Deviation Residual Matrix Snapshot and Adaptive Dynamic Boundary Threshold Vector The snapshot is packaged and latched to generate a dataset that triggers transient events. .

[0010] In the preferred embodiment, step S3 includes the following sub-steps: S31: Monitor the dataset of triggered transient events Write status, activate the simplified initial reporting process, and capture the event timestamp. Physical location coordinates of the event Receive initial manual qualitative text Temporarily stored in the application layer memory queue; S32: Call the initial feature set of the exogenous text modality Compared with the initial human qualitative text Perform cross-modal semantic space projection to extract initial manually defined text. High-density core text features The first non-linear dependency between them is used to calculate high-density core text features. Initial feature set of exogenous text modal The second nonlinear dependency between them and the introduction of semantically preserving Lagrange multipliers Perform scaling; subtract the scaled second nonlinear dependency from the first nonlinear dependency to construct the mutual information minimization objective function. The objective function is optimized by backpropagation to minimize mutual information. To achieve convergence, subjective and invalid expressions are stripped away, and high-density core text features are output. ; S33: Call the physical location coordinates of the event Equipment process threshold mapping relationship table Initiate a spatial topology addressing query and retrieve the corresponding process code in reverse. The process code is encoded using a string hash concatenation algorithm. Event timestamp Physically concatenate with a secure random sequence code to generate a unique event identifier primary key with embedded spatiotemporal process constraints. ; S34: Instantiate a dual-loop sequential state machine Towards a dual-closed-loop sequential state machine Inject transition instructions to force a dual-loop sequential state machine. The running state transitions from idle state to emergency suspension state. A high-precision timer is triggered by the rising edge signal of the transition synchronization to start the supplementary timer. ; S35: Receive the operator-selected device identifier Operator selected procedure number Extract the unique event identifier primary key Included process codes Using first-paradigm filtering conditions, the list of interactions with associated devices is used. Perform dynamic view clipping to generate a subset of cascading constraint devices. ;Execute Boolean logic comparison instructions to determine the device identifier selected by the operator. Does it belong to a subset of cascaded constraint devices? When the internal element of the function is determined to be logically false, it outputs a hardware latch-up signal to the front-end I / O controller. Cut off the underlying touch event response and forcibly block the path for dirty data to enter the database; S36: In a dual-closed-loop sequential state machine Determine if the Boolean logic comparison instruction is logically true and add a timer. If the overflow has not yet reached zero, the primary key will be uniquely identified by the event. High-density core text features Operator selects equipment identifier Operator selects procedure number and the dataset that triggers transient events Perform deep packet encapsulation to generate a two-level structured reporting dataset. .

[0011] In the preferred embodiment, step S4 includes the following sub-steps: S41: Extract the two-dimensional structured reporting dataset Encapsulated unique event identifier primary key For unique event identifier primary key Perform reverse hashing of the string to extract the process code. Timestamp of the event Read process code The corresponding standard production cycle parameters are based on the event occurrence timestamp. Extending bidirectionally from the origin to both the historical and future time domains, an adaptively aligned data retrieval time window is constructed. ; S42: Joint Process Code With the operator selecting the device identifier Perform primary-foreign key cascading addressing in a relational database, utilizing process coding. In the equipment process threshold mapping table Initiate an inner join query to retrieve the list of interactions between related devices. Defined physical sensor measurement points and production execution management control points; compare the operator-selected equipment identifiers. By matching the physical sensor measurement points, the register address of the underlying programmable logic controller and the production batch number of the manufacturing execution system are extracted, along with the data retrieval time window. Perform formatted packaging to generate a list of data retrieval requirements for heterogeneous systems. ; S43: Activate the protocol conversion socket and use the mapped hash table structure to retrieve the data retrieval request list from heterogeneous systems. The included qualitative fields are mapped to node identifiers, and a direct memory access request is initiated to the workshop's underlying equipment monitoring and management system, according to the data retrieval time window. Copy the high-frequency operating parameters of the target device with defined start and end scales, and output heterogeneous timing raw data. ; S44: Call heterogeneous timing raw data With two-order structured reporting datasets Perform null value matching based on the event timestamp. Extracting heterogeneous timing data as the reference axis The included physical waveform curves are used to extract the time-series mean vector of the physical waveform curves within the truncated time window, and the corresponding continuous derivative term of the physical waveform curves is calculated. This continuous derivative term is then compared with a pre-configured topology attenuation coefficient. Perform multiplication and add the result to the time-series mean vector to solve for deterministic physical characteristics, which are then back-injected into empty form fields as interpolated physical scalars. Reorganize to generate a fully encoded interpolated time series data matrix. ; S45: The system consistency check kernel is started to perform a three-level serial error prevention audit and compare the interpolated timing data matrix. Includes a list of actual network card physical addresses and associated device interactions. Includes operator-selected device identifier The binding relationship is audited using hardware identifiers, comparing the batch parameters returned by the manufacturing execution system with the process codes. The mapping relationship is used to perform production batch auditing and compare the interpolated time series data matrix. Includes key equipment operating parameters and static initial baseline constraint vectors The defined reasonable fluctuation tolerance zone is used for tolerance interval auditing; when the results of hardware identification audit, production batch audit, and tolerance interval audit are all logically true, the interpolated time series data matrix will be used. High-density core text features Operator selects procedure number With triggering transient event dataset Structured assembly is performed in the high-security area of ​​system memory to generate a high-fidelity multimodal event dataset. .

[0012] In the preferred embodiment, step S5 includes the following sub-steps: S51: Reading a high-fidelity multimodal event dataset In the high-security area of ​​system memory, a high-fidelity multimodal event dataset is processed according to a preset pointer offset protocol. Perform modal separation and unpack the interpolated time series data matrix. High-density core text features and the dataset that triggers transient events ; Interpolate the time series data matrix The input is a feature encoding neural network containing one-dimensional causal convolutional layers. A sliding convolution operation is performed along the time axis to extract waveform slope, peak-to-peak value, and local oscillation frequencies. The output is continuous time-series modal features to generate a temporal manifold feature matrix. ; S52: Transform the temporal manifold feature matrix High-density core text features Initial feature set of exogenous text modal The input is fed into the cross-modal alignment engine, where a content condensation and reconstruction mechanism is introduced to calculate the temporal manifold feature matrix. High-density core text features The content condensation network maps the posterior conditional distribution with respect to the condensed cross-modal feature matrix under the joint prior distribution. The log-likelihood ratio approximating the marginal distribution is combined with the expected value and a pre-defined regularization penalty constant. Constrained condensation of cross-modal feature matrix Initial feature set of exogenous text modal The Frobenius norm squared terms between the terms are used to construct the mutual information condensation and reconstruction constraint objective function. Redundant modal information is filtered by successively updating the network weights to minimize the mutual information condensation and reconstruction constraint objective function, and the reconstructed output is a condensed cross-modal feature matrix. ; S53: Condensing the cross-modal feature matrix The feature vectors in the text and the initial feature set of the exogenous text modality The standard operating procedure control clause vectors contained therein are subjected to dot product and cosine similarity transformation, and the inverse of the similarity is mapped to feature distance to generate a semantic distance matrix. ;Utilizing a preset semantic recall cutoff constant in the semantic distance matrix The process involves filtering candidate clauses, extracting complete real text clause data that meets the filtering criteria from the relational database, and combining them to generate a set of reverse semantic recall control clauses. ; S54: Construct a four-dimensional causal topology primitive skeleton within the video memory space, dividing the four-dimensional causal topology primitive skeleton into a threshold dimension principal axis, a trend dimension principal axis, a process weight dimension principal axis, and a process connection dimension principal axis; trigger transient event datasets. Includes adaptive dynamic boundary threshold vector Filling to the threshold dimension principal axis will trigger the transient event dataset. Included deviation residual matrix Fill the trend dimension main axis with the process priority coefficient. Fill the interpolated time series data matrix to the principal axis of the process weight dimension. Fill in the main axis of process connection; control the set of reverse semantic recall clauses. The text semantic tags are parsed, and the text control procedures are anchored to the secondary branches corresponding to the threshold dimension main axis, trend dimension main axis, process weight dimension main axis, and process connection dimension main axis, outputting a structured lesion mapping map. ; S55: Comparison of structured lesion mapping maps Adaptive dynamic boundary threshold vector on the principal axis of the threshold dimension With static initial reference bound vector In determining the adaptive dynamic boundary threshold vector The contraction amplitude exceeds the set deviation bandwidth and the interpolated time series data matrix When non-abrupt degradation is observed, the root cause is determined to be critical missed detection failure due to threshold setting deviation; in determining the adaptive dynamic boundary threshold vector... A set of steady-state and reverse semantic recall control clauses When the text contains text indicating operational violations, the root cause is identified as a loophole in the operational procedures for process connection. The attribute identifier code corresponding to the root cause, the feature matrix involved in the judgment, and the attached text clause number are structurally encapsulated to generate the final root cause judgment result matrix. .

[0013] In the preferred embodiment, step S6 includes the following sub-steps: S61: Read the final root cause determination result matrix The final root cause determination matrix is ​​processed at the kernel level. The attribute identifier code encapsulated in the system is used to perform bitwise AND logical operations to determine the fundamental attribute of the event; when the attribute identifier code is mapped to a threshold setting deviation causing a critical missed alarm failure or false alarm trigger, the system's internal algorithm dynamic parameter adjustment bus is activated to enter the control parameter adaptive correction branch; when the attribute identifier code is mapped to a loophole in the process connection operation specification or ambiguity in the control text, the system's standard operating procedure revision push queue is activated to enter the management procedure text optimization branch. S62: Execute the adaptive correction branch for control parameters, based on the final root cause determination result matrix. Extracting interpolated time series data matrix from mounted evidence chain The physical waveform extrema within, combined with the static initial reference constraint vector Calculate the peak relative deviation scalar ; Compare the peak relative deviation scalar The absolute value and adaptive dead zone deviation threshold ; in the peak relative deviation scalar The absolute value is less than the adaptive dead zone deviation threshold. At that time, the weighting coefficients are fed back to the step size. Forced to zero to maintain absolute static hard locking of the weighting coefficients; in the peak relative deviation scalar The absolute value is greater than or equal to the adaptive dead zone deviation threshold. When calculating the peak relative deviation scalar The absolute value and adaptive dead zone deviation threshold The difference is calculated and then compared with the convergence step size adjustment factor. Perform multiplication calculations to determine the weight coefficient feedback step size. Feedback the weighting coefficients to the step size Accumulated to the original weight coefficient The above outputs the updated weight vector. Based on the magnitude of missed and false alarms, the preset ratio of the triggering conditions is adjusted synchronously, and an updated dynamic parameter set is generated and encapsulated. ; S63: Execution Management Procedure Text Optimization Branch, Extracting High-Density Core Text Features With the set of reverse semantic recall control clauses ;Activate the natural language generation model to recall the set of control clauses through reverse semantics. Based on the modified object, high-density core text features are used. The exact physical phenomenon primitives contained therein are used as supplementary entities to perform text concatenation and grammatical reconstruction, and the reconstructed revision suggestion text and process code are then combined. And serialize and encode the globally unique identifier to generate a structured revision push vector. ; S64: Creates an isolated sandbox memory area in physical memory that is physically isolated unidirectionally from the main control bus. Batch extraction of historical operating condition data snapshots and triggered transient event datasets from relational databases. In the isolated sandbox memory area Internally reassembled into historical test data packets In the isolated sandbox memory area The internally instantiated virtual synthesis degradation index calculation engine will update the dynamic parameter set. Or structured revision push vector Injecting the virtual integrated degradation index calculation engine into historical test data packets Perform a time-series replay operation; if the replay operation result indicates safe convergence and no new false positives or negatives are generated, verify the pass of the Boolean-type backtracking test. The value is assigned to logical true, and the backtracking test verification is passed when the replay operation results diverge or conflict. The logic is false, and a termination rollback interrupt signal is sent to the central processing unit's main control module. S65: Continuous monitoring of isolated sandbox memory area The output backtracking test verification passed the flag. Upon detection of the backtracking test verification pass flag Activating database write locks and enabling global transaction management when logically true; utilizing the update dynamic parameter set. Physical coverage of the static initial baseline constraint vector stored in the standard operating procedure base database The structured revision may be pushed to the vector. The text tensors in the text are remapped to the manifold space to physically replace the corresponding initial feature sets of the exogenous text modalities. The original vector entries in the database are used to complete the semantic base hard update; the return status code and confirmation timestamp of the database table structure change operation are captured, and a closed-loop update log is generated by packaging them using a cryptographic hash algorithm. The data is persistently stored in the system audit storage node, and the control system returns to steady-state monitoring mode.

[0014] In a preferred embodiment, the present invention also provides an adverse event analysis system based on a Standard Operating Procedure (SOP) system, comprising: The SOP basic data preset module is used to control the terms and conditions of the standard operating procedure text. Establishing an initial feature set for exogenous text modalities And construct a mapping table of equipment process thresholds. ; The 3D trigger model construction and trigger module is used to acquire physical sensor parameters to construct process monitoring data vectors. Based on process monitoring data vector Latch-triggered transient event dataset ; The step-by-step reporting management module is used to report data based on the initial manual qualitative text. Initial feature set of exogenous text modal Generate a unique event identifier primary key The level signal is locked by the interface hardware. Block and output a two-level structured reporting dataset ; The cross-system data automatic association module is used to identify primary keys using unique events. Retrieve heterogeneous time series raw data And reverse-inject a two-level structured reporting dataset. Generate a high-fidelity multimodal event dataset ; Fishbone diagram root cause localization module, used to locate the root cause of events based on a high-fidelity multimodal event dataset. Generating the temporal manifold feature matrix The temporal manifold feature matrix Initial feature set of exogenous text modal Mapping to a set of recall reverse semantic recall control clauses And generate a structured lesion mapping atlas. Determine the final root cause determination matrix. ; The triggering rules and SOP closed-loop optimization module is used to parse the final root cause determination result matrix. isolated sandbox memory area When the backtracking test converges safely, update the static initial baseline constraint vector. .

[0015] In a preferred embodiment, the present invention also provides a computer device comprising at least one processor coupled to at least one memory storing at least one computer program or instruction, wherein the computer program or instruction is loaded and executed by the processor to implement the steps of the adverse event analysis method based on the SOP system as described in any of the preceding embodiments.

[0016] In a preferred embodiment, the present invention further provides a computer-readable storage medium, wherein a computer program or instructions are stored on the computer-readable storage medium, and when the computer program or instructions are executed by a processor, the steps of the adverse event analysis method based on the SOP system as described in any of the above embodiments are implemented.

[0017] In a preferred embodiment, the present invention further provides a computer program product, including a computer program or instructions, wherein when the computer program or instructions are executed by a processor, they implement the steps of the adverse event analysis method based on the SOP system as described in any of the preceding embodiments.

[0018] This invention provides an adverse event analysis system and method based on a Standard Operating Procedure (SOP) system. Through the coordination of the above-mentioned structures, compared with existing methods, it has the following advantages: First, by introducing an adaptive dynamic threshold and trend joint tracking mechanism, the unsupervised prediction model is called to solve the deviation residual matrix online, and the historical deviation variance and process priority coefficient within the dynamic time sliding window are organically integrated to realize the dynamic expansion and contraction of the adaptive dynamic boundary threshold vector envelope and the temporal divergence measurement. This enables the device to perform high-sensitivity capture of slowly changing drift faults that have not reached the static red line but show a micro-divergent deterioration trend in multiple consecutive sliding statistical periods, thereby reducing the probability that micro-critical risks will be directly filtered out by the existing system and lead to the underreporting of potential hidden dangers. Secondly, by constructing a two-stage asynchronous qualitative data entry control chain, in the initial simplified data entry stage performed by the operator, a content condensation and reconstruction neural network based on the mutual information minimization strategy is activated. In the manifold space, the colloquial noise and subjective redundant expressions in the initial manual text are directly stripped away, and high-density core text features are solved and output. At the same time, the unique event identifier primary key with embedded spatiotemporal process attributes drives the database primary and foreign key cascading triggers. Dynamic view clipping and interface hardware locking are performed on the equipment identifier and procedure number in the subsequent secondary data entry to prevent errors. This not only decouples the constraints of complex field data entry on the timeliness of the initial response to emergencies, but also reduces the high-frequency noise caused by manual data entry based on experience at the source of data entry, ensuring the fidelity of the data base throughout the system's entire life cycle. Third, by using the unique event identifier primary key as a retrieval beacon, the cross-system data automatic routing module is driven, the protocol conversion socket is activated to actively subscribe to and directly access the physical register nodes of the multi-source heterogeneous control system, and the nearest neighbor time-series mean data interpolation algorithm based on the process mainline topology is called to perform deterministic physical reverse injection of missing form fields, including waveform derivatives and deviation slopes. Finally, the series verification of equipment, batch and tolerance tolerance band is completed through a three-level linkage strong consistency boundary audit mechanism; thus, high-precision automated complete extraction of micro-transient underlying physical time-series data streams before and after sudden safety events is achieved. Fourth, by activating the cross-perspective text fusion alignment engine, the temporal manifold feature matrix and the condensed cross-modal feature matrix are projected into the same manifold semantic space to perform feature matrix dot product multiplication. The semantic distance matrix between the quantitative physical waveform and the qualitative standard operating procedure text clauses is calculated in real time. Based on the truncation constant, the set of control clauses is recalled in reverse semantics and mounted onto the main axis of the four-dimensional causal topology primitive to perform cross-causal auditing and qualitative attribution. This enables the precise location of specific defect clauses or threshold deviations that cause production deviations. Fifth, by introducing a feedback correction strategy with an adaptive dead-zone control function at the end of the feedback loop, and utilizing the cascaded comparison results of the peak relative deviation scalar and the adaptive dead-zone deviation threshold, the weighted constant coefficients of each characteristic channel and the trigger preset ratio in the calculation of the comprehensive degradation index of the online self-adjusting front end are used. Long-cycle historical monitoring data packets are imported into the isolated sandbox memory area that is physically isolated from the main control bus to perform historical monitoring data replay backtracking tests. After outputting the backtracking test verification pass mark, the database primary and foreign key mapping table is refreshed in a cascaded manner. This effectively ensures the long-cycle high stability and mathematical safety of the entire control process during the adaptive evolution process. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the invention. Figure 2 This is a system structure diagram of the present invention; Figure 3 This is a schematic diagram of the structure of the computer device of the present invention. Detailed Implementation

[0020] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features described herein can be combined with each other without conflict. The exemplary embodiments disclosed herein will be described below with reference to the accompanying drawings, including specific technical details disclosed to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures are omitted in the following description.

[0021] To improve the accuracy and automation level of production line anomaly identification, existing technologies have proposed a rolling adjustment scheme for production line simulation based on digital twins, such as the rolling optimization system and method for production line simulation disclosed in Chinese Patent Publication No. CN113361139B. This scheme periodically extracts recent production data and real-time event signals from the workshop manufacturing execution system (MES) and equipment monitoring and management system through a data integration interface. It then uses mathematical distribution models such as normal distribution, Ellang distribution, and binomial distribution to fit and reconstruct the feature values ​​of the process cycle analysis model, equipment fault analysis model, and production quality analysis model, thereby updating the digital twin model to correct simulation errors.

[0022] However, existing systems have two shortcomings in solving the aforementioned industry problems: First, the risk triggering mechanism is simplistic and mechanical. Iterative evolution triggering relies heavily on a single static deviation limit between actual monitored values ​​and the macroscopic simulation model, or on explicit sudden equipment downtime anomalies, as defined in cited document D1, where the deviation exceeds 5%. Due to the lack of cumulative integral characteristics of the temporal deviation magnitude and weighted inputs for the rigidity of process topology constraints, the existing system cannot detect micro-critical risks that have not reached the static limit but have shown a continuous deterioration trend over multiple consecutive sliding statistical periods. This results in potential hazards being directly filtered out and missed by the existing system before reaching the limit deviation.

[0023] Second, there is a lack of mechanisms for cross-system integration and seamless completion verification of micro-transient data. Data integration in citation document systems only performs distribution fitting and reconstruction based on statistical averages over macro-historical time periods, lacking an automatic extraction mechanism for micro-transient underlying physical time-series data streams that adaptively extend across subsystems and time axes for specific sudden event identifiers. This deficiency prevents the system from adapting to the progressive information processing requirements of first providing extremely simple second-level responses and then asynchronously supplementing structured data. It also fails to automatically and intelligently interpolate and perform strong consistency static verification on manually entered fields using primary and foreign key error-proofing constraints in relational databases. Furthermore, it cannot directly map, feed back, and replace the root cause analysis results with specific standard operating procedure (SOP) text control clauses in the basic database.

[0024] To overcome the above-mentioned shortcomings, this technical solution introduces an adaptive dynamic threshold and trend joint tracking mechanism as well as a cross-modal semantic automatic mapping mechanism, forming an adverse event analysis system and method based on the SOP system.

[0025] Example 1: like Figure 1 As shown in the figure, this embodiment provides a method for analyzing adverse events based on a SOP system. The specific steps are as follows: S1: Pre-set the basic data matrix of standard operating procedures required for adverse event analysis in the business relationship database, establish the underlying static topology mapping between production equipment, process nodes and control limits, and complete the initialization of the baseline parameters of dynamic boundary tracking and semantic mapping space.

[0026] S11: Extract the standard operating procedures, quality control requirements, and equipment operation standards for each production process, and establish a static initial baseline constraint vector. .

[0027] In practice, quality-related processes need to specify a minimum product qualification rate, and equipment-related processes need to specify a baseline value for single equipment downtime. The system processor will define a vector for each static initial baseline. Corresponding process code Perform unique memory address binding to form a basic parameter dictionary containing control red lines.

[0028] S12: Organize all production equipment required for each production process and construct a structured list of related equipment interactions. .

[0029] Specifically, the processor interacts with the associated device list. Record the equipment model of each production machine , To which process node Equipment Management Identification and equipment operating parameter acquisition port information The processor will then associate the device interaction list. Each piece of production equipment is associated with a corresponding process code. The static initial reference constraint vector that has been bound to this process. Perform three-dimensional primary and foreign key association binding to generate a table mapping relationship between equipment process thresholds. And map the equipment process thresholds to the table. Persistent storage is performed in the business relationship database, establishing a physical addressing foundation for accurate routing and retrieval of underlying physical time-series data streams across systems.

[0030] S13: Quantify the impact weight of each production process on the final product quality and the constraint stiffness on the advancement of subsequent processes, and set and map process priority coefficients. .

[0031] In one feasible approach, the processor assigns a process priority coefficient. It is divided into three discrete level scores.

[0032] The first priority corresponds to the bottleneck process that directly determines the product quality characteristics and cannot be carried out in any subsequent process when interrupted. Second-level priority corresponds to related processes that affect product quality characteristics and require subsequent processes to be paused and waited for when there is a delay. When a level 3 priority error occurs, subsequent processes can be adjusted to proceed normally through a temporary buffer to the final process.

[0033] The processor assigns priority coefficients to each process. Corresponding process code One-to-one mapping.

[0034] S14: Initialize the data distribution of the long-period baseline state for the adaptive dynamic boundary triggering mechanism, and construct the long-period data feature matrix. .

[0035] Specifically, to overcome the mechanical limitations of a fixed linear threshold, the processor retrieves a long-term historical sequence of physical parameters under continuous, error-free operating conditions through the data integration interface of the workshop manufacturing execution system. .

[0036] Preferably, a long-period physical parameter history sequence It covers stable operation data for 72 consecutive hours.

[0037] The processor calculates the long-term physical parameter history sequence Long-period mean vector and long-period variance matrix .

[0038] Long-period mean vector With long-period variance matrix Together they form the initial long-period data feature matrix. This is used to provide an adaptively evolving health baseline for unsupervised joint anomaly detection in subsequent steps, ensuring that the dynamic decision boundary can automatically track the normal physical drift of the equipment's production cycle.

[0039] S15: Initialize the exogenous semantic base of the standard operating procedure text for the cross-modal semantic automapping mechanism, and generate the initial feature set of the exogenous text modality. .

[0040] In practice, the processor reads all standard operating procedure text control clauses stored in the relational database. The processor utilizes a pre-trained natural language processing backbone network to process standard operating procedure text control clauses. The process involves word segmentation, stop word removal, and semantic feature extraction to generate an initial feature set of exogenous text modalities that includes process management constraints and operational limitations. Initial feature set of exogenous text modalities It is stored in a high-dimensional manifold space for subsequent steps to perform mutual information minimization compression and cross-modal feature distance calculation with the physical time-series data stream.

[0041] S2: Based on the adaptive dynamic threshold and trend joint tracking mechanism, it continuously parses the multi-dimensional process monitoring data stream online through the underlying hardware interface, calculates the adaptive dynamic boundary threshold vector and comprehensive deterioration index in real time, performs time-series divergence measurement on micro-critical risks, and determines and executes the adaptive triggering of the adverse event analysis process.

[0042] S21: Real-time retrieval of physical sensor timing parameters for each process node through the workshop data acquisition layer to construct process monitoring data vectors. .

[0043] Specifically, the central processing unit uses the equipment process threshold mapping relationship table established in step S12. Dynamically identify the list of interactions with associated devices Equipment management identification for each piece of production equipment Corresponding device operating parameter acquisition port information .

[0044] The central processing unit extracts pressure, flow, temperature, time amplitude, and timing waveform data of each controlled hardware device through the industrial bus interface at a preset hard real-time sampling period.

[0045] Preferably, the preset hard real-time sampling period is set to 50 milliseconds.

[0046] The central processing unit is at the current time step. The retrieved multidimensional physical sensor parameters are then spatiotemporally aligned and denoised, and normalized to assemble and generate the time step. Step-by-step process monitoring data vector Process monitoring data vector As a multidimensional feature column vector, it is temporarily stored in the system's central processing unit cache in real time, serving as the real-time input source for subsequent adaptive dynamic boundary iteration and degradation measurement.

[0047] S22: Call the long-period data feature matrix initialized in step S14. An unsupervised prediction model is established based on a time-series sliding window to solve the current time step. The expected vector of the step prediction .

[0048] In one feasible approach, the central processing unit allocates a segment of memory of length [length missing] in running memory to track normal physical drift of the production cycle. A dynamic time-series sliding window.

[0049] Preferably, the length of the dynamic time-series sliding window Set to 3600 consecutive time steps.

[0050] The central processing unit reads historical process monitoring data within a dynamic time-series sliding window and combines it with a long-cycle data feature matrix. Long-period mean vector With long-period variance matrix A Gaussian mixture distribution is fitted to calculate the evolution probability density of the current operating condition under no-anomaly conditions. By performing maximum likelihood estimation on the state evolution probability density, the central processing unit calculates the current time step online. Predicted expected vector at the statistical health baseline Predict the expected vector Dimensions and process monitoring data vectors The dimensions are completely consistent, representing the theoretical physical state value when there is no abnormal deviation in the current process node.

[0051] S23: Compare the measured values ​​from time-series facts with the predicted values ​​from unsupervised theory, measure the micro-transient deviation characteristics of actual operating conditions, and generate the deviation residual matrix. .

[0052] In practice, the central processing unit retrieves the current time step from the cache. Step-by-step process monitoring data vector The predicted expectation vector calculated in step S22 The central processing unit performs vector subtraction to obtain the process monitoring data vector. With the predicted expected vector The transient deviation residuals are used to capture the cross-coupling and divergent characteristics between multidimensional physical parameters. The central processing unit (CPU) performs an outer product matrix operation on the transient deviation residuals and their transposes to construct the deviation residual matrix. Deviation from residual matrix The mathematical expression is as follows: (1); in, For the current time step The deviation residual matrix of the step; For time step number Step-by-step process monitoring data vector; For the current time step The expected vector of the step prediction; All are deviation residual matrices The matrix elements in the matrix represent the first... Monitoring parameters of the dimensional process and the first Monitoring parameters for the maintenance process at time step 1 The intensity of the crossover deviation of the step; This represents the total feature dimension of the process monitoring data. Deviation residual matrix. It fully locks the nonlinear anomaly amplitude at the microscopic level of the current physical pipeline.

[0053] S24: Combine the static initial reference constraint vector in step S11 Based on the nearest time-series waveform distribution, perform adaptive dynamic boundary iteration to solve the adaptive dynamic boundary threshold vector in real time. .

[0054] Specifically, to eliminate critical false negatives caused by fixed dead thresholds, the decision boundary must be flexibly adjusted according to the time-varying operating conditions of the production line. The central processing unit introduces a degradation trend negative feedback damping mechanism to adaptively adjust the dynamic boundary threshold vector. The formula is as follows: (2); in, For the current time step The adaptive dynamic boundary threshold vector of the step; The static initial reference constraint vector established in step S11; The Hadamard product represents the digit-wise multiplication of corresponding elements of a matrix or vector. It is a unit column vector consisting entirely of 1s; The adaptive boundary adjustment coefficient is set to a continuous real number between 0.05 and 0.15. The length of the dynamic time-series sliding window; Characterization extraction time step Deviation residual matrix of step The deviation variance column vector formed by the diagonal elements.

[0055] Adaptive dynamic boundary threshold vector By accumulating the deviation variance within the dynamic time-series sliding window through logarithmic integration, the decision boundary can adaptively and dynamically shrink the envelope for slowly changing drift faults, thus achieving high-sensitivity tracking of the boundary for microscopic deterioration trends.

[0056] S25: Fusion Deviation Residual Matrix Adaptive dynamic boundary threshold vector With the process priority coefficient in step S13 The comprehensive degradation index is calculated by using a multidimensional hybrid linear weighting and nonlinear compression function. .

[0057] In practice, the central processing unit calls the deviation residual matrix generated in step S23. Extract the trace term and recall the adaptive dynamic boundary threshold vector solved in step S24. Perform norm measurement and simultaneously introduce the process priority coefficient from step S13. As a structural stiffness weighting coefficient. To prevent amplitude overload of discrete ordinal levels, the central processing unit uses a hyperbolic tangent function to weight the process priority coefficients. Perform nonlinear continuous mapping. Overall degradation index. The calculation formula is as follows: (3); in, For the current time step The comprehensive deterioration index of the step is a dimensionless composite risk measurement index. For deviation residual matrix The trace represents the total deviation energy of the multidimensional monitoring features at the current time step; For adaptive dynamic boundary threshold vector Relative static initial reference bound vector The L2 distance represents the significance of the cumulative degradation trend of boundary contraction; The process priority coefficient mapped in step S13; It is the hyperbolic tangent activation function; , , These are the weighted constant coefficients for the corresponding dimensions.

[0058] Preferably, the weighting constant coefficients satisfy... Normalization constraints. Overall degradation index. The production characteristics of heterogeneous and heterogeneous quantities are deeply coupled into a single flow evaluation value.

[0059] S26: Overall Deterioration Index The system performs a tiered condition check control chain audit. When a risk threshold is crossed, the hardware trigger level is toggled, bus data is blocked, and the trigger transient event dataset is latched. .

[0060] In one feasible approach, the central processing unit calculates the overall degradation index of the output. Continuous comparison is performed against the preset safety threshold value. When the overall degradation index... If the time span exceeding the safety threshold reaches the preset anti-jitter filter time limit, the central processing unit (CPU) determines that a deviation from the production line has occurred. The CPU immediately sends a hardware interrupt instruction to the physical bus actuator, toggling the trigger level to block material dispatching on the current production line, and simultaneously resetting the current time step. Historical process monitoring data vector before and after the step Sequence, Deviation Residual Matrix Snapshot and Adaptive Dynamic Boundary Threshold Vector Snapshots are packaged at the snapshot level to generate a dataset that triggers transient events. .

[0061] Preferably, the transient event dataset is triggered. The time span covers the production cycle time domain, specifically two hours before and after the trigger. (Triggering transient event dataset) Write it into the system flash memory as the global master data source for subsequent step-by-step asynchronous reporting and cross-system joint debugging.

[0062] S3: Based on the two-stage asynchronous qualitative filling and multimodal content condensation and reconstruction of the trigger event data, a unique event identifier primary key containing spatiotemporal process constraints is generated, and cascaded hard lock error prevention verification is performed.

[0063] In practical implementation, to address the multimodal disconnect between the qualitative text data subjectively and asynchronously filled in by frontline operators and the physical time-series data stream of the process control bus, this step introduces a mutual information minimization strategy and a content condensation and reconstruction mechanism. This mechanism aims to achieve a simplified data entry stage where operators respond first and then refine, immediately filtering out subjective redundancy in the text at the underlying level. This aligns the cross-modal semantic benchmark between the manually entered text and the static procedures, and utilizes primary and foreign key cascade triggers in a relational database to perform hard locking for subsequent inputs, ensuring that the dataset transmitted to the mid-to-back-end has absolute physical and semantic fidelity.

[0064] S31: The system's central processing unit detects the trigger transient event dataset output in step S2. After writing to the flash memory, the simplified initial reporting process of the front-end industrial tablet is activated to obtain the initial human interaction parameters.

[0065] Preferably, the front-end interface implements strong dimensionality reduction and locking, opening only three core interaction ports. The central processing unit automatically captures the current event timestamp through the operating system's underlying clock and the global positioning system interface. Physical location coordinates of the event It also receives initial human-generated qualitative text input by the operator via voice transcription or keyboard. Initial manual qualitative text This typically includes highly colloquial, subjective, and redundant descriptions of the fault phenomenon. At this point, the central processing unit timestamps the event. Physical location coordinates of the event Compared with the initial human qualitative text It is temporarily stored in the application layer memory queue.

[0066] S32: Initiate a cross-perspective text fusion and content condensation and reconstruction mechanism for the initial manually qualitative text. Perform mutual information minimization mapping to solve and output high-density core text features. .

[0067] At this stage, in order to eliminate data noise introduced by artificially inflated numbers, the central processing unit calls the exogenous text modality initial feature set that was initialized in step S15. , with the initial manually defined text in the memory queue Perform cross-modal semantic space projection. The central processing unit constructs a content condensation and reconstruction neural network, employing a mutual information minimization strategy to remove the initial manually defined text. Redundant subjective expressions, such as "I think" or "it might be broken," are removed. Only operational actions and physical phenomenon primitives with strong mapping relationships to the management constraints of the standard operating procedure are retained. The mathematical expression of the mutual information minimization objective function is as follows: (4); in, The objective function is to minimize mutual information; It is a mutual information metric operator used to calculate the nonlinear dependence between two random variable domains; The initial manual qualitative text obtained in step S31; The high-density core text features to be output by the content condensation neural network; The initial feature set of the exogenous text modality serves as a guiding benchmark for semantic preservation; Semantic-preserving Lagrange multipliers are used to balance compression ratio and feature fidelity.

[0068] Optimization through backpropagation Upon reaching convergence, the central processing unit outputs high-density core text features filtered out from spoken noise. High-density core text features As a highly structured semantic tensor, it directly replaces the original messy text, eliminating the semantic gap barrier for cross-modal fishbone diagram alignment in the subsequent step S5.

[0069] S33: Based on location mapping routing rules, extract the underlying process semantics and combine them with time variables to generate a unique event identifier primary key with embedded spatiotemporal process constraints. .

[0070] Specifically, the central processing unit calls the physical location coordinates of the event extracted in step S31. The persistent storage of the device process threshold mapping table in step S12 Initiate a spatial topology addressing query and retrieve the process code corresponding to the physical location in reverse order. Subsequently, the central processing unit uses a string hash concatenation algorithm to encode the process. Event timestamp The six-bit secure random sequence code generated by the hardware true random number generator is physically concatenated to generate a unique event identifier primary key. Unique event identifier primary key It possesses global uniqueness and embeds specific node attributes and time section attributes of the physical pipeline directly within its hash table layer, serving as the underlying data root node that runs through the upstream and downstream data flow of the system.

[0071] S34: Instantiate a dual-loop sequential state machine Based on the unique event identifier primary key The execution process controls the state transition and starts a countdown hardware clock.

[0072] Generate a unique event identifier primary key Within a microsecond-level instruction cycle, the central processing unit instantiates a dual-loop sequential state machine corresponding to the current event in memory. The central processing unit (CPU) directs the data to a dual-closed-loop sequential state machine. Inject a jump instruction to force its running state to transition from idle state to emergency suspended state. The rising edge signal of the state transition triggers the high-precision timer on the system motherboard, starting the supplementary timer. .

[0073] Preferably, a supplementary timer is provided. The countdown full-scale configuration is 20 minutes. A supplementary timer is available. During the golden window of opportunity before the system reaches zero, a dual-closed-loop sequential state machine is used. Maintain emergency suspension status The system also opens up additional fields for the front-end industrial tablet, allowing operators to enter specific device associations and compliance procedure numbers for a second time.

[0074] S35: Apply a unique event identifier primary key Drive the database primary and foreign key cascading triggers to dynamically prune and cascade hardware locking to prevent errors in the structured interactive commands input at secondary levels.

[0075] Within this restriction window, the operator must submit the selected device identifier via the drop-down menu on the front-end interface. Operator selected procedure number To prevent data pollution caused by human error, the central processing unit extracts a unique event identifier primary key. Process coding in Coded by this process As a first-normal form filtering condition, the associated device interaction list in step S12 is used. Perform dynamic view clipping to generate a subset of cascaded constraint devices that belong only to the current process. .

[0076] The central processing unit executes a Boolean logic comparison instruction at the application kernel layer to determine the operator-selected device identifier passed from the front end. Does it belong to a subset of cascaded constraint devices? The internal elements, i.e., the calculations If the judgment result is logically false, meaning the operator arbitrarily selected an irrelevant device not related to the current process, the central processing unit's low-level interrupt request module immediately outputs a hardware latching signal to the front-end I / O controller. Interface hardware latching level signal Directly cut off the underlying touch event response of the "Submit" button on the front end, forcibly blocking the path of dirty data entering the database until the operator changes to the matching device identifier.

[0077] S36: Assemble multimodal data within the countdown constraint and output a two-order structured reporting dataset. Complete the construction of a closed data link.

[0078] When a dual closed-loop sequential state machine Determine if the Boolean comparison logic is true, and add a timer. Before the overflow reaches zero, the CPU releases the interface suspension and executes the data stream assembly. The CPU assigns a unique event identifier to the primary key. The high-density core text features generated in step S32 Operators who have passed verification select the device identifier. Operator selected procedure number and the dataset of triggering transient events extracted in step S2 Perform deep data packet encapsulation to uniformly generate a two-level structured reporting dataset. Two-order structured reporting dataset The data is pushed downstream via a high-speed memory bus to the data pre-alignment and cross-system automatic interpolation modules, becoming the only legitimate data input source for subsequent microscopic transient physical time series extraction and fishbone diagram root cause analysis.

[0079] S4: Based on the unique event identifier primary key, it crosses heterogeneous industrial buses, performs spatiotemporal adaptive data routing and intelligent interpolation of underlying microscopic physical parameters, implements multidimensional strong consistency auditing, and assembles high-fidelity multimodal event datasets.

[0080] In practical implementation, addressing the shortcomings of lacking a micro-level transient cross-system joint debugging and seamless completion verification mechanism, and the fact that data integration only performs distribution fitting based on statistical averages within a macro-level historical time period, failing to adapt to asynchronous structured supplementation after a second-level simplified response, this step uses the primary key generated in step S3 as a beacon to construct a dynamic relational routing parsing target subsystem. On-demand data retrieval commands are directly issued to the underlying industrial control hardware through a dedicated protocol conversion socket, and high-precision intelligent interpolation is performed on missing fields in the form using the slope of the physical waveform and the mean of nearest neighbors. Finally, boundary error prevention verification is performed through a relational database mapping matrix.

[0081] S41: Parse the unique event identifier primary key The semantic hash bits are used to reverse extract spatial topology nodes and temporal cross-sectional parameters, and the data retrieval time window is dynamically configured. .

[0082] Specifically, the central processing unit calls the two-level structured reporting dataset output in step S3. Extracting the two-order structured reporting dataset The header encapsulates a unique event identifier primary key. The central processing unit (CPU) initiates the underlying event identifier parsing engine to uniquely identify the primary key of the event. Perform a reverse hash operation on the string. The central processing unit (CPU) uses bitmasking techniques to precisely extract the process code to which the event belongs. Timestamp of the event .

[0083] Get the event timestamp Then, the central processing unit reads the pre-set process code from the basic database. The corresponding standard production cycle parameters. The central processing unit uses the event timestamp. Using the origin as a baseline, a standard production cycle duration is extended bidirectionally into both the historical and future time domains to construct an adaptively aligned data retrieval time window. Data retrieval time window Along with process codes It is pushed to a temporary buffer area to serve as the global spatiotemporal constraint boundary for subsequent cross-system data retrieval instructions.

[0084] S42: Joint Process Code With the operator selecting the device identifier Perform primary and foreign key cascading addressing in the relational database to generate a list of data retrieval requirements for heterogeneous systems. .

[0085] In one feasible approach, the central processing unit retrieves the procedure codes from the temporary cache in parallel. Data retrieval time window and two-order structured reporting dataset Operator selects device identifier The central processing unit utilizes process coding. The equipment process threshold mapping table constructed in step S1 Initiate an inner join query to extract the list of interactions between related devices involved in this process. All physical sensor measurement points and production execution management control points defined in the document.

[0086] The central processing unit further compares the identifier of the device selected by the operator. The physical sensor measurement points returned by the query will be matched with the underlying programmable logic controller (PLC) register address and manufacturing execution system (MES) production batch number, along with the data retrieval time window. Unified packaging to form a formatted list of data retrieval requirements from heterogeneous systems. .

[0087] S43: Activate protocol conversion socket, based on heterogeneous system data retrieval requirement list Perform seamless cross-system handshakes and hard real-time data acquisition on industrial buses to obtain heterogeneous timing raw data. .

[0088] In practice, the central processing unit starts a pre-configured JSON-OPCUA protocol conversion socket to receive a list of data retrieval requests from heterogeneous systems. The protocol conversion socket utilizes a built-in mapped hash table structure to retrieve data retrieval requests from heterogeneous systems. The qualitative fields are automatically mapped to node identifiers (NodeIDs) that conform to the IEC 62541 international standard.

[0089] Preferably, when sending a read request to the workshop-level equipment monitoring and management system, the protocol-converted socket adopts a dual-track parallel transmission mode of active polling and event subscription. The central processing unit directly initiates a direct memory access (DMA) request to the DB data block register area of ​​the underlying programmable logic controller, according to the data retrieval time window. With defined start and end scales, the system fully copies the high-frequency operating parameters of the target device (such as instantaneous current, spindle speed, fault alarm codes, etc.) and outputs heterogeneous timing raw data. Heterogeneous time series raw data By bypassing any artificial intermediary layer, the objectivity and immutability of the physical fact snapshot are ensured.

[0090] S44: Execute the nearest-neighbor time-series mean intelligent interpolation algorithm based on the process mainline topology for the two-order structured reporting dataset. Deterministic physical feature inverse injection is performed on the missing form fields to generate an interpolated time series data matrix. .

[0091] In one specific implementation, for critical null value fields such as specific fault breakdown codes and instantaneous fluctuation extreme values ​​that front-line operators may have omitted during emergency response, the central processing unit calls heterogeneous timing raw data. With two-order structured reporting datasets Perform null value matching. The central processing unit uses the event timestamp. Using the reference axis, extract the raw heterogeneous time series data. For physical waveform curves within a very short time span, calculate the waveform derivative and deviation slope, and then forcibly inject the calculated deterministic eigenvalues ​​into the missing fields. Interpolation physical scalar. The mathematical model is as follows: (5); in, For the first The interpolated physical scalars corresponding to each missing field; To extract heterogeneous time series raw data within a time window The time-series mean vector; The topology attenuation coefficient is pre-configured based on the inertia constant of a specific process in the physical topology. For the continuous derivative terms of the physical time-series waveform; The timestamp of the event; This is the derivative sampling offset.

[0092] By solving formula (5), the central processing unit accurately reconstructs the physical dynamic characteristics at the moment the fault occurred, eliminating information gaps caused by human omissions in filling out forms. The central processing unit then converts all generated interpolated physical scalars... Reorganize with the original data to output an interpolated time series data matrix with all fields fully encoded. Interpolated time series data matrix The values ​​in the database are in a read-only locked state at the application layer, prohibiting manual modification.

[0093] S45: Establish a three-level linkage strong consistency boundary audit mechanism to compare the interpolated time series data matrix. A high-fidelity multimodal event dataset is generated by assembling the dataset after verification, based on static topological constraints. .

[0094] Specifically, after the data splicing is completed, the central processing unit starts the system consistency verification kernel and performs a three-level serial error prevention audit: Level 1 Hardware Identification Audit: Comparison of Interpolated Timing Data Matrix The actual network card physical address included in the list of devices associated with step S12 Operator selects device identifier The binding relationship was verified to confirm that no logical misalignment occurred at the device level; Level 2 Production Batch Audit: Compare the batch parameters returned by the Manufacturing Execution System with the process codes. The mapping relationship confirms that the production process timeline has not drifted; Level 3 tolerance interval audit: Comparison of interpolated time series data matrix The key equipment operating parameters are verified to ensure that these parameters strictly fall within the vector defined by the static initial reference in step S11. Within the defined reasonable fluctuation tolerance zone.

[0095] The CPU determines that the currently retrieved and interpolated data is valid only if all three levels of audit results are logically true. Two-level structured reporting dataset High-density core text features in the middle of the transmission Operator selects procedure number and the dataset of triggering transient events passed in step S2 The data packets are then structurally assembled within a high-security area of ​​the system memory. The assembled packets are then labeled as a high-fidelity multimodal event dataset. .

[0096] S5: Based on a high-fidelity multimodal event dataset Perform cross-modal semantic automatic mapping and content condensation reconstruction, construct structured four-dimensional causal topology primitives, and achieve high-precision semantic alignment and root cause determination of underlying physical time sequence and management procedure text.

[0097] In practical implementation, to address the deficiency in conventional systems that can only perform natural language processing word frequency statistics on reported text, resulting in a multimodal disconnect between subjectively asynchronous qualitative text data and the high-frequency physical time-series data stream of the process control bus, this stage introduces a cross-perspective text fusion and content condensation and reconstruction mechanism. The central processing unit (CPU) forces the exogenous text modality and the continuous time-series modality to be placed in the same semantic manifold space to calculate feature distance, automatically filters redundant subjective invalid expressions, and uses underlying physical waveform features to reverse semantically recall the control clauses that actually violated the rules from the standard operating procedure knowledge base, thus completing the noise reduction mapping from physical quantitative features to textual qualitative procedures.

[0098] S51: High-fidelity multimodal event datasets Perform data unpacking and modal separation, and perform temporal manifold feature encoding on the underlying physical parameters to generate a temporal manifold feature matrix. .

[0099] Specifically, the central processing unit reads the high-fidelity multimodal event dataset output in step S4. The central processing unit (CPU) processes the high-fidelity multimodal event dataset in the high-security area of ​​system memory according to a preset pointer offset protocol. Perform modal separation. The central processing unit unpacks the interpolated timing data matrix output from step S4. The high-density core text features output in step S3 and the dataset of triggering transient events output in step S2 .

[0100] The central processing unit calls a feature encoding neural network containing one-dimensional causal convolutional layers to receive interpolated time-series data matrices. As input, the feature encoding neural network performs a sliding convolution operation along the time axis to extract the interpolated time-series data matrix. It includes the waveform slope, peak-to-peak value, and local oscillation frequency. The central processing unit outputs continuous time-series modal features, i.e., time-series manifold feature matrix, through a feature encoding neural network. Temporal manifold characteristic matrix Discrete physical sampling points are transformed into continuous feature vectors in a high-dimensional semantic space, providing a quantitative physical foundation for subsequent cross-modal interactions.

[0101] S52: Combining high-density core text features Initial feature set of exogenous text modal Perform mutual information minimization and content condensation reconstruction to solve the condensed cross-modal feature matrix. .

[0102] In one feasible approach, to eliminate potential subjective cognitive biases from operators, the central processing unit (CPU) initiates a content condensation and reconstruction mechanism. The CPU then converts the temporal manifold feature matrix... High-density core text features Compared with the initial feature set of exogenous text modalities initialized in step S1 Parallel input is fed into the cross-modal alignment engine.

[0103] The cross-modal alignment engine introduces a mutual information minimization strategy, which mandates the use of high-density core text features. With the characteristic matrix of the time-series manifold In the jointly generated representation, features that are similar to the initial feature set of the exogenous text modality should be filtered out as much as possible. (i.e., absolutely objective standard operating procedure control clauses) irrelevant redundant modal information. The central processing unit calculates and optimizes the mutual information condensation and reconstruction of the constraint objective, the constraint objective function formula is as follows: (6); in, To condense and reconstruct the constraint objective function for mutual information; The mathematical expectation operator for the joint probability distribution; The characteristic matrix of the time-series manifold High-density core text features The joint prior distribution of; The posterior conditional distribution of the content condensation network mapping; To condense the cross-modal feature matrix The distribution near the edge; This is the regularization penalty constant, used to control the smoothness of the mapping space; For the squared term of the Frobenius norm.

[0104] The constraint objective function is reconstructed by successively updating network weights to minimize mutual information concentration. The central processing unit outputs a condensed cross-modal feature matrix that completely eliminates artificial interference noise. Condensed cross-modal feature matrix It achieves modal-level high-fidelity reconstruction from physical time-series indicators to high-density text.

[0105] S53: Compute condensed cross-modal feature matrices in the same semantic manifold space Initial feature set of exogenous text modal Feature distance, generating semantic distance matrix And execute the set of reverse semantic recall control clauses. .

[0106] In practice, the central processing unit will condense the cross-modal feature matrix. Each row of feature vectors in the dataset is compared with the initial feature set of the exogenous text modality. The central processing unit (CPU) performs dot product and cosine similarity calculations on the vectors of all standard operating procedure control clauses included in the calculation. The CPU then maps the inverse of the calculated similarity to feature distances, generating a multi-dimensional semantic distance matrix. .

[0107] semantic distance matrix The semantic matching distance between the current sudden anomaly's physical waveform and each text control procedure in the knowledge base is fully recorded. The central processing unit (CPU) sets a semantic recall truncation constant. The CPU then processes the semantic distance matrix. In the process of filtering out all candidate options whose feature distance is less than the semantic recall cutoff constant, the corresponding real text clause data of these candidate options are completely extracted from the relational database and combined to generate a set of reverse semantic recall control clauses. A set of reverse semantic recall control clauses It replaced the procedures that were subjectively filled out by operators and became the objective legal basis for violations related to accuracy.

[0108] S54: Construct a four-dimensional causal topological primitive skeleton, and integrate the set of reverse semantic recall control clauses. And trigger transient parameters to be loaded into multidimensional branches to generate structured lesion mapping atlases. .

[0109] Preferably, the central processing unit constructs a four-dimensional causal topology primitive structure within the video memory space. The four-dimensional causal topology primitive structure is divided into four logical axes: threshold dimension axis, trend dimension axis, process weight dimension axis, and process connection dimension axis.

[0110] The central processing unit begins executing the branch data loading operation: The central processing unit will trigger the transient event dataset. Adaptive dynamic boundary threshold vector within Fill to the threshold dimension principal axis; The central processing unit will trigger the transient event dataset. Deviation residual matrix within Fill to the main axis of the trend dimension; The central processing unit will assign the process priority coefficient in step S13. Fill to the main axis of the process weight dimension; The central processing unit will interpolate the timing data matrix Fill to the main axis of the process connection dimension.

[0111] Furthermore, the central processing unit controls the set of reverse semantic recall clauses. The text semantic tags are parsed, and each text control procedure is precisely anchored and attached to the secondary branches corresponding to the four main axes. After the central processing unit completes the binding of all nodes and edge weights, it outputs a structured lesion mapping map. Structured lesion mapping atlas It enables parallel review of digital physical quantities and standard management contracts at the visualization level.

[0112] S55: Comparison of structured lesion mapping maps Within the dynamic evolution boundary and static benchmark, perform qualitative attribution and output the final root cause determination matrix. .

[0113] Specifically, the central processing unit maps the structured lesion atlas. Adaptive dynamic boundary threshold vector on the principal axis of the threshold dimension With the static initial reference constraint vector in step S1 Perform the comparison.

[0114] For testing the slow-varying drift faults common in industrial systems, due to the adaptive dynamic boundary threshold vector The adaptive shrinkage of the envelope is achieved in the adaptive dynamic boundary iteration of step S2. If the central processing unit calculates and finds the adaptive dynamic boundary threshold vector... The contraction amplitude exceeds the set deviation bandwidth, and the interpolated time series data matrix The non-mutation type of degradation was observed, and the central processing unit determined that the fundamental attribute of this abnormal event was a critical missed failure caused by a threshold setting deviation.

[0115] If the adaptive dynamic boundary threshold vector Relatively stable, while the set of reverse semantic recall control clauses The central processing unit determined that the root cause of the non-compliant operation text, such as insufficient material buffer delay, was a loophole in the process connection operation specification.

[0116] The central processing unit (CPU) uniformly formats and encapsulates the attribute identifiers, the feature matrices involved in the determination, and the attached text clause numbers to generate the final root cause determination result matrix. Final root cause determination matrix The complete quantitative and qualitative evidence chain of this anomaly was recorded and written to the read-only storage area to be read by the subsequent feedback adjustment engine.

[0117] S6: Based on the final root cause determination result, perform parameter feedback correction and procedure text revision with adaptive dead zone control, complete historical data replay and backtracking verification in the isolated sandbox memory area, and build a highly secure control clause-level closed-loop evolution engine.

[0118] In practical implementation, to address the technical bias in industrial control where real-time feedback from high-frequency fluctuating sensor time-series physical quantities to modify control procedures can easily lead to oscillations and divergences in control loop parameters, this step introduces a positive correlation proportional adaptive feedback algorithm with dead-zone limitations and a historical monitoring data replay backtesting mechanism. This stage transforms the qualitative and quantitative evidence chains output from step S5 into specific mathematical corrections and compensations. Only after deterministic mathematical closed-loop verification is completed in a local sandbox environment can the cascaded update of the production database be executed, thereby achieving highly secure adaptive iteration of standard operating procedure control clauses.

[0119] S61: Analyzing the final root cause determination result matrix The attribute identifier code is used to perform hardware routing distribution for the feedback optimization branch.

[0120] Specifically, the central processing unit reads the final root cause determination result matrix output in step S5. The central processing unit (CPU) processes the final root cause determination matrix at the kernel level. The attribute identifiers encapsulated in the code are subjected to bitwise AND logical operations to determine the fundamental attributes of the current event.

[0121] If the attribute identifier code is mapped to a threshold setting deviation that causes a critical missed alarm or false alarm trigger, the central processing unit activates the internal algorithm dynamic parameter adjustment bus and enters the control parameter adaptive correction branch. If the attribute identifier code maps to a loophole in the process connection operation specification or ambiguity in the control text, the central processing unit activates the system's standard operating procedure revision push queue and enters the management procedure text optimization branch.

[0122] S62: Based on the deviation residual magnitude and adaptive dead zone deviation constraints, execute the adaptive correction branch of control parameters and solve for the updated weight vector. With updating dynamic parameter sets .

[0123] In one feasible approach, to address the threshold setting deviation problem, the central processing unit must adjust the weighting constant coefficients in step S2 (weighted coefficients in the threshold dimension). (For example) Online correction is performed. To prevent frequent triggering of cascaded optimizations due to micro-parameter perturbations, the central processing unit introduces a dead-time control function.

[0124] The central processing unit (CPU) analyzes the final root cause determination matrix. Extract the interpolated time series data matrix from the attached evidence chain. The physical waveform extrema within, combined with the static initial reference constraint vector in step S11 Calculate the peak relative deviation scalar The central processing unit (CPU) uses a scalar based on the peak relative deviation. Solving for the weight coefficient feedback step size Its control formula is as follows: (7); in, The step size is used to calculate the weight coefficients. To adjust the convergence step size, a fixed damping constant is pre-configured based on the risk tolerance of the industrial site. To extract from the interpolated time series data matrix Extract the calculated peak relative deviation scalar; This is the adaptive dead zone deviation threshold.

[0125] When the peak relative deviation scalar The adaptive dead zone deviation threshold was not crossed. At that time, the step size of the weight coefficient feedback Forced to zero, the weighting coefficients remain absolutely static and hard-locked. When the threshold is crossed, the central processing unit feeds back the weighting coefficients to the step size. Accumulated to the original weight coefficient The algorithm updates the weights of other dimensions simultaneously while satisfying normalization constraints, and outputs an updated weight vector. The central processing unit will update the weight vector. The preset ratio of trigger conditions for synchronous fine-tuning based on the magnitude of missed / false alarms is uniformly encapsulated into an updated dynamic parameter set. It is pushed to the sandbox verification queue.

[0126] S63: Integrates high-density core text features with reverse semantic recall content, executes management procedure text optimization branches, and generates structured revision push vectors. .

[0127] In practice, if the management procedure text optimization branch is entered, the central processing unit extracts the set of reverse semantic recall control clauses generated in step S5. and the high-density core text features output in step S3 The central processing unit initiates a natural language generation model to retrieve the set of control clauses through reverse semantic recall. Based on the modified object, high-density core text features are used. The specific physical phenomenon primitives included are used as supplementary entities to perform text concatenation and grammatical reconstruction. Preferably, the specific physical phenomenon primitive can be set to a conveyor belt delay preheating time of not less than 120 seconds.

[0128] The central processing unit will reconstruct the revised revision proposal text and the relevant process codes. The original control clause's globally unique identifier (GUID) is serialized and encoded to generate a structured revision push vector. Structured revision push vector As text control instructions awaiting activation, they are also pushed to the sandbox verification queue to await auditing.

[0129] S64: Import historical test data packages into the isolated sandbox memory area and utilize the updated dynamic parameter set. Or structured revision push vector Perform a historical monitoring data replay backtest and output a backtest verification pass flag. .

[0130] Preferably, to eliminate logical oscillations in the production system caused by parameter modifications, the central processing unit (CPU) allocates an isolated sandbox memory area in physical memory that is unidirectionally and physically isolated from the main control bus. The central processing unit extracts in batches historical operating condition data snapshots containing characteristics of missed and false alarm events from the relational database over the past 30 days, as well as the trigger transient event dataset generated in step S2. In the isolated sandbox memory area Internally reassembled into historical test data packets .

[0131] The central processing unit is in the isolated sandbox memory area The comprehensive degradation index calculation engine is instantiated in step S2. The central processing unit will update the dynamic parameter set. Inject this virtual computing engine into historical test data packets. Perform a complete high-speed timing replay operation. The central processing unit verifies the replay operation results: if the new parameter set successfully triggers the historical test data packet... If a missed event is marked in the middle and does not generate new false alarms or triggers on the normal waveform, then the adaptive correction logic for the control parameters is considered to have converged; similarly, for the structured revision push vector... The system verifies whether it has disrupted the original equipment process topology dependencies.

[0132] The CPU will mark a Boolean-type backtracking test verification pass if and only if the replay backtracking test within the sandbox fully meets the safety convergence criteria. If a divergence or conflict occurs, the value is assigned to logical true (True); if a divergence or conflict occurs, the value is assigned to logical false (False), and a termination rollback interrupt signal is sent to the central processing unit main control module.

[0133] S65: Based on the backtracking test verification results, perform a cascading mapping refresh of the database, physically replace the underlying static baseline and semantic base, and generate a closed-loop update log. .

[0134] In one feasible approach, the central processing unit continuously monitors the isolated sandbox memory region. The output backtracking test verification passed the flag. When the backtracking test verification is detected, the flag is displayed. When the condition is logically true, the central processing unit activates the database write lock and enables global transaction management.

[0135] If the current operation is parameter correction, the CPU will update the dynamic parameter set. The static initial reference constraint vector stored in the physical coverage step S11 And the weight configuration file in the preset algorithm module; If the current operation is text revision, the CPU will push the structured revision vector. The text tensors in the text are remapped to the manifold space, and the corresponding exogenous text modality initial feature set in step S15 is physically replaced. The original vector entries in the semantic base are used to complete the hard update of the semantic base.

[0136] Finally, the central processing unit captures the return status codes of all the database table structure change operations, the parameter mirror images before and after the modification, and the confirmation timestamps. It then packages these data using a cryptographic hash algorithm to generate an immutable closed-loop update log. Closed-loop update log The data is persistently stored in the system audit storage node. At this point, the entire process of adaptive closed-loop iterative self-evolution of standard operating procedure control clauses driven by online physical waveform timing fact snapshots is completed, and the system returns to steady-state monitoring mode.

[0137] Example 2: Figure 2 This application provides a schematic diagram of the structure of an adverse event analysis system based on a SOP system. The adverse event analysis system based on a SOP system includes a SOP basic data preset module, a three-dimensional trigger model construction and triggering module, a step-by-step reporting management module, a cross-system data automatic association module, a fishbone diagram root cause localization module, and a trigger rule and SOP closed-loop optimization module.

[0138] The SOP basic data preset module is responsible for presetting the SOP basic data required for adverse event analysis in the system, and completing the binding and storage of data with processes, providing basic data support for subsequent modules.

[0139] In one implementation, the SOP basic data preset module may include an application processor, a system memory bus, a non-volatile storage unit, and a high-speed data interface.

[0140] In this application, the SOP basic data preset module is used to preset the standard operating procedure (SOP) basic data matrix required for adverse event analysis in the business relationship database, establish the underlying static topology mapping between production equipment, process nodes, and control limits, and complete the baseline parameter initialization of dynamic boundary tracking and semantic mapping space. The SOP basic data preset module extracts the SOP quality control requirements and equipment operation standards for each production process and establishes a static initial baseline limit vector. .

[0141] The SOP basic data preset module clearly defines the lower limit value of product qualification rate and the benchmark value of single equipment downtime, and limits each static initial benchmark vector. Corresponding process code Perform unique memory address binding. The SOP basic data preset module organizes all production equipment required for each production process and constructs a structured list of related equipment interactions. The SOP basic data preset module is in the associated device interaction list. Record the equipment model of each production machine , To which process node Equipment Management Identification and equipment operating parameter acquisition port information .

[0142] The SOP basic data preset module will associate the device interaction list. Each piece of production equipment is associated with a corresponding process code. The static initial reference constraint vector that has been bound to this process. Perform 3D primary and foreign key association binding to generate a device process threshold mapping table. The SOP basic data preset module maps equipment process thresholds to a table. Persistently store the data in the business relational database.

[0143] The SOP basic data preset module quantifies the impact weight of each production process on the final product quality and the constraint stiffness on the advancement of subsequent processes, and sets and maps process priority coefficients. The SOP basic data preset module will set the process priority coefficient. The process is divided into three discrete score levels, and the priority coefficient of each process is assigned accordingly. Corresponding process code One-to-one mapping.

[0144] The SOP basic data preset module initializes the data distribution of the long-period baseline state for the adaptive dynamic boundary triggering mechanism, and constructs the long-period data feature matrix. The SOP basic data preset module retrieves a long-term historical sequence of physical parameters under continuous, error-free operating conditions. Calculate the historical sequence of long-period physical parameters Long-period mean vector and long-period variance matrix Together, they constitute the initial long-period data feature matrix. .

[0145] The SOP basic data preset module initializes the exogenous semantic base of the standard operating procedure (SOP) text for the cross-modal semantic automatic mapping mechanism, and reads all SOP text control clauses stored in the relational database. The SOP basic data preset module utilizes a preset text feature extraction and processing unit to process the standard operating procedure text control clauses. Perform word segmentation to generate an initial feature set of exogenous text modalities. .

[0146] The SOP basic data preset module calculates the static initial baseline constraint vector. List of associated device interactions Equipment process threshold mapping table Process priority coefficient Long-period data feature matrix and the initial feature set of exogenous text modalities Transmitted to the 3D trigger model construction and triggering module.

[0147] In one implementation, taking the example of a pharmaceutical company's injection production workshop that needs to strictly follow Good Manufacturing Practices (GMP), the SOP basic data pre-setting module defines static initial baseline constraint vectors for different types of processes. .

[0148] Preferably, the solution preparation process is configured with a process code as a quality-related node. For Solution Preparation-01, the SOP basic data preset module is set to a single solution preparation failure rate of no more than 2%. A single solution preparation failure rate of no more than 2% corresponds to a minimum pass rate of 98%.

[0149] Preferably, the light inspection process is configured with a process code as a quality-related node. For the light inspection-03, the SOP basic data preset module sets the lower limit of the pass rate to 99.5%.

[0150] Preferably, the filling process is configured as an equipment node with a process code. For Filling-02, the SOP basic data preset module sets the baseline downtime for a single operation of the filling equipment to 1 hour. For the dispensing tank, as an equipment node, the SOP basic data preset module sets the baseline downtime to 1.5 hours.

[0151] All static initial baseline bounding vectors The specific values ​​are all related to the corresponding process codes. Unique binding. The SOP basic data preset module assigns process priority coefficients based on the impact of each process on product quality and subsequent processes. The filling process is configured with a priority level of 1, corresponding to a quantification value of 3. This process directly determines the accuracy of the injection dosage; if it is interrupted, the visual inspection and packaging processes will both be halted. The solution preparation process is configured with a priority level of 2, corresponding to a quantification value of 2. This process affects the concentration of the active ingredient; if it is delayed, the filling process will be waiting for materials. The visual inspection process is configured with a priority level of 3, corresponding to a quantification value of 1. This process only affects the identification of visual defects; if an abnormality occurs, subsequent processes can be ensured by temporarily adding manual re-inspection. The SOP basic data preset module organizes the interaction list of related equipment according to the process. And complete the binding. The filling process is associated with the first filling equipment and the first conveyor belt.

[0152] Preferably, the first filling equipment is model GZ-200, and the equipment management identification is... SY-003, device operating parameter acquisition port information. It is COM5. The first conveyor belt is model SD-300, equipment management identifier. It is SY-004.

[0153] The SOP basic data preset module binds the equipment management identifier SY-003 with the filling-02 process and the 1-hour downtime threshold. The liquid preparation process is associated with the first liquid preparation tank.

[0154] Preferably, the first liquid preparation tank is model PLG-500, and the equipment management identification is... It is SY-001.

[0155] The SOP basic data preset module binds the equipment management identifier SY-001 with the solution preparation-01 process and the 98% pass rate threshold. All binding relationships between equipment, processes, and thresholds are stored in the equipment-process-threshold mapping table. .

[0156] The 3D trigger model construction and triggering module is used to construct a 3D trigger model based on basic parameters, which includes threshold, trend change, and process weight. It calculates dynamic trigger values ​​in real time, judges trigger conditions according to preset logic, and automatically triggers the adverse event analysis process and stores the trigger data if the conditions are met, providing a trigger basis for root cause analysis.

[0157] In one implementation, the 3D trigger model construction and triggering module may include a central processing unit, a coprocessor, an industrial bus interface unit, and a data comparator.

[0158] In this application, the 3D trigger model construction and triggering module is used to continuously parse multi-dimensional process monitoring data streams online through the underlying hardware interface based on the adaptive dynamic threshold and trend joint tracking mechanism, and to calculate the adaptive dynamic boundary threshold vector in real time. With the overall degradation index It performs time-series divergence measurement on micro-critical risks, determines and executes adaptive triggering of adverse event analysis processes.

[0159] The 3D trigger model construction and triggering module retrieves the timing parameters of physical sensors at each process node in real time through the workshop data acquisition layer. It then performs spatiotemporal alignment and noise reduction normalization on the multi-dimensional physical sensor parameters, assembling them to generate the time step. Step-by-step process monitoring data vector The 3D trigger model is constructed, and the trigger module calls the long-period data feature matrix. An unsupervised prediction model is established based on a time-series sliding window.

[0160] The 3D trigger model construction and trigger module read historical process monitoring data within a dynamic time-series sliding window and combine it with long-period mean vectors. With long-period variance matrix Calculate the state evolution probability density and solve for the prediction expectation vector. The three-dimensional triggering model is constructed and the triggering module compares the measured values ​​of time-series facts with the predicted values ​​of unsupervised theory to measure the micro-transient deviation characteristics of actual working conditions and generate a deviation residual matrix. .

[0161] 3D trigger model construction and trigger module for obtaining process monitoring data vectors With the predicted expected vector The transient deviation residuals are then subjected to an outer product matrix operation with their transpose vectors. The 3D trigger model construction and trigger module are combined with the static initial baseline constraint vector. Based on the distribution of nearest-neighbor time-series waveforms, adaptive dynamic boundary iteration is performed, and a degradation trend negative feedback damping mechanism is introduced to solve the adaptive dynamic boundary threshold vector in real time. .

[0162] 3D trigger model construction and trigger module fusion deviation residual matrix Adaptive dynamic boundary threshold vector With process priority coefficient The comprehensive degradation index is calculated by multidimensional hybrid linear weighting and nonlinear compression function. 3D trigger model construction and trigger module call deviation from residual matrix Extract trace terms and recall adaptive dynamic boundary threshold vectors Perform norm measurement, using the hyperbolic tangent function to evaluate the process priority coefficients. Perform nonlinear continuous mapping.

[0163] The construction of the 3D trigger model and the trigger module for the comprehensive degradation index Perform a step-condition check control chain audit and adjust the overall degradation index. Continuous comparison is performed with the preset safety threshold value.

[0164] When a risk threshold is crossed, the 3D trigger model construction and triggering module sends a hardware interrupt command to the physical bus actuator, toggling the hardware trigger level, blocking bus data, and latching the trigger transient event dataset. .

[0165] The 3D trigger model construction and trigger module will vectorize historical process monitoring data. Sequence, Deviation Residual Matrix Snapshot and Adaptive Dynamic Boundary Threshold Vector Snapshots are packaged at the snapshot level to generate datasets that trigger transient events. The specific calculation process is as follows: ; ; The 3D trigger model construction and trigger module will calculate the dataset of transient trigger events. Transmitted to the step-by-step reporting management module.

[0166] In one implementation, the 3D trigger model construction and triggering module constructs a 3D trigger model based on basic parameters, comprising a threshold dimension, a trend change dimension, and a process weight dimension. The threshold dimension is defined by a preset static initial baseline vector. This serves as the calculation benchmark. The trend change dimension is based on a production rhythm of 2 hours per batch for the injection solution, with a statistical period of 30 minutes per unit. The process weight dimension directly calls the preset process priority coefficient. .

[0167] The 3D trigger model construction and trigger module uses a single-dimensional order reduction formula to calculate the dynamic trigger value constant. The specific calculation process is as follows: .

[0168] Based on the key points of drug quality risk control, the three-dimensional trigger model is constructed, and the first weight coefficient corresponding to the preset threshold dimension of the trigger module is determined. The second weighting coefficient is 0.4, corresponding to the trend change dimension. The value is 0.3, which is the third weighting coefficient corresponding to the process weighting dimension. It is 0.3.

[0169] At 9:00 AM on a certain production day, the 3D trigger model construction and trigger module obtains the actual downtime of the filling equipment in real time through a preset interface. The time is 1.2 hours, and the static initial baseline constraint vector is called. numerical values The time interval is 1 hour, and the calculated relative deviation ratio is 0.2.

[0170] The 3D trigger model construction and trigger module obtained downtime data for the past three statistical periods, showing an upward trend from 0.5 hours to 0.8 hours to 1.2 hours. The number of upward statistical periods was calculated. Equals 3, total number of cycles The trend ratio equals 3, and the trend ratio equals 1. The priority coefficient for the 3D trigger model construction and trigger module call process is also considered. numerical values The value is 3.

[0171] The 3D trigger model construction and trigger module substitutes the above data into the 1D reduction formula to calculate the dynamic trigger value constant. It equals 1.28. The 3D trigger model construction and trigger module verify trigger conditions according to the logic of threshold priority, trend assistance, and weighted fallback.

[0172] The three-dimensional trigger model construction and trigger module first determine whether the relative deviation ratio of 0.2 reaches the preset risk tolerance ratio of 20% for the filling process, and the result meets the requirement.

[0173] The three-dimensional trigger model construction and trigger module further determine the number of rising statistical periods within a continuous statistical period. Is it equal to the total number of cycles? The result satisfies the requirements. The three-dimensional trigger model is constructed, and the trigger module ultimately determines the process priority coefficient. Does the value 3 reach the preset weight threshold of the Level 1 process? The result is satisfactory.

[0174] Once all three conditions are met, the 3D trigger model construction and trigger module automatically triggers the adverse event analysis process, and simultaneously sets the dynamic trigger value constant. The corresponding basic data and monitoring data sequences are associated and stored in the dataset that triggers transient events. .

[0175] The step-by-step reporting management module is used to divide the reporting process into two stages: initial reporting and supplementary improvement. It designs corresponding forms and generates unique event identifiers, verifies the completeness of fields and data matching at each stage, ensures that the reported data is timely and accurate, and provides identifiers and structured data for data association.

[0176] In one implementation, the step-by-step reporting management module may include a microprocessor, a clock generator, a display driver, and a human-computer interaction touch interface.

[0177] In this application, the step-by-step reporting management module is used for two-stage asynchronous qualitative reporting and multimodal content condensation and reconstruction based on trigger event data, generating a unique event identifier primary key containing spatiotemporal process constraints. And perform cascaded hard lock error prevention verification.

[0178] The step-by-step reporting management module detected a dataset that triggered transient events. After writing to the flash memory, the simplified initial reporting process of the front-end industrial tablet is activated to obtain the initial human interaction parameters.

[0179] The step-by-step reporting management module automatically retrieves the timestamp of the current event. Physical location coordinates of the event and receive initial manual qualitative text. .

[0180] The step-by-step reporting management module initiates a cross-perspective text fusion and content condensation and reconstruction mechanism for the initial manually qualitative text. Perform mutual information minimization mapping.

[0181] The step-by-step reporting management module calls the initial feature set of exogenous text modalities. The initial manual qualitative text was stripped away using a mutual information minimization strategy. Solve for and output high-density core text features by analyzing subjective invalid expressions in the text. The step-by-step reporting management module extracts the underlying process semantics based on location mapping routing rules and combines them with time variables to generate a unique event identifier primary key with embedded spatiotemporal process constraints. .

[0182] The step-by-step reporting management module calls the physical location coordinates of the event. Equipment process threshold mapping relationship table Initiate a spatial topology addressing query and retrieve the corresponding process code in reverse. The step-by-step reporting management module uses a string hash concatenation algorithm to encode the process. Event timestamp And concatenate secure random sequence codes.

[0183] The step-by-step reporting management module instantiates a dual-closed-loop sequential state machine. Based on the unique event identifier primary key Execution process control of state transitions. Step-by-step reporting management module to dual closed-loop sequential state machine. Inject jump instruction to start supplementary timer .

[0184] The step-by-step reporting management module uses a unique event identifier primary key. The system drives a database primary and foreign key cascading trigger, dynamically pruning and cascading hardware locking to prevent errors in structured interactive commands input at secondary levels. The step-by-step reporting management module receives the operator-selected device identifier. Operator selected procedure number The step-by-step reporting management module extracts the process code. List of interactions with associated devices Perform dynamic view clipping to generate a subset of cascading constraint devices. .

[0185] The step-by-step reporting management module determines which device identifier the operator has selected. Does it belong to a subset of cascaded constraint devices? If the judgment result is logical false, the hardware latch-up level signal is output to the interface. Forcefully block data entry into the database. The step-by-step reporting management module assembles multimodal data within a countdown constraint and outputs a two-level structured reporting dataset. The step-by-step reporting management module will uniquely identify the primary key of the event. High-density core text features Operator selects equipment identifier Operator selects procedure number and the dataset that triggers transient events Unified encapsulation, the specific calculation process is as follows: .

[0186] The step-by-step reporting management module will calculate the two-level structured reporting dataset. Transmitted to the cross-system data automatic association module.

[0187] In one implementation, the phased reporting management module initiates the phased reporting process according to the principle of responding first and then improving, balancing the timeliness of adverse event response with data accuracy. The phased reporting management module configures a simple form in the initial reporting phase, including the event's timestamp. Physical location coordinates of the event Initial manual qualitative text Three core fields. After the initial form submission, the step-by-step reporting management module automatically generates a unique event identifier primary key. .

[0188] Preferably, a unique event identifier primary key The value is GZ-02-202405200900-123456. This identifier consists of the process code GZ-02, a timestamp accurate to the minute, and a 6-digit random sequence.

[0189] The step-by-step reporting management module will use a unique event identifier primary key. Events are stored in the event database in real time and displayed synchronously on the user interface. The phased reporting management module will allow for the addition of extended fields during the supplementation and improvement phase.

[0190] Preferably, the operator selects the procedure number. The drop-down menu only displays sop-GZ-02-001 corresponding to the filling-02 process.

[0191] Preferably, the operator selects the device identifier. The dropdown menu only displays the equipment process threshold mapping table. SY-003 and SY-004 are associated with Filling-02.

[0192] The step-by-step reporting management module features a timeline-style interface, displaying three stages at the top: initial reporting completion, supplementary information entry, and information verification passed. The module automatically activates the supplementary information entry stage after the initial report is submitted, with a 1-hour supplementary information timer displayed on the right side of the interface. Time Limit. After the fields are filled in, the step-by-step reporting management module will automatically perform verification. The verification operator selects the procedure number. With unique event identifier primary key After verifying whether the process code GZ-02 matches, and both checks pass, activate the information verification passed node and generate a two-level structured reporting dataset. .

[0193] The cross-system data automatic association module is used to parse unique event identifiers, retrieve corresponding data from the equipment management system and production execution system, fill in the missing fields in the report form, verify the consistency between the data and the basic data, and integrate them to form a complete adverse event dataset.

[0194] In one implementation, the cross-system data auto-association module may include an application processor, a protocol translation socket controller, a direct memory access engine, and a database connection pool.

[0195] In this application, the cross-system automatic data association module is used to perform spatiotemporal adaptive data routing and underlying microscopic physical parameter data interpolation across heterogeneous industrial buses based on unique event identifier primary keys, implement multidimensional strong consistency auditing, and assemble high-fidelity multimodal event datasets. The cross-system data automatic association module parses unique event identifier primary keys. The semantic hash bits are used to reverse extract spatial topology nodes and temporal cross-sectional parameters, and the data retrieval time window is dynamically configured. .

[0196] The cross-system data automatic association module uses unique event identifiers for primary keys. Perform reverse hashing of the string to extract the process code. Timestamp of the event The cross-system automatic data association module extends bidirectionally to both historical and future time domains, constructing an adaptively aligned data retrieval time window. Cross-system automatic data association module for joint process coding. With the operator selecting the device identifier Perform primary and foreign key cascading addressing in the relational database to generate a list of data retrieval requirements for heterogeneous systems. .

[0197] The cross-system data automatic association module utilizes process coding In the equipment process threshold mapping table The system initiates an inner join query, packages the successfully matched underlying programmable logic controller register addresses with the manufacturing execution system production batch numbers to form a heterogeneous system data retrieval request list. The cross-system data automatic association module activates protocol conversion sockets to retrieve data from heterogeneous systems based on a data retrieval requirement list. Perform seamless cross-system handshakes and hard real-time data acquisition on industrial buses to obtain heterogeneous timing raw data. .

[0198] The cross-system automatic data association module executes a nearest-neighbor time-series mean data interpolation algorithm based on the process mainline topology for the two-order structured reporting dataset. Deterministic physical feature inverse injection is performed on the missing form fields to generate an interpolated time series data matrix. The cross-system data automatic association module calculates the waveform derivative and deviation slope of the physical waveform curve, and forcibly injects deterministic feature values ​​into the missing fields to solve for the interpolated physical scalar. The cross-system data automatic association module establishes a three-level linkage strong consistency boundary audit mechanism to compare the interpolated time series data matrix. With static topological constraints.

[0199] The cross-system automatic data association module performs Level 1 hardware identifier auditing, Level 2 production batch auditing, and Level 3 tolerance range auditing. After verification, the module assembles and generates a high-fidelity multimodal event dataset. The cross-system data automatic correlation module will interpolate the time-series data matrix. High-density core text features Operator selects procedure number and the dataset that triggers transient events The structured assembly process involves the following calculations: .

[0200] The cross-system data automatic association module will calculate the high-fidelity multimodal event dataset. Transmitted to the fishbone diagram root cause localization module.

[0201] In one implementation, the cross-system data auto-association module starts the event identifier parsing engine to extract the unique event identifier primary key. The process code GZ-02 in the code determines the process to which the adverse event belongs, and the event occurrence timestamp is extracted. The data retrieval time window is set to 202405200900 and combined with the standard production cycle of 2 hours per batch for the filling process. Extended to 08:00 to 10:00 on May 20, 2024.

[0202] The cross-system data automatic association module maps equipment process thresholds to process codes GZ-02. Retrieve the interaction list of associated devices This will create a list of data retrieval requirements for heterogeneous systems. The cross-system data automatic association module retrieves data from heterogeneous systems according to the OPCUA industry standard protocol, based on the required list. Transform it into a request command that the interface can recognize.

[0203] The cross-system automatic data association module is configured with a data receiving caching mechanism to classify and store the SY-003 runtime data and production execution system data returned by the interface to obtain heterogeneous time-series raw data. .

[0204] Preferably, the acquired SY-003 operating data includes the filling pressure of 0.3 MPa at 9:00 and the fault alarm E01.

[0205] Preferably, the acquired production execution system data includes an average filling volume of 10.1 ml from 8:30 to 9:00.

[0206] The cross-system data automatic association module identifies missing form fields from heterogeneous time-series raw data. Extracting fault alarm E01 and solving interpolated physical scalars Fill the equipment fault code field. The cross-system data automatic association module calculates the filling volume deviation value from the production data. The value is 10.1 minus 10 equals 0.1 ml, which is then filled into the filling volume deviation value field.

[0207] The cross-system data automatic association module performs a consistency check, verifying the management identifiers SY-003 and SY-004 in the device operation data against the associated device interaction list. To verify a complete match, check if the process code in the production data is consistent with GZ-02, and verify if the filling pressure of SY-003 (0.3 MPa) is within the static initial reference limit vector. Within a reasonable fluctuation range. A high-fidelity multimodal event dataset is generated after all validation items pass. .

[0208] The fishbone diagram root cause localization module is used to retrieve trigger rule parameters and datasets, construct fishbone diagram analysis dimensions and generate fishbone diagrams, mark related SOP clauses, investigate threshold setting deviations and process connection loopholes, and locate the root cause of adverse events.

[0209] In one implementation, the fishbone diagram root cause localization module may include a graphics processor, a tensor operation core, a video memory unit, and a parallel computing stream processor. In this application, the fishbone diagram root cause localization module is used based on a high-fidelity multimodal event dataset. Perform cross-modal semantic automatic mapping and content condensation reconstruction to construct a structured 4D causal topology primitive, achieving high-precision semantic alignment and root cause determination of the underlying physical time series and management procedure texts. The fishbone diagram root cause localization module performs high-fidelity multimodal event dataset analysis. Perform data unpacking and modal separation to extract the interpolated time series data matrix. High-density core text features and the dataset that triggers transient events The fishbone diagram root cause localization module performs temporal manifold feature encoding on the underlying physical parameters, generating a temporal manifold feature matrix. .

[0210] Fishbone diagram root cause localization module combined with high-density core text features Initial feature set of exogenous text modal Perform mutual information minimization and content condensation reconstruction to solve the condensed cross-modal feature matrix. The fishbone diagram root cause localization module calculates and optimizes mutual information condensation to reconstruct the constraint objective function. Filter out artificial interference noise.

[0211] The fishbone diagram root cause localization module computes the condensed cross-modal feature matrix in the same semantic manifold space. Initial feature set of exogenous text modal Feature distance, generating semantic distance matrix And execute the set of reverse semantic recall control clauses. The fishbone diagram root cause localization module filters all candidates whose feature distance is less than the semantic recall cutoff constant.

[0212] The fishbone diagram root cause localization module constructs a 4D causal topology primitive skeleton, which integrates the set of reverse semantic recall control clauses. And trigger transient parameters to be loaded into multidimensional branches to generate structured lesion mapping atlases. The fishbone diagram root cause localization module will adaptively adapt the dynamic boundary threshold vector. Fill to the threshold dimension principal axis, deviating from the residual matrix. Fill to the trend dimension main axis, process priority coefficient Fill the interpolated time series data matrix to the main axis of the process weight dimension. Fill to the main axis of the process connection dimension.

[0213] Fishbone diagram root cause localization module compares structured lesion mapping atlas Within the dynamic evolution boundary and static benchmark, perform qualitative attribution and output the final root cause determination matrix. .

[0214] The fishbone diagram root cause localization module uniformly formats and encapsulates the determined attribute identifier code, the feature matrix involved in the determination, and the attached text clause number. The specific calculation process is as follows: .

[0215] The fishbone diagram root cause localization module calculates the final root cause determination result matrix. Transmitted to the triggering rules and SOP closed-loop optimization module.

[0216] In one implementation, the fishbone diagram root cause localization module uses fishbone diagram analysis to locate the root cause of adverse events. The module determines the four main analytical dimensions of the fishbone diagram and attaches a set of reverse semantic recall control clauses to the corresponding branches of each main axis. .

[0217] The fishbone diagram root cause localization module labels the static initial baseline constraint vector in the threshold dimension. The corresponding clause 5.2 of sop-GZ-02-001.

[0218] Preferably, the text content of Article 5.2 is that if the filling equipment is shut down for more than 1 hour at a time, an abnormality handling procedure must be initiated.

[0219] The fishbone diagram root cause localization module marks the associated process nodes in SOP-GZ-02-001 item 7.3 in the process connection dimension.

[0220] Preferably, the text of Clause 7.3 states that after the filling process is completed, the conveyor belt must transport the product to the light inspection process within 10 minutes.

[0221] The fishbone diagram root cause positioning module verifies data in each dimension, comparing the monitored data exceeding the 0.2% threshold with the static threshold set for 1 hour. No material deterioration was observed during this 1.2-hour downtime, and subsequent product testing showed a 99.2% pass rate.

[0222] In the process connection dimension, the fishbone diagram root cause localization module analysis revealed that the delayed start of the conveyor belt was due to the operator's failure to preheat in a timely manner. Combining the trend dimension (although downtime increased, it did not affect quality) and the process weight dimension (priority level 1, requiring caution but not overly sensitive data), the fishbone diagram root cause localization module ultimately determined the final root cause judgment matrix for the adverse event. Set a deviation for the static threshold of the filling equipment.

[0223] The fishbone diagram root cause localization module will generate a matrix of the final root cause determination results. Associated storage.

[0224] The triggering rules and SOP closed-loop optimization module is used to adjust the triggering algorithm parameters according to the root cause and generate SOP revision suggestions, verify the optimization results, and update the basic data after passing the verification, forming an analysis-optimization-update closed loop to improve the system's subsequent management and control capabilities.

[0225] In one implementation, the triggering rules and SOP closed-loop optimization module may include a baseband processor, an isolated sandbox physical memory area, a transaction manager, and a flash controller.

[0226] In this application, the triggering rules and the SOP closed-loop optimization module are used to optimize the final root cause determination result matrix. Perform parameter feedback correction and procedure text revision with adaptive dead-time control, and complete historical data replay and backtracking verification in the isolated sandbox memory area.

[0227] Triggering rules and SOP closed-loop optimization module analyze the final root cause determination result matrix The attribute identifier code is used to perform hardware routing distribution for the feedback optimization branch. The triggering rule and SOP closed-loop optimization module, based on the deviation residual magnitude and adaptive dead zone deviation constraints, executes an adaptive correction branch for control parameters to solve for and update the weight vector. With updating dynamic parameter sets .

[0228] Triggering rules and SOP closed-loop optimization module calculate peak relative deviation scalar Based on the peak relative deviation scalar Solving for the weight coefficient feedback step size Triggering rules and the SOP closed-loop optimization module integrate high-density core text features. With the set of reverse semantic recall control clauses Optimize the execution management procedure text branch and generate a structured revision push vector. .

[0229] Triggering rules and SOP closed-loop optimization module in isolated sandbox memory area Import historical test data packages The triggering rules and SOP closed-loop optimization module utilizes the updated dynamic parameter set. Or structured revision push vector Perform a historical monitoring data replay backtest and output a backtest verification pass flag. .

[0230] The triggering rules and SOP closed-loop optimization module executes a database cascading mapping refresh based on the backtracking test verification results, physically replacing the underlying static baseline and semantic base, and generating a closed-loop update log. The specific calculation process is as follows: .

[0231] In one implementation, the triggering rules and the SOP closed-loop optimization module initiate the closed-loop optimization process. The triggering rules and the SOP closed-loop optimization module extract the root cause of the threshold setting deviation and correct the dynamic triggering algorithm parameters of the one-dimensional reduction formula.

[0232] The triggering rules and SOP closed-loop optimization module will use the first weight coefficient The value was increased from 0.4 to 0.5, with the increase being a scalar of the relative deviation from the peak value. Positive correlation. The triggering rule and SOP closed-loop optimization module increases the preset ratio of triggering conditions from 0.2 to 0.3.

[0233] The trigger rules and SOP closed-loop optimization module execution management procedure text optimization branch were revised, and section 5.2 of SOP-GZ-02-001 was revised to generate a structured revision push vector. The downtime threshold for filling equipment was adjusted from 1 hour to 1.3 hours. The trigger rule and SOP closed-loop optimization module were activated, and the comprehensive verification module was launched to verify the optimization results, incorporating historical test data packets. Based on historical monitoring data, the dynamic trigger value constant of the corrected one-dimensional order reduction formula is calculated. .

[0234] Preferably, the calculation process is 0.2 multiplied by 0.5 plus 1 multiplied by 0.3 plus 3 multiplied by 0.3, and the value is equal to 1.3.

[0235] At this point, the relative deviation ratio of 0.2 is less than 0.3, which does not meet the triggering condition. The triggering rule and the backtracking test verification of the SOP closed-loop optimization module output pass the checkmark. If the logic is true, the false triggering event is excluded. The triggering rules and SOP closed-loop optimization module verify that the operator selected the procedure number. The field's dropdown options are updated to the revised structured revision push vector. Once the triggering rules and the SOP closed-loop optimization module confirm the optimization is effective, the standard operating procedure (SOP) basic database is updated synchronously, and a closed-loop update log is generated. .

[0236] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the adverse event analysis system based on the SOP system. In other embodiments of this application, the adverse event analysis system based on the SOP system may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0237] Example 3: This embodiment describes the mathematical derivation process in an adverse event analysis system and method based on a SOP system disclosed in Embodiments 1 and 2. The system constructed in this embodiment operates under a model inference and industrial control application architecture. The inference input data specification in this embodiment includes two independent heterogeneous data streams. The first input is a multi-channel quantitative process timing parameter stream acquired in real time through an industrial bus interface; the second input is an unstructured qualitative manual declaration text received through a front-end human-machine interaction touch interface.

[0238] In practice, the input features corresponding to the quantitative process time sequence parameter flow contain four physical dimensions, which are assembled into a multi-dimensional feature column vector.

[0239] Preferably, the multidimensional feature column vector includes the instantaneous operating pressure of the filling transmitter, the sterile flow rate of pipeline cleaning, the environmental temperature and humidity variation values, and the cumulative downtime of the controlled hardware equipment.

[0240] The sensor acquisition card has a fixed hard real-time sampling period of 50 milliseconds, corresponding to a sampling frequency of 20 Hz. The quantitative input data block within a single inference execution window is constructed into a shape... The temporal feature tensor, where Channel dimensions represent quantitative characteristics. This represents the length of the dynamic time-series sliding window, corresponding to a transient physical quantity control matrix for a continuous online observation time domain of 180 seconds. The raw data format for qualitative manual declaration text input is an unstructured string.

[0241] Preferably, the qualitative manual declaration text input is forcibly limited to a maximum character sequence length of 128 tokens through a front-end text input mask. Any part exceeding this length is automatically discarded by the truncation unit, and any part below this length is aligned to the high-dimensional boundary by fixed-length zero-padding units.

[0242] Before the central processing unit (CPU) executes the core inference operations, the system's data cleaning layer performs high-fidelity real-time data preprocessing logic to ensure a perfect match between the inference input format and the fixed parameter matrix. For the input time-series feature tensor, the preprocessing module calls the pre-stored global runtime history boundary parameters in the system configuration layer to perform minimax normalization calculations, linearly mapping the raw floating-point values ​​of all channels' sensors to open intervals. Within the module, absolute amplitude interference between different physical dimensions is eliminated. For the input qualitative manual declaration text string, the preprocessing module first converts the text string to all lowercase characters, then calls the WordPiece word segmenter, which is strictly consistent with the fixed model vocabulary, to perform discrete word segmentation, converting the text string into a format... Integer indexed sequence tensor.

[0243] To ensure that those skilled in the art can fully reproduce and run this system without creative effort, this embodiment completely decrypts and transparently discloses the internally fixed algorithm topology. The internally fixed models include an unsupervised dynamic boundary anomaly detection network and a cross-modal temporal text semantic alignment network (MindTS architecture). The exogenous semantic processing backbone network for handling text modalities adopts the publicly released lightweight DistilBERT pre-trained model architecture, specifically the distilbert-base-uncased version from the Hugging Face open-source model library, with fixed weights in ONNX format. This exogenous semantic processing backbone network contains 6 transformer layers and 12 self-attention mechanism heads, with a hidden layer semantic dimension width of 768 dimensions. After the exogenous text modality is output through the pooling layer of this backbone network, a shape is generated. The high-dimensional text feature vectors are obtained. To connect the semantic dimension of the text with the temporal feature dimension in the same dimension, a linear fully connected projection matrix layer is cascaded at the end of the exogenous text network. This projection matrix layer consists of a dense weight projection matrix. With a one-dimensional static bias vector Composition. Dense weighted projection matrix It has a rigorous matrix topological definition, and the specific matrix format expression is as follows: ; The dense weight projection matrix With bias vector Linearly map the 768-dimensional text semantic vector to a dimension of... In the shared manifold semantic space. The quantitative temporal feature processing backbone network adopts a multi-channel one-dimensional temporal convolutional network (TCN) architecture, containing four cascaded residual convolutional blocks. Each residual convolutional block contains a dilated convolutional layer, with the kernel sizes of the dilated convolutional layers fixed at 3, 5, 7, and 9, and the dilation factors set at 1, 2, 4, and 8. The one-dimensional temporal convolutional network also uses a shape... The time alignment layer will integrate the temporal manifold feature matrix. Mapping to the same dimension In the shared manifold semantic space.

[0244] During the online inference and root cause determination execution phase, the central processing unit (CPU) continuously runs the algorithm control chain on hardware computing nodes with a clock speed of at least 3.0 GHz. The CPU invokes unsupervised prediction models to process the real-time input process monitoring data vectors. Perform Gaussian mixture distribution likelihood estimation to calculate the expected prediction vector. The deviation residual matrix is ​​calculated by combining the outer products. The central processing unit (CPU) utilizes an adaptive dynamic boundary iteration formula to perform logarithmic cumulative integration on the variance of the deviation residuals within a dynamic time-series sliding window, thereby shrinking the adaptive dynamic boundary threshold vector in real time. The envelope of the content condensation and reconstruction network. Simultaneously, the content condensation and reconstruction network solves the mutual information minimization constraint objective function. It automatically filters out high-density core text features within the high-security area of ​​system memory. Subjective colloquial noise in the text is eliminated, and the output is a highly convergent condensed cross-modal feature matrix. Next, the semantic space distance calculation unit performs a condensation of the cross-modal feature matrix within the shared manifold semantic space. Initial feature set of exogenous text modal Dot product summation and cosine similarity transform are used to generate semantic distance matrices in real time. The semantic distance matrix The mathematical mapping calculation uses a matrix transpose multiplication approach, and the specific matrix format expression is as follows: ; The central processing unit compares the semantic distance matrix With truncation constants, the set of automatic reverse semantic recall control terms And mounted to the structured lesion mapping atlas Quantitative and qualitative cross-causal auditing is performed on the 4-dimensional principal axis. If a threshold setting deviation failure is detected, the parameter feedback correction module immediately activates adaptive feedback control logic with dead-zone limitation to solve the weight coefficient feedback step size online. To adjust the overall degradation index for the next condition assessment .

[0245] To ensure the system's inference performance meets the mathematical stability standards for high-precision industrial control, the core algorithm parameters in this embodiment are based on deep scientific inference and solidified design using historical steady-state data distribution. Among these, the objective function used for content condensation and reconstruction constrains... Mutual information regularization skin constant in computation Strict locking is a fixed constant of 0.025.

[0246] Preferably, the penalty constant In gradient sensitivity testing during the R&D phase, the value was constrained to a range of 0.01 to 0.05.

[0247] When the penalty constant When the value of is greater than 0.05, excessive compression of the mathematical model will lead to condensation of the cross-modal feature matrix. Severe semantic information collapse occurs, resulting in the loss of specific process operation physical primitives; when the penalty constant... When the value is less than 0.01, the constraint strength is insufficient and it is impossible to completely eliminate manually entered text noise within the manifold space. Therefore, in this embodiment, a fixed value of 0.025 is used to ensure the semantic distance matrix... The optimal derived value with the highest splitting independence. Adaptive boundary adjustment coefficient. The solidification factor is set to 0.12 to provide optimal boundary convergence damping. The overall degradation index is then calculated. initial weight vector It is defined as a standard normalized row vector, with the specific matrix format expression as follows: ; Among them, the first weight coefficient For deviation residual matrix Trace weights, second weight coefficients The third weight coefficient is the distance weight for the dynamic boundary shrinkage norm. Process priority coefficient The nonlinear activation weights are used. The safety threshold value in the system cascaded error prevention and control chain is fixed as a scalar constant of 1.25. The time limit of the hardware anti-jitter filter for triggering the state is set to 20 consecutive time steps, that is, it only triggers when the comprehensive degradation index is... When the CPU deviates from the safety threshold of 1.25 for 20 consecutive frames, it activates the hardware interrupt level, thereby completely eliminating false triggering caused by electromagnetic pulses or transient network jitter in the industrial field from the perspective of mathematical control principle, and ensuring the extremely high stability of the closed-loop self-evolution control of the entire system.

[0248] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the adverse event analysis system based on the SOP system. In other embodiments of this application, the adverse event analysis system based on the SOP system may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0249] Example 4: This embodiment applies Example 1 to an online audit scenario of anomalies in a pharmaceutical injection production line. It selects the physical monitoring parameters of the process control bus in the controlled production line of a large process manufacturing enterprise's biopharmaceutical injection filling workshop during the production cycle from 08:00 to 10:00 on May 20, 2026, as the entity application scenario for target evaluation and data extrapolation. Addressing the unique challenges of this scenario, such as the high-frequency discrete fluctuations of multi-dimensional process equipment parameters online, and the shortcomings of traditional fixed single-variable deviation limits (deviation greater than 5%) in failing to sensitively detect microscopic continuous slow drift hazards in equipment, leading to missed critical risks, a deep and concrete demonstration of the full-link reasoning application process is presented.

[0250] S1: Pre-set the basic data matrix of standard operating procedures required for adverse event analysis in the business relationship database, establish the underlying static topology mapping between production equipment, process nodes and control limits, and complete the initialization of the baseline parameters of dynamic boundary tracking and semantic mapping space.

[0251] The specific input data types in this embodiment include the instantaneous output value of the pressure transmitter in the filling process, the value measured by the high-frequency electromagnetic flowmeter in the sterile pipeline, the sampled value of the environmental precision temperature and humidity sensor, and the cumulative physical downtime seconds of the filling machine's programmable logic controller (PLC). The physical dimensions are defined as megapascals (MPa), liters per hour (L / h), degrees Celsius (°C), and hours (h), respectively.

[0252] In this embodiment, the system main processor first establishes the process code in the main memory register area. This is an independent storage grid for Fill-02. Using the method in Example 1, the control references for each monitoring channel involved in this process are combined and assembled to generate a static initial reference constraint vector. The specific initial matrix element values ​​are defined as follows: ; Among them, the first item The filling pressure reference setting, in MPa; Item 2 The baseline setpoint for aseptic pipeline flow rate, in L / h; Item 3 The ambient temperature reference setting, in °C; Item 4 This is the upper limit of equipment downtime specified in the standard operating procedures, in hours (h).

[0253] The system processor retrieves the workshop asset topology list, identifies the first filling equipment and the first conveyor belt associated with the filling process, and extracts the equipment management identifiers for each. For SY-003 and equipment management identification It is SY-004.

[0254] In one feasible approach, the processor establishes a device process threshold mapping table. To solidify the underlying static topology.

[0255] In addition, the system processor retrieves the process priority coefficient defined in step S13. Since the filling process directly restricts the progress of subsequent light inspection and packaging processes, it is considered a bottleneck process. Therefore, the priority coefficient of this process is [not specified]. The discrete score is set to the scalar integer value 3.

[0256] In this embodiment, to activate envelope tracking based on unsupervised joint anomaly detection, the system processor extracts the historical sequence of the production line under stable operation over the past 72 hours through the data integration interface, calculates and initializes the long-cycle data feature matrix. The long-period mean vector is calculated. The system synchronously reads text segments containing all standard operating procedure clauses for the workshop pre-stored in the system's local non-volatile memory, extracts and encodes them through word segmentation features, and outputs an initial feature set of exogenous text modalities. Initial feature set of exogenous text modalities It is stored as a set of continuous manifold vectors with a hidden layer width of 256 dimensions in a high-security memory partition.

[0257] Intermediate Output: The initial operating condition configuration and dynamic baseline matrix specifically generated in this step are detailed in Table 1 below: Table 1 Initial Topology and Control Baseline Matrix for Filling Process

[0258] Specifically, Table 1 above shows that the low downtime (0.2h) and low pressure deviation characteristics in the long-period mean vector constitute a reference system for the system to determine the potential for microscopic continuous slow drift deterioration of equipment.

[0259] S2: Based on an adaptive dynamic threshold and trend joint tracking mechanism, it continuously parses multi-dimensional process monitoring data streams online through the underlying hardware interface and calculates the adaptive dynamic boundary threshold vector in real time. With the overall degradation index It performs time-series divergence measurement on micro-critical risks, determines and executes adaptive triggering of adverse event analysis processes.

[0260] In this embodiment, the three key hyperparameters of the unsupervised decision-making stage are set as follows: adaptive boundary adjustment coefficient. Set to a fixed constant of 0.12; weighted constant coefficient vector Set as ; Length of the dynamic timing sliding window The time step is set to 3600 consecutive time steps, corresponding to a continuous 180-second hard real-time monitoring section. The safety threshold value is fixed at 1.25.

[0261] At 09:00 on May 20, 2026, the system's central processing unit (CPU) used the industrial bus to parse and construct the process monitoring data vector for the current time step from the data register area of ​​the SY-003 device in real time. .

[0262] Specifically, the measured values ​​of the various physical quantities are as follows: ; The first item, 1.2 MPa, indicates an overshoot in the instantaneous filling pressure, and the fourth item, 1.2 h, indicates that the cumulative downtime seconds have exceeded the static control limit. The central processing unit calls the Gaussian mixture distribution unsupervised prediction model described in step S22, and calculates the theoretical prediction expectation vector at the current moment below the health baseline based on the historical distribution evolution trajectory within the dynamic time-series sliding window. .

[0263] First, the central processing unit (CPU) obtains the process monitoring data vector. With the predicted expected vector The transient deviation residuals between them are calculated by performing vector subtraction to obtain the difference vector. Referring to formula (1) in Example 1, perform the outer product operation of the difference vector and its own transpose vector to calculate and output the deviation residual matrix of the current time step. : ; Therefore, the deviation residual matrix is ​​extracted. The diagonal elements are used to obtain the current deviation variance of each feature channel. The central processing unit performs an accumulation and summation calculation to obtain the deviation residual matrix. The trace is: ; Secondly, the central processing unit calls the rolling timing cache to extract the cumulative historical deviation variance mean vector within the current dynamic timing sliding window. Referring to formula (2) in Example 1, and substituting the static initial reference constraint vector... and adaptive boundary adjustment coefficient Solve for the adaptive dynamic boundary threshold vector at the current time step. : ; Then, the central processing unit performs vector subtraction and Euclidean distance metric to calculate the adaptive dynamic boundary threshold vector. Relative static initial reference bound vector The deviation distance of the L2 norm. The numerical calculation process is as follows: ; Next, the process priority coefficient for curing in step S13 is invoked. The nonlinear hyperbolic tangent activation function is applied to it to obtain the continuous weight tensor value: ; Finally, referring to formula (3) in Example 1, and substituting the above-calculated intermediate variable floating-point numbers, the overall degradation index of the current time step is calculated in parallel. : ; The system data comparator will use the currently calculated overall degradation index. The floating-point value is compared with the preset safety threshold of 1.25. Because... Furthermore, this out-of-range state remained unchanged after 20 consecutive time steps of strong auditing by the hardware anti-jitter filter. The system determined that it was in a major abnormal deviation condition, immediately flipped the hardware trigger level of the control board, blocked the real-time bus, and encapsulated the multi-dimensional physical timing waveforms and control snapshots in the time domain for two hours before and after 09:00, latching the output trigger transient event dataset. .

[0264] S3: Based on two-stage asynchronous qualitative data entry and multimodal content condensation and reconstruction using triggered event data, a unique event identifier primary key containing spatiotemporal process constraints is generated. And perform cascaded hard lock error prevention verification.

[0265] In this step, the semantic preservation Lagrange multipliers of the content condensation reconstruction network are used. The curing setting is 0.05. Add a timer. The full-range time limit countdown is set to 20 minutes.

[0266] First, the system's central processing unit detects the triggering transient event dataset. The initial dimensionality reduction form is pushed to the industrial tablet held by the operator via the display driver the instant the data is written to the flash memory. The central processing unit automatically injects the captured event timestamp. and the physical location coordinates of the event It also receives initial manual qualitative text input by the operator. .

[0267] Preferably, the unstructured, conversational text to be filled in is: "Around nine o'clock, the filling machine suddenly stopped. The filling pressure seemed a bit unstable, and the conveyor belt next to it also seemed a bit stuck and delayed." Secondly, the central processing unit calls the content condensation and reconstruction neural network to convert the initial manually classified text... The initial feature set of exogenous text modalities in step S15 Import the shared manifold space. Referencing formula (4) from Example 1, perform Lagrange multiplier-based operations. The mutual information is minimized through iterative approximation computation. The network filters out subjective qualitative modifiers from the text sequence and reconstructs highly structured, non-qualitative, high-density core text features. .

[0268] Preferably, the core semantics of the output are explicitly represented as: ["filling equipment shutdown", "filling pressure divergence", "conveyor belt timing delay"].

[0269] Then, the central processing unit calls the unique primary key generation algorithm to extract the physical location coordinates of the event. In the equipment process threshold mapping table Process codes retrieved from The key “GZ-02” is concatenated with the event timestamp 202405200900 and the six-digit sequence “123456” generated by the true random number generator. This concatenation is used to assemble and output a globally unique non-standard business key, which is defined as the unique event identifier key. "GZ-02-202405200900-123456".

[0270] Next, the main processor instantiates the dual-loop sequential state machine. This causes its state to transition from idle to emergency suspended state. Synchronously activate supplementary timer A 20-minute countdown begins. During the window, the operator manually completes the input of the selected device identifier. “SY-003” and the operator-selected procedure number “sop-GZ-02-001”. The system processor immediately reads the unique event identifier primary key. The prefix character "GZ-02" in the table applies conditional filtering to the list of associated equipment interactions for the corresponding process in the equipment process threshold mapping table, thus extracting a subset of cascaded constraint equipment. [“SY-003”,“SY-004”]. Central Processing Unit Execution Unit Element Inclusion Determination ( Since "SY-003" belongs to a subset of this cascaded constraint device... The core internal element outputs a logical true result upon verification, at which point the system unlocks the interface.

[0271] The system will uniquely identify the primary key for each event. High-density core text features Compliant operators select equipment identifiers Operator selects procedure number and the dataset that triggers transient events Perform deep, high-density packaging to output a two-level structured reporting dataset. .

[0272] S4: Based on a unique event identifier primary key, it traverses heterogeneous industrial buses, performs spatiotemporal adaptive data routing and intelligent interpolation of underlying microscopic physical parameters, implements multidimensional strong consistency auditing, and assembles a high-fidelity multimodal event dataset. .

[0273] Topological attenuation coefficient Based on the fluid dynamics and physical inertial constant of the pharmaceutical liquid delivery pipeline, the pre-configured constant is fixed at 0.85. The number of retries for interface calls is limited to 3, with a retry interval set to 10 seconds.

[0274] First, the system's central processing unit calls step S3 to generate a two-level structured reporting dataset with sequential shifting. Extract the unique event identifier primary key “GZ-02-202405200900-123456”. The processor's reverse hash decomposition yields the process code. "GZ-02" and the time stamp of the event Based on a standard production rhythm of two hours per batch, the processor uses 202405200900 as the central axis, symmetrically and adaptively expanding forward and backward by one hour to dynamically generate precise data retrieval time windows. The period is: "2024-05-20 08:00:00 to 2024-05-20 10:00:00".

[0275] Secondly, the central processing unit will assign process code GZ-02, operator-selected equipment identifier SY-003, and data retrieval time window. Inject into the database proxy connection pool. Invoke the equipment process threshold mapping table. Perform a primary key-foreign key-inner join query to automatically parse the underlying PLC register pointer address and data acquisition point mapping table of the target physical controlled device, and format and assemble the output heterogeneous system data retrieval requirement list. .

[0276] Then, the central processing unit activates the underlying JSON-OPCUA protocol conversion socket to retrieve the data retrieval request list from the heterogeneous system. The data is compiled into NodeID addressing messages conforming to the IEC 62541 standard and a direct memory access request is initiated directly to the field IoT gateway via the workshop industrial Ethernet. The socket continuously captures all high-frequency underlying physical data of the SY-003 machine from 08:00 to 10:00. Due to an accidental shutdown at 09:00, the specific underlying hardware status code at that time was not filled in the operator form. The system data cleaning layer detects this missing item and immediately initiates the timing intelligent interpolation mechanism.

[0277] Specifically, the processor retrieves heterogeneous timing raw data. Continuous dynamic pressure waveform data was extracted for 10 milliseconds before 09:00, and the time-series mean scalar value within this microscopic window was measured. MPa, the continuous transient first derivative of the pressure time series signal with respect to time (i.e., the pressure deviation slope) is calculated as follows: MPa / s. Referring to formula (5) in Example 1, substitute the topological attenuation coefficient... Perform deterministic physical feature inverse interpolation solution: ; The central processing unit (CPU) uses the calculated deterministic physical interpolation scalar of 0.3000 MPa as high-fidelity factual data, and forcibly injects and fills the missing form field "Instantaneous filling pressure at the time of equipment failure" in the two-level structured reporting dataset. Simultaneously, it retrieves the hardware fault code at the same timestamp, extracting the underlying hardware alarm code as "E01," indicating conveyor belt overload. After all fields are filled, the processor outputs the interpolated time-series data matrix. .

[0278] Next, the central processing unit calls the consistency check kernel to process the interpolated timing data matrix. Implement a three-tiered, sequential, and robust audit.

[0279] Level 1 Hardware Identifier Audit: Verify the extracted network card hardware MAC address against the associated device interaction list. The registration code for SY-003 was checked to ensure it was 100% identical to prevent misalignment caused by industrial WiFi signal drift. The result was satisfactory. Second-level production batch audit: Verify the matching between the injection batch number issued by the MES system in this time window and the GZ-02 process; the result is satisfactory. Third-level tolerance range audit: Verify whether the instantaneous pressure of 0.30 MPa injected through reverse interpolation is within the static initial reference limit vector of step S11. The defined tolerance zone for reasonable pressure fluctuations (0.4 MPa) Within 0.1 MPa, after comparison, 0.30 MPa falls exactly at the lower limit critical point, and the audit is passed.

[0280] Intermediate Output: After completing the multi-dimensional strong audit, the processor outputs the legally cleaned interpolated time-series data matrix. High-density core text features centrally transmitted in the two-level structured reporting dataset Operator selects procedure number and the dataset that triggers transient events Perform structured assembly of the high-security zone to officially output a high-fidelity multimodal event dataset. .

[0281] S5: Based on a high-fidelity multimodal event dataset, perform cross-modal semantic automatic mapping and content condensation reconstruction to construct a structured four-dimensional causal topology primitive, and achieve high-precision semantic alignment and root cause determination of the underlying physical time series and management procedure text.

[0282] This step involves running the MindTS cross-modal feature space alignment engine from the ICLR 2026 conference paper. Its regularization penalty constant... The strict lockout setting is set to a fixed constant of 0.025. The upper limit of the token length in the exogenous semantic processing backbone network is hard-coded to 128 dimensions.

[0283] First, the central processing unit receives the high-fidelity multimodal event dataset with global dependencies sent down in the previous step. Start the feature encoding neural network to interpolate the time-series data matrix in the dataset. Performing a one-dimensional temporal convolutional network (TCN) sliding residual convolution operation, stretching along the time step to extract the nonlinear temporal manifold features of the waveform, and outputting a high-dimensional temporal manifold feature matrix. .

[0284] Secondly, the central processing unit activates the cross-modal alignment engine. This involves aligning the temporal manifold feature matrix... High-density core text features and the initial feature set of exogenous text modalities in step S1 Unified compression into the aligned hidden space. Referring to formula (6) in Example 1, substitute the regularization penalty constant. It performs content condensation and reconstruction optimization based on the expectation operator and the squared term of the Frobenius norm, completely filtering out unstructured redundancy in the text, and solving the output into a condensed cross-modal feature matrix that converges absolutely in the semantic manifold space. .

[0285] Then, the semantic space distance calculation unit performs matrix transpose multiplication to calculate the condensed cross-modal feature matrix. Initial feature set of exogenous text modal The inverse of the cosine similarity at the row vector level is used to solve for and output a semantic distance matrix in real time, representing the distance between the physical features of factual anomalies and the written management clauses. The central processing unit retrieves the semantic distance matrix. For feature entries whose feature distance is less than the truncation threshold, reverse engineering can accurately recall the corresponding standard operating procedure text clauses.

[0286] Preferably, the set of standard text clauses for successful recall is defined as the set of reverse semantic recall control clauses. The document contains two core written regulations: one is "SOP-GZ-02-001, Clause 5.2: If the filling equipment is shut down for more than 1 hour at a time, the abnormal handling process control must be activated"; the other is "SOP-GZ-02-001, Clause 7.3: After the filling process is completed, the conveyor belt must transport the product to the next process within 10 minutes".

[0287] Next, the central processing unit initializes a four-dimensional causal topological primitive skeleton in the graphics memory. It then performs targeted branching and loading of all the aforementioned multi-source heterogeneous factual data: the adaptive dynamic boundary threshold vector... Numerical values ​​are filled into the threshold dimension principal axis; deviation from the residual matrix Snapshot values ​​are entered into the trend dimension main axis; process priority coefficients. Enter the process weight dimension of the main axis; interpolate the time series data matrix. The conveyor belt delay start time and fault code E01 are included in the process connection dimension main axis. The set of reverse semantic recall control clauses is simultaneously added. The text semantic tags are respectively attached and aligned to the secondary graphic bone nodes of the threshold dimension and process connection dimension, and the graphics are merged to output a structured lesion mapping map. .

[0288] Finally, the central processing unit invokes the intelligent attribution verification and auditing kernel to map the structured lesion map. Cross-verification of internal evolution data: Audited data shows that, within the threshold dimension branch, the actual equipment downtime... Although the 1.0-hour delay deviated from the initial static limit, a comparison of the final quality inspection report of the entire batch of products returned by the MES revealed that the actual product qualification rate of the injection solution was as high as 99.2%, significantly higher than the 98% strict control threshold specified in step S11. This proved that the short-term shutdown of 1.2 hours did not cause any substantial physical quality damage to the sterile stable phase of the drug solution. The originally set absolute static threshold of 1 hour was too low and did not match the risk tolerance for short-term flashing shutdowns in the actual manufacturing process. Combining the convergence characteristics of the deviation energy in the trend dimension branch, the central processing unit performed qualitative attribution and determined that the root cause of this adverse event was a design flaw in the control procedure.

[0289] The central processing unit (CPU) encapsulates the attribution coding and supporting evidence data chain in a structured manner, and outputs the final root cause determination matrix. This clearly identifies the fundamental attribute of the current event as a deviation in the static threshold setting of the filling equipment.

[0290] S6: Based on the final root cause determination result, perform parameter feedback correction and procedure text revision with adaptive dead zone control, complete historical data replay and backtracking verification in the isolated sandbox memory area, and build a highly secure control clause-level closed-loop evolution engine.

[0291] Convergence step size adjustment factor in adaptive feedback control loop The setting is 0.05. Adaptive dead zone deviation threshold. The firmware setting is 0.15. Historical test data packages. The backtracking extraction span is configured to be the full historical process control log of the past 30 consecutive calendar days.

[0292] First, the central processing unit reads the final root cause determination result matrix output in step S5. Because the attribute identifier code parsing points to "threshold setting deviation", the system automatically triggers hardware routing switching, activates the dynamic parameter adjustment bus of the targeted activation algorithm, and performs adaptive self-adjustment of control parameters.

[0293] Secondly, the central processing unit invokes the adaptive feedback control algorithm. This is based on the final root cause determination result matrix. Extract the actual downtime at the moment the accident occurred from the quantitative chain of evidence. Hours and the original static reference limit Calculate the peak relative deviation scalar over hours. The central processing unit executes the size comparison logic, because... This means that the adaptive dead zone deviation threshold is exceeded, and the parameter feedback correction module is officially activated. Referring to formula (7) in Example 1, subtraction and multiplication operations are performed to solve for the weight coefficient feedback step size of the current optimization cycle. : ; The central processing unit feeds this weight coefficient back in with a step size of 0.0025 and adds it to the original threshold dimension weight coefficient. The online tuning was modified to 0.4025. Simultaneously, to ensure that the costly system root cause diagnosis process is only passively triggered under more severe fault deviations, the CPU fine-tuned the trigger condition preset ratio from the original 0.20 to 0.30. The revised parameters, along with the linked parameters, adaptively revised the downtime limit in the standard operating procedure's paper text from 1.0 hour to 1.3 hours (i.e.,... Structured revision push vector The unified assembly output is used to update the dynamic parameter set. .

[0294] Then, the central processing unit allocates a sandbox memory area in main memory that is physically isolated unidirectionally from the physical process control network. The historical monitoring data and frequently triggered snapshots accumulated over the past 30 days were imported to a disk and reconstructed within the sandbox to create a historical test data package. The central processing unit (CPU) is located in the isolated sandbox memory area. The unsupervised synthesis degradation index calculation module in the internal re-instantiation step S2 will update the dynamic parameter set. The fixed configuration was injected into it, and the historical monitoring data replay and retrospective test was fully launched.

[0295] Specifically, the central processing unit re-inputs the actual downtime deviation residual data that occurred at 09:00 into the virtual degradation calculation engine. At this point, because the preset trigger condition ratio has been adaptively increased to 0.30, and the current actual relative deviation ratio of this event is 0.20, the calculation result directly outputs that the relative deviation ratio is less than the newly set risk threshold of 0.30. This minor, benign short-term downtime anomaly is accurately intercepted and filtered in the sandbox virtual test, without causing false triggers. Simultaneously, the backtesting engine replay shows that the remaining 27 historical degradation events that truly led to major quality defects can still be accurately captured and triggered by the new parameters 100%, proving that the optimized parameter set, while possessing high convergence, did not cause any hidden dangers to be missed.

[0296] Next, the sandbox backtest fully met the preset safety and stability indicators, and the central processing unit output a Boolean-type backtest verification mark. True.

[0297] Finally, the central processing unit (CPU) captures the truth beacon, and the master transaction manager executes a global transaction commit. This will update the dynamic parameter set. The newly fixed threshold and adjusted weighting coefficients are formally written into the standard operating procedure's basic database, physically replacing and overwriting the original static initial baseline constraint vector in step S11, completing the cascading hard update of the database and tables. The central processing unit automatically extracts the operation return status codes of each underlying physical module and the database transaction mirror, outputting a closed-loop update log. Store in the hard audit disk partition.

[0298] Example 5: Further explanation in conjunction with Example 1, such as Figure 3 The structure shown is... Figure 3 A schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes: Processor, memory, communication bus, and computer programs stored in memory that can run on the processor.

[0299] The processor can call a computer program in memory, and when executing the program, implement the adverse event analysis system and method based on the SOP system provided in the above embodiments. The method includes: S1: according to the standard operating procedure text control clauses Establishing an initial feature set for exogenous text modalities And construct a mapping table of equipment process thresholds. ; S2: Obtain physical sensor parameters to construct process monitoring data vectors Based on process monitoring data vector Latch-triggered transient event dataset ; S3: Based on the initial manual qualitative text Initial feature set of exogenous text modal Generate a unique event identifier primary key The level signal is locked by the interface hardware. Block and output a two-level structured reporting dataset ; S4: Use a unique event identifier for the primary key Retrieve heterogeneous time series raw data And reverse-inject a two-level structured reporting dataset. Generate a high-fidelity multimodal event dataset ; S5: Based on a high-fidelity multimodal event dataset Generating the temporal manifold feature matrix The temporal manifold feature matrix Initial feature set of exogenous text modal Mapping to a set of recall reverse semantic recall control clauses And generate a structured lesion mapping atlas. Determine the final root cause determination matrix. ; S6: Analyze the final root cause determination result matrix isolated sandbox memory area When the backtracking test converges safely, update the static initial baseline constraint vector. .

[0300] Furthermore, computer equipment also includes: The Communications Interface (CI) is used for communication between the memory and the processor.

[0301] The memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0302] If the memory, processor, and communication interface are implemented independently, they can be interconnected via a bus to communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0303] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0304] Display devices are used to display images, videos, etc. Display devices may include display panels, which may employ liquid crystal displays (LCDs), organic light-emitting diodes (OLEDs), active-matrix organic light-emitting diodes (AMOLEDs), flexible light-emitting diodes (FLEDs), MiniLEDs, MicroLEDs, Micro-OLEDs, quantum dot light-emitting diodes (QLEDs), etc.

[0305] Alternatively, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, then the memory, processor, and communication interface can communicate with each other through an internal interface.

[0306] On the other hand, embodiments of this application also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described adverse event analysis system and method based on the SOP system. The method includes: S1: Controlling the terms of the standard operating procedure text. Establishing an initial feature set for exogenous text modalities And construct a mapping table of equipment process thresholds. ; S2: Obtain physical sensor parameters to construct process monitoring data vectors Based on process monitoring data vector Latch-triggered transient event dataset ; S3: Based on the initial manual qualitative text Initial feature set of exogenous text modal Generate a unique event identifier primary key The level signal is locked by the interface hardware. Block and output a two-level structured reporting dataset ; S4: Use a unique event identifier for the primary key Retrieve heterogeneous time series raw data And reverse-inject a two-level structured reporting dataset. Generate a high-fidelity multimodal event dataset ; S5: Based on a high-fidelity multimodal event dataset Generating the temporal manifold feature matrix The temporal manifold feature matrix Initial feature set of exogenous text modal Mapping to a set of recall reverse semantic recall control clauses And generate a structured lesion mapping atlas. Determine the final root cause determination matrix. ; S6: Analyze the final root cause determination result matrix isolated sandbox memory area When the backtracking test converges safely, update the static initial baseline constraint vector. .

[0307] In another aspect, embodiments of this application also provide a computer program product, which includes a computer program that can be stored on a computer-readable storage medium. The computer program can execute computer instructions, and when executed by a processor, the computer can perform the adverse event analysis system and method based on the SOP system provided by the above methods. The method includes: S1: Controlling the terms of the standard operating procedure text. Establishing an initial feature set for exogenous text modalities And construct a mapping table of equipment process thresholds. ; S2: Obtain physical sensor parameters to construct process monitoring data vectors Based on process monitoring data vector Latch-triggered transient event dataset ; S3: Based on the initial manual qualitative text Initial feature set of exogenous text modal Generate a unique event identifier primary key The level signal is locked by the interface hardware. Block and output a two-level structured reporting dataset ; S4: Use a unique event identifier for the primary key Retrieve heterogeneous time series raw data And reverse-inject a two-level structured reporting dataset. Generate a high-fidelity multimodal event dataset ; S5: Based on a high-fidelity multimodal event dataset Generating the temporal manifold feature matrix The temporal manifold feature matrix Initial feature set of exogenous text modal Mapping to a set of recall reverse semantic recall control clauses And generate a structured lesion mapping atlas. Determine the final root cause determination matrix. ; S6: Analyze the final root cause determination result matrix isolated sandbox memory area When the backtracking test converges safely, update the static initial baseline constraint vector. .

[0308] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0309] For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. Furthermore, a computer-readable medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0310] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0311] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0312] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0313] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0314] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0315] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for adverse event analysis based on a Standard Operating Procedure (SOP) system, characterized in that, Includes the following steps: S1: Control clauses in accordance with standard operating procedure text Establishing an initial feature set for exogenous text modalities And construct a mapping table of equipment process thresholds. ; S2: Obtain physical sensor parameters to construct process monitoring data vectors Based on process monitoring data vector Latch-triggered transient event dataset ; S3: Based on the initial manual qualitative text Initial feature set of exogenous text modal Generate a unique event identifier primary key The level signal is locked by the interface hardware. Block and output a two-level structured reporting dataset ; S4: Use a unique event identifier for the primary key Retrieve heterogeneous time series raw data And reverse-inject a two-level structured reporting dataset. Generate a high-fidelity multimodal event dataset ; S5: Based on a high-fidelity multimodal event dataset Generating the temporal manifold feature matrix The temporal manifold feature matrix Initial feature set of exogenous text modal Mapping to a set of recall reverse semantic recall control clauses And generate a structured lesion mapping atlas. Determine the final root cause determination matrix. ; S6: Analyze the final root cause determination result matrix isolated sandbox memory area When the backtracking test converges safely, update the static initial baseline constraint vector. .

2. The adverse event analysis method based on the SOP system according to claim 1, characterized in that, Step S1 includes the following sub-steps: S11: Extract the standard operating procedures, quality control requirements, and equipment operation standards for each production process to establish a static initial benchmark constraint vector. For quality-related processes, a minimum product qualification rate is defined; for equipment-related processes, a baseline value for single equipment downtime is defined; and each static initial baseline is defined as a vector. Corresponding process code Perform unique memory address binding; S12: Compile a structured list of all production equipment required for each production process and create a list of related equipment interactions. In the list of interactions with associated devices Record the equipment model of each production machine , To which process node Equipment Management Identification and equipment operating parameter acquisition port information ; List of interactions with associated devices Each piece of production equipment is associated with a corresponding process code. The static initial reference constraint vector bound to the corresponding process. Generate a device process threshold mapping table by performing three-dimensional primary and foreign key association binding. Map the equipment process thresholds to the table. Persistently store the data in the business relationship database; S13: Quantify the impact weight of each production process on the final product quality and the constraint stiffness on the advancement of subsequent processes; set process priority coefficients. The process priority coefficient The process is divided into three discrete score levels, and the priority coefficient of each process is assigned accordingly. Corresponding process code Perform mapping; S14: Retrieve the historical sequence of long-cycle physical parameters under continuous, error-free operation conditions through the data integration interface of the workshop manufacturing execution system. Calculate the historical sequence of long-period physical parameters Long-period mean vector With long-period variance matrix Combining long-period mean vectors With long-period variance matrix Generate long-period data feature matrix This is to initialize the data distribution for the adaptive dynamic boundary triggering mechanism. S15: Read all standard operating procedure text control clauses stored in the business relationship database. Using a pre-trained natural language processing backbone network to control standard operating procedure text clauses Word segmentation, stop word removal, and hidden layer mapping are performed to extract semantic features and generate an initial feature set of exogenous text modalities stored in the manifold space. This completes the initialization of the exogenous semantic base for the cross-modal semantic mapping mechanism.

3. The adverse event analysis method based on the SOP system according to claim 1 or 2, characterized in that, Step S2 includes the following sub-steps: S21: The time-series parameters of the physical sensors of the production equipment are retrieved in real time through the workshop data acquisition layer. The multi-dimensional physical sensor parameters are then spatiotemporally aligned and denoised and normalized to assemble and generate the process monitoring data vector for the current time step. ; S22: Read the dynamic timing sliding window in the running memory and combine it with the long-period data feature matrix. An unsupervised prediction model is established, and the expected prediction vector at the current time step is calculated using the state evolution probability density function. ; S23: Calculate process monitoring data vector With the predicted expected vector The transient deviation residuals are used to construct the deviation residual matrix by performing an outer product operation on the transient deviation residuals and their transposes. ; S24: Extract the deviation residual matrix Logarithmic integration is performed on the diagonal elements within the dynamic time-series sliding window to generate a degradation trend negative feedback damping factor, which then constrains the static initial benchmark vector. Combined with the negative feedback damping factor of the degradation trend, a preset adaptive boundary adjustment coefficient is used. Execute Hadama The adaptive dynamic boundary threshold vector is calculated. ; S25: Extract the deviation residual matrix The trace term is used to calculate the adaptive dynamic boundary threshold vector. Relative static initial reference bound vector The L2 distance is used to apply the hyperbolic tangent function to the process priority coefficient. Perform a nonlinear continuous mapping; combine the trace term, the L2 distance, and the mapped process priority coefficients. Perform multidimensional mixed linear weighting to calculate and output a dimensionless comprehensive degradation index. ; S26: Overall Deterioration Index Perform a step-condition check control chain audit when the overall degradation index is... When the time span exceeding the safety threshold reaches the time limit of the hardware anti-jitter filter, a hardware interrupt command is sent to the physical bus actuator to toggle the hardware trigger level, blocking real-time bus data and vectorizing the process monitoring data. Sequence, Deviation Residual Matrix Snapshot and Adaptive Dynamic Boundary Threshold Vector The snapshot is packaged and latched to generate a dataset that triggers transient events. .

4. The adverse event analysis method based on the SOP system according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31: Monitor the dataset of triggered transient events Write status, activate the simplified initial reporting process, and capture the event timestamp. Physical location coordinates of the event Receive initial manual qualitative text Temporarily stored in the application layer memory queue; S32: Call the initial feature set of the exogenous text modality Compared with the initial human qualitative text Perform cross-modal semantic space projection to extract initial manually defined text. High-density core text features The first non-linear dependency between them is used to calculate high-density core text features. Initial feature set of exogenous text modal The second nonlinear dependency between them and the introduction of semantically preserving Lagrange multipliers Perform scaling; subtract the scaled second nonlinear dependency from the first nonlinear dependency to construct the mutual information minimization objective function. The objective function is optimized by backpropagation to minimize mutual information. To achieve convergence, subjective and invalid expressions are stripped away, and high-density core text features are output. ; S33: Call the physical location coordinates of the event Equipment process threshold mapping relationship table Initiate a spatial topology addressing query and retrieve the corresponding process code in reverse. The process code is encoded using a string hash concatenation algorithm. Event timestamp Physically concatenate with a secure random sequence code to generate a unique event identifier primary key with embedded spatiotemporal process constraints. ; S34: Instantiate a dual-loop sequential state machine Towards a dual-closed-loop sequential state machine Inject transition instructions to force a dual-loop sequential state machine. The running state transitions from idle state to emergency suspension state. A high-precision timer is triggered by the rising edge signal of the transition synchronization to start the supplementary timer. ; S35: Receive the operator-selected device identifier Operator selected procedure number Extract the unique event identifier primary key Included process codes Using first-paradigm filtering conditions, the list of interactions with associated devices is used. Perform dynamic view clipping to generate a subset of cascading constraint devices. ; Execute Boolean logic comparison instructions to determine the operator-selected device identifier Does it belong to a subset of cascaded constraint devices? When the internal element of the function is determined to be logically false, it outputs a hardware latch-up signal to the front-end I / O controller. Cut off the underlying touch event response and forcibly block the path for dirty data to enter the database; S36: In a dual-closed-loop sequential state machine Determine if the Boolean logic comparison instruction is logically true and add a timer. If the overflow has not yet reached zero, the primary key will be uniquely identified by the event. High-density core text features Operator selects equipment identifier Operator selects procedure number and the dataset that triggers transient events Perform deep packet encapsulation to generate a two-level structured reporting dataset. .

5. The adverse event analysis method based on the SOP system according to any one of claims 1 to 4, characterized in that, Step S4 includes the following sub-steps: S41: Extract the two-dimensional structured reporting dataset Encapsulated unique event identifier primary key For unique event identifier primary key Perform reverse hashing of the string to extract the process code. Timestamp of the event Read process code The corresponding standard production cycle parameters are based on the event occurrence timestamp. Extending bidirectionally from the origin to both the historical and future time domains, an adaptively aligned data retrieval time window is constructed. ; S42: Joint Process Code With the operator selecting the device identifier Perform primary-foreign key cascading addressing in a relational database, utilizing process coding. In the equipment process threshold mapping table Initiate an inner join query to retrieve the list of interactions between related devices. Defined physical sensor measurement points and production execution management control points; compare the operator-selected equipment identifiers. By matching the physical sensor measurement points, the register address of the underlying programmable logic controller and the production batch number of the manufacturing execution system are extracted, along with the data retrieval time window. Perform formatted packaging to generate a list of data retrieval requirements for heterogeneous systems. ; S43: Activate the protocol conversion socket and use the mapped hash table structure to retrieve the data retrieval request list from heterogeneous systems. The included qualitative fields are mapped to node identifiers, and a direct memory access request is initiated to the workshop's underlying equipment monitoring and management system, according to the data retrieval time window. Copy the high-frequency operating parameters of the target device with defined start and end scales, and output heterogeneous timing raw data. ; S44: Call heterogeneous timing raw data With two-order structured reporting datasets Perform null value matching based on the event timestamp. Extracting heterogeneous timing data as the reference axis The included physical waveform curves are used to extract the time-series mean vector of the physical waveform curves within the truncated time window, and the corresponding continuous derivative term of the physical waveform curves is calculated. This continuous derivative term is then compared with a pre-configured topology attenuation coefficient. Perform multiplication and add the result to the time-series mean vector to solve for deterministic physical characteristics, which are then back-injected into empty form fields as interpolated physical scalars. Reorganize to generate a fully encoded interpolated time series data matrix. ; S45: The system consistency check kernel is started to perform a three-level serial error prevention audit and compare the interpolated timing data matrix. Includes a list of actual network card physical addresses and associated device interactions. Includes operator-selected device identifier The binding relationship is audited using hardware identifiers, comparing the batch parameters returned by the manufacturing execution system with the process codes. The mapping relationship is used to perform production batch auditing and compare the interpolated time series data matrix. Includes key equipment operating parameters and static initial baseline constraint vectors The defined reasonable fluctuation tolerance zone is used for tolerance interval auditing; when the results of hardware identification audit, production batch audit, and tolerance interval audit are all logically true, the interpolated time series data matrix will be used. High-density core text features Operator selects procedure number With triggering transient event dataset Structured assembly is performed in the high-security area of ​​system memory to generate a high-fidelity multimodal event dataset. .

6. The adverse event analysis method based on the SOP system according to claim 1, characterized in that, Step S5 includes the following sub-steps: S51: Reading a high-fidelity multimodal event dataset In the high-security area of ​​system memory, a high-fidelity multimodal event dataset is processed according to a preset pointer offset protocol. Perform modal separation and unpack the interpolated time series data matrix. High-density core text features and the dataset that triggers transient events ; Interpolate the time series data matrix The input is a feature encoding neural network containing one-dimensional causal convolutional layers. A sliding convolution operation is performed along the time axis to extract waveform slope, peak-to-peak value, and local oscillation frequencies. The output is continuous time-series modal features to generate a temporal manifold feature matrix. ; S52: Transform the temporal manifold feature matrix High-density core text features Initial feature set of exogenous text modal The input is fed into the cross-modal alignment engine, where a content condensation and reconstruction mechanism is introduced to calculate the temporal manifold feature matrix. High-density core text features The content condensation network maps the posterior conditional distribution with respect to the condensed cross-modal feature matrix under the joint prior distribution. The log-likelihood ratio approximating the marginal distribution is combined with the expected value and a pre-defined regularization penalty constant. Constrained condensation of cross-modal feature matrix Initial feature set of exogenous text modal The Frobenius norm squared terms between the terms are used to construct the mutual information condensation and reconstruction constraint objective function. Redundant modal information is filtered by successively updating the network weights to minimize the mutual information condensation and reconstruction constraint objective function, and the reconstructed output is a condensed cross-modal feature matrix. ; S53: Condensing the cross-modal feature matrix The feature vectors in the text and the initial feature set of the exogenous text modality The standard operating procedure control clause vectors contained therein are subjected to dot product and cosine similarity transformation, and the inverse of the similarity is mapped to feature distance to generate a semantic distance matrix. ; Using a preset semantic recall cutoff constant in the semantic distance matrix The process involves filtering candidate clauses, extracting complete real text clause data that meets the filtering criteria from the relational database, and combining them to generate a set of reverse semantic recall control clauses. ; S54: Construct a four-dimensional causal topology primitive skeleton within the video memory space, dividing the four-dimensional causal topology primitive skeleton into a threshold dimension principal axis, a trend dimension principal axis, a process weight dimension principal axis, and a process connection dimension principal axis; trigger transient event datasets. Includes adaptive dynamic boundary threshold vector Filling to the threshold dimension principal axis will trigger the transient event dataset. Included deviation residual matrix Fill the trend dimension main axis with the process priority coefficient. Fill the interpolated time series data matrix to the principal axis of the process weight dimension. Fill in the main axis of process connection; control the set of reverse semantic recall clauses. The text semantic tags are parsed, and the text control procedures are anchored to the secondary branches corresponding to the threshold dimension main axis, trend dimension main axis, process weight dimension main axis, and process connection dimension main axis, outputting a structured lesion mapping map. ; S55: Comparison of structured lesion mapping maps Adaptive dynamic boundary threshold vector on the principal axis of the threshold dimension With static initial reference bound vector In determining the adaptive dynamic boundary threshold vector The contraction amplitude exceeds the set deviation bandwidth and the interpolated time series data matrix When non-abrupt degradation is observed, the root cause is determined to be critical missed detection failure due to threshold setting deviation; in determining the adaptive dynamic boundary threshold vector... A set of steady-state and reverse semantic recall control clauses When the text contains text indicating operational violations, the root cause is identified as a loophole in the operational procedures for process connection. The attribute identifier code corresponding to the root cause, the feature matrix involved in the judgment, and the attached text clause number are structurally encapsulated to generate the final root cause judgment result matrix. .

7. The adverse event analysis method based on the SOP system according to claim 1, characterized in that, Step S6 includes the following sub-steps: S61: Read the final root cause determination result matrix The final root cause determination matrix is ​​processed at the kernel level. The attribute identifiers encapsulated in the code are subjected to bitwise AND logical operations to determine the fundamental attributes of the event; When the attribute identifier code is mapped to the threshold setting deviation, resulting in critical missed alarm failure or false alarm triggering, the internal algorithm dynamic parameter adjustment bus is activated to enter the control parameter adaptive correction branch. When the attribute identifier code is mapped to a loophole in the process connection operation specification or ambiguity in the control text, the system's standard operating procedure revision push queue is activated to enter the management procedure text optimization branch. S62: Execute the adaptive correction branch for control parameters, based on the final root cause determination result matrix. Extracting interpolated time series data matrix from mounted evidence chain The physical waveform extrema within, combined with the static initial reference constraint vector Calculate the peak relative deviation scalar ; Comparison of peak relative deviation scalar The absolute value and adaptive dead zone deviation threshold ; scalar of relative peak deviation The absolute value is less than the adaptive dead zone deviation threshold. At that time, the weighting coefficients are fed back to the step size. Forced to zero to maintain absolute static hard locking of the weighting coefficients; in the peak relative deviation scalar The absolute value is greater than or equal to the adaptive dead zone deviation threshold. When calculating the peak relative deviation scalar The absolute value and adaptive dead zone deviation threshold The difference is calculated and then compared with the convergence step size adjustment factor. Perform multiplication calculations to determine the weight coefficient feedback step size. Feedback the weighting coefficients to the step size Accumulated to the original weight coefficient The above outputs the updated weight vector. Based on the magnitude of missed and false alarms, the preset ratio of the triggering conditions is adjusted synchronously, and an updated dynamic parameter set is generated and encapsulated. ; S63: Execution Management Procedure Text Optimization Branch, Extracting High-Density Core Text Features With the set of reverse semantic recall control clauses ; Launch a natural language generation model to retrieve control clause sets via reverse semantics. Based on the modified object, high-density core text features are used. The exact physical phenomenon primitives contained therein are used as supplementary entities to perform text concatenation and grammatical reconstruction, and the reconstructed revision suggestion text and process code are then combined. And serialize and encode the globally unique identifier to generate a structured revision push vector. ; S64: Creates an isolated sandbox memory area in physical memory that is physically isolated unidirectionally from the main control bus. Batch extraction of historical operating condition data snapshots and triggered transient event datasets from relational databases. In the isolated sandbox memory area Internally reassembled into historical test data packets In the isolated sandbox memory area The internally instantiated virtual synthesis degradation index calculation engine will update the dynamic parameter set. Or structured revision push vector Injecting the virtual integrated degradation index calculation engine into historical test data packets Perform a time-series replay operation; if the replay operation result indicates safe convergence and no new false positives or negatives are generated, verify the pass of the Boolean-type backtracking test. The value is assigned to logical true, and the backtracking test verification is passed when the replay operation results diverge or conflict. The logic is false, and a termination rollback interrupt signal is sent to the central processing unit's main control module. S65: Continuous monitoring of isolated sandbox memory area The output backtracking test verification passed the flag. Upon detection of the backtracking test verification pass flag Activating database write locks and enabling global transaction management when logically true; utilizing the update dynamic parameter set. Physical coverage of the static initial baseline constraint vector stored in the standard operating procedure base database The structured revision may be pushed to the vector. The text tensors in the text are remapped to the manifold space to physically replace the corresponding initial feature sets of the exogenous text modalities. The original vector entries in the database are used to complete the semantic base hard update; the return status code and confirmation timestamp of the database table structure change operation are captured, and a closed-loop update log is generated by packaging them using a cryptographic hash algorithm. The data is persistently stored in the system audit storage node, and the control system returns to steady-state monitoring mode.

8. An adverse event analysis system based on a Standard Operating Procedure (SOP) system, characterized in that, include: The SOP basic data preset module is used to control the terms and conditions of the standard operating procedure text. Establishing an initial feature set for exogenous text modalities And construct a mapping table of equipment process thresholds. ; The 3D trigger model construction and trigger module is used to acquire physical sensor parameters to construct process monitoring data vectors. Based on process monitoring data vector Latch-triggered transient event dataset ; The step-by-step reporting management module is used to report data based on the initial manual qualitative text. Initial feature set of exogenous text modal Generate a unique event identifier primary key The level signal is locked by the interface hardware. Block and output a two-level structured reporting dataset ; The cross-system data automatic association module is used to identify primary keys using unique events. Retrieve heterogeneous time series raw data And reverse-inject a two-level structured reporting dataset. Generate a high-fidelity multimodal event dataset ; Fishbone diagram root cause localization module, used to locate the root cause of events based on a high-fidelity multimodal event dataset. Generating the temporal manifold feature matrix The temporal manifold feature matrix Initial feature set of exogenous text modal Mapping to a set of recall reverse semantic recall control clauses And generate a structured lesion mapping atlas. Determine the final root cause determination matrix. ; The triggering rules and SOP closed-loop optimization module is used to parse the final root cause determination result matrix. isolated sandbox memory area When the backtracking test converges safely, update the static initial baseline constraint vector. .

9. A computer device comprising at least one processor coupled to at least one memory storing at least one computer program or instruction, characterized in that, The computer program or instructions are loaded and executed by the processor to implement the steps of the adverse event analysis system and method based on the SOP system as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions, which, when executed by a processor, implement the steps of the adverse event analysis system and method based on the SOP system as described in any one of claims 1 to 7.

Images