An intelligent auditing system based on sample analysis data of a LIMS system

Abstract

This invention relates to the field of laboratory information management and data processing, and discloses an intelligent auditing system based on sample analysis data from a LIMS system. The system includes an underlying rule parameter library, a data acquisition and timing scheduling module, a basic auditing rule calculation module, a rule conflict resolution and confidence fusion module, a tiered alarm and closed-loop intervention module, and an automatic decision-making interception module. The data acquisition and timing scheduling module utilizes an asynchronous suspension and joint wake-up scheduler to perform timing alignment; the basic auditing rule calculation module calls operators to perform multi-component and cross-sample conversion rate calculations for the same sample; the rule conflict resolution and confidence fusion module uses logarithmic mapping to generate a comprehensive anomaly confidence score; and the automatic decision-making interception module compares the confidence score with a dynamic threshold, generates a status code object, and performs data table status locking, achieving accurate issuance of review work orders through a fit score. This solves the problems of verification timing misalignment and rule conflicts, improving the accuracy of anomaly interception.

Description

Technical Field

[0001] This invention relates to the field of laboratory information management and data processing, specifically to an intelligent auditing system based on sample analysis data from a LIMS system. Background Technology

[0002] In existing Laboratory Information Management Systems (LIMS), the review of sample analysis data typically relies solely on simple static upper and lower limit judgments, making it difficult to identify data that falls within the normal quality indicator range but exhibits abnormal fluctuations. In real-world testing scenarios, the generation and entry of sample analysis data are often asynchronous, with significant time differences between different test components or upstream and downstream samples. Existing systems lack effective data time-series alignment mechanisms, and directly applying correlation rule checks to incomplete data can easily lead to failed correlation calculations or frequent false alarms.

[0003] When multiple rules are introduced to compensate for the shortcomings of a single lower limit judgment, existing systems often output discrete, multi-dimensional deviation values ​​that may have conflicting physical meanings. This approach leads to individual rules triggering alarms independently, causing significant interference to the reviewers' judgment and failing to provide objective, comprehensive quantitative indicators of anomalies. Due to the lack of automated multi-dimensional judgment algorithms, the current review of data anomalies within the quality indicator range can only be conducted manually by technical personnel afterward, retrieving nearly ten batches of historical data and combining it with their personal experience. This process is labor-intensive and inefficient.

[0004] Existing systems often employ static, fixed standards for interception, making them ill-suited to the dynamic changes in production stability. Furthermore, the lack of automated data freezing mechanisms and review work order assignment mechanisms after manual detection of data anomalies results in a high reliance on manual intervention for abnormal data, leading to a prolonged overall workflow. This model, dependent on post-event manual investigation and static verification, severely restricts the efficiency and accuracy of inspection and auditing work. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent review system for sample analysis data based on a LIMS system. This system solves the problems of existing LIMS systems, which often rely on static thresholds and single-dimensional manual review methods when processing sample analysis data. These methods struggle to handle asynchronously generated sample data with complex temporal dependencies. Furthermore, when faced with parallel verification of multiple testing indicators, they are prone to rule conflicts and alarm storms, lacking a confidence assessment mechanism that dynamically adjusts based on production conditions, resulting in low data review efficiency and high false alarm rates. In addition, existing systems lack automated interception methods and graphical traceability explanations after anomalies are detected, failing to achieve intelligent closed-loop control for anomaly diagnosis and reassignment of data for review.

[0006] To address the above problems, the present invention provides the following technical solution:

[0007] This invention provides an intelligent auditing system based on sample analysis data from a LIMS system, comprising an underlying rule parameter library, a data acquisition and time-series scheduling module, a basic auditing rule calculation module, a rule conflict resolution and confidence fusion module, a hierarchical alarm and closed-loop intervention module, and an automatic decision-making and interception module.

[0008] The underlying rule parameter library constructs a multidimensional parameter mapping table using the sample inspection method standard number, product brand, and device tag number as joint primary keys. This table stores parameters required for logical verification, such as rule dependencies, conditions and limits, and alarm classification thresholds. The system monitors the submission payloads of external rule configuration interfaces through version control and hot update components, parses and generates incremental update scripts, and performs lock-free reload operations on the multidimensional parameter mapping table during verification intervals. This allows parameters to dynamically take effect without interrupting the current system review process.

[0009] The data acquisition and timing scheduling module monitors the LIMS business database redo logs based on a change data capture mechanism, acquires sample analysis data, and extracts metadata including timestamps. The system constructs a finite state machine based on in-memory computing, and assigns initial, suspended, and verifiable states to the analysis data based on rule dependencies. When the set of associated identifiers corresponding to the data matches and the difference in sampling timestamps is less than the operating condition alignment tolerance threshold, a joint wake-up instruction is triggered to update the state of the analysis data to the verifiable state. Simultaneously, the dynamic timeout degradation decoupling engine periodically polls the suspended state data, and when the actual dwell time exceeds the tolerance threshold, the state of the analysis data is updated to the verifiable state. This solves the verification blocking problem caused by missing dependencies in asynchronously generated data, ultimately outputting concatenated and encapsulated integrated data.

[0010] The basic audit rule calculation module receives and integrates data, extracts a set of baseline configuration parameters based on conditions and limits, calls the maximum value operator to perform denominator boundary protection, and distributes safe computational load objects. The system generates individual deviation values ​​through non-correlation rule calculation; it performs numerical interpolation and summation through correlation rule calculation with the same sample, outputting the total concentration deviation and the deviation of primary and secondary auxiliary indicators; it calculates the absolute deviation of upstream and downstream sampling timestamps through correlation rule calculation with different samples, and activates the material conversion rule calculation channel to output the conversion rate deviation when the deviation meets the conditions. The system merges and encapsulates the deviations from these different dimensions to output a deviation dataset, achieving multi-dimensional deep verification of complex operating condition data.

[0011] The rule conflict resolution and confidence fusion module extracts the rule baseline weights and amplification factors from the underlying rule parameter library, infers the initial dynamic weights using a multilayer perceptron model, generates effective rule baseline weight values ​​using Hadamard product operations, and calls the global normalization operator to output a structured weight matrix object. The system receives the deviation dataset and the weight matrix object to construct a computation array, and performs fusion calculations through a parallel dual-channel architecture: the first channel performs maximum deviation extremum bypass extraction; the second channel calls the logarithmic mapping engine to transform the computation array into the natural logarithm space for addition dimensionality reduction calculation, achieving nonlinear multiplication and generating a joint anomaly confidence score. The system compares the extracted maximum deviation scalar with the joint anomaly confidence score, extracting the maximum value as the comprehensive anomaly confidence score, thus solving the weight allocation and conflict resolution problems when multiple rules are triggered in parallel.

[0012] The automatic decision-making interception module extracts production stability features, generates a baseline dynamic threshold through a feedforward neural network, and outputs the dynamic threshold after being limited and pruned by a boundary constraint operator. The system compares the comprehensive anomaly confidence level with the dynamic threshold to generate a status code object. When the status code object contains an automatic interception or manual review identifier, the system initiates a status freeze command to lock the sample analysis data status through a built-in database transaction lock mechanism. After locking, the dynamic dispatch solver polls and calculates the dynamic load equivalent index and comprehensive fit score of online personnel nodes, dispatches review work order objects to online personnel nodes whose comprehensive fit scores reach the maximum value, and thus triggers the review workflow. The system extracts the same-site measured dataset based on the sampling time and calculates and generates dynamic baseline control upper and lower limits. The dynamic baseline control upper and lower limits and the same-site measured dataset are packaged to generate a comprehensive visual rendering matrix message. Subsequently, it is fused with the structured diagnostic prompt string generated by the causal rule encoding mapping as an audit conclusion output containing graphical tracing explanation, providing direct evidence for manual review.

[0013] The tiered alarm and closed-loop intervention module receives the comprehensive anomaly confidence score, extracts the alarm tiering thresholds to construct a stepped threshold interval matrix, compares the comprehensive anomaly confidence score with each alarm tiering threshold in the matrix, and outputs a discretized alarm level scalar. The system extracts the bound response strategy template to generate a task instruction containing message push channel identifiers and approval node flow blocking identifiers, pushes it into the system task scheduling queue for issuance, and executes the tiered alarm and closed-loop intervention.

[0014] This invention provides an intelligent review system based on sample analysis data from a LIMS system. It has the following beneficial effects:

[0015] 1. This invention introduces an asynchronous suspension and joint wake-up scheduler based on memory computing to allocate state machine nodes to asynchronously generated sample analysis data. By comparing the absolute difference of sampling timestamps to trigger a joint wake-up command, it solves the problem of rule verification failure or false alarm caused by the time difference of data entry for multiple components or upstream and downstream in traditional LIMS systems. This ensures that the correlation rule calculation for the same sample or across samples is executed under the condition of data time sequence alignment, thus guaranteeing the accuracy of basic data review.

[0016] 2. The rule conflict resolution and confidence fusion module of this invention combines the dynamic weights and adaptive amplification factors derived by the multilayer perceptron to perform Hadamard product operations, and uses logarithmic mapping to transform the calculation array into the natural logarithmic space for dimensionality reduction and fusion. This transforms discrete, multidimensional deviation values ​​with conflicting physical meanings into a single comprehensive anomaly confidence level, objectively quantifying the overall deviation of the sample data and avoiding the judgment interference caused by multiple single rules independently triggering alarms.

[0017] 3. This invention utilizes an automatic decision-making interception module. It employs a feedforward neural network to generate dynamic thresholds based on production stability characteristics for numerical comparison, and calls a database transaction lock mechanism to freeze the state of abnormal data. A dynamic dispatch solver calculates the dynamic load equivalent index and comprehensive fit score of online personnel for node optimization. Combined with intercepted historical datasets of the same location, it generates dynamic baseline control limits and outputs audit conclusions with graphical traceability explanations. This achieves automated processing from anomaly identification and data interception to accurate issuance of review work orders, reducing reliance on human experience for abnormal data intervention, providing intuitive decision-making basis for subsequent manual operations, and shortening the business cycle. Attached Figure Description

[0018] Figure 1 This is a diagram illustrating the overall architecture of the intelligent verification system for sample analysis data in the LIMS system of this invention.

[0019] Figure 2 This is a data flow logic diagram of the asynchronous suspension and joint wake-up scheduler of the present invention;

[0020] Figure 3 This is a diagram of the underlying computational architecture of the nonlinear multiplication solver for comprehensive anomaly confidence in this invention.

[0021] Figure 4 This is a schematic diagram of the dynamic threshold adaptive evolution and anomaly capture trajectory of the present invention;

[0022] Figure 5 This is a schematic diagram illustrating the comparison of the effects of large-scale historical data backtracking in this invention.

[0023] The module includes: 1. Data acquisition and time-series scheduling module; 2. Basic audit rule calculation module; 3. Rule conflict resolution and confidence fusion module; 4. Hierarchical alarm and closed-loop intervention module; 5. Underlying rule parameter library; and 6. Automatic decision-making and interception module. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see the appendix Figure 1 and Figure 2 This invention provides an intelligent review system based on sample analysis data from a LIMS system. This intelligent review system is embedded in a laboratory information management system and is used to perform data processing and flow control on the entered sample analysis data.

[0026] The intelligent auditing system based on LIMS system sample analysis data includes a data acquisition and time-series scheduling module 1, a basic auditing rule calculation module 2, a rule conflict resolution and confidence fusion module 3, a hierarchical alarm and closed-loop intervention module 4, a low-level rule parameter library 5, and an automatic decision-making and interception module 6.

[0027] The data acquisition and timing scheduling module 1 receives sample analysis data and corresponding metadata entered by laboratory personnel at the front end of the laboratory information management system. The metadata includes component type, sampling time, and entry timestamp. The data acquisition and timing scheduling module 1 communicates with the underlying rule parameter library 5 to obtain the rule dependencies corresponding to the currently entered data.

[0028] The data acquisition and timing scheduling module 1 determines the verification timing of the currently entered data based on dependencies. For irrelevant data that does not depend on other data, the data acquisition and timing scheduling module 1 sends it directly to the basic audit rule calculation module 2. For relevant data with dependencies, the data acquisition and timing scheduling module 1 performs a status suspension or wake-up operation based on the entry timestamp. When the associated data record is complete or the set timeout condition is met, the data acquisition and timing scheduling module 1 sends the integrated data to the basic audit rule calculation module 2.

[0029] The basic audit rule calculation module 2 receives the scheduled data and calls the calculation conditions and parameter limits in the underlying rule parameter library 5. The basic audit rule calculation module 2 performs parallel calculations of irrelevant rules, similar-item relevance rules, and different-sample relevance rules. The basic audit rule calculation module 2 maps the condition judgment results of different business rules to deviation values ​​within a continuous interval.

[0030] The basic audit rule calculation module 2 generates a dataset containing multiple deviation values ​​and outputs the dataset to the rule conflict resolution and confidence fusion module 3.

[0031] The rule conflict resolution and confidence fusion module 3 receives the aforementioned deviation value dataset and extracts the rule weight coefficients and nonlinear amplification factors configured in the underlying rule parameter library 5. The rule conflict resolution and confidence fusion module 3 uses a nonlinear multiplication algorithm to mathematically calculate the deviation values, generating a comprehensive anomaly confidence score for the current batch of analysis data. The rule conflict resolution and confidence fusion module 3 then sends the calculated comprehensive anomaly confidence score to the hierarchical alarm and closed-loop intervention module 4.

[0032] The graded alarm and closed-loop intervention module 4 receives the comprehensive anomaly confidence level and compares it with the warning threshold and interception threshold set in the underlying rule parameter library 5. Based on the threshold comparison result, the graded alarm and closed-loop intervention module 4 sends corresponding status task instructions to the main control program of the laboratory information management system. Task instructions include data normal release instructions, interface auxiliary prompt instructions, and data automatic interception instructions. After receiving the task instructions, the laboratory information management system executes the corresponding interface rendering and data table status locking operations to complete the system control for sample analysis data review.

[0033] In this embodiment, the data acquisition and timing scheduling module 1 integrates an analysis data and metadata monitoring and extraction unit. This unit serves as the traffic entry point for the entire auditing system, and its technical principle is based on real-time parsing and asynchronous capture of the underlying data communication protocol of the laboratory information management system. To ensure the immediacy and non-intrusiveness of data capture, the analysis data and metadata monitoring and extraction unit is physically deployed between the application layer and the database layer of the laboratory information management system, achieving deep packet inspection of data packets through the construction of a logically transparent traffic gateway.

[0034] The data analysis and metadata monitoring and extraction unit acquires data awaiting review from the laboratory information management system through a data interception mechanism. This mechanism is based on database change data capture technology, obtaining data update trajectories by monitoring database redo logs. As a preferred approach, this unit is also equipped with a Hypertext Transfer Protocol (HTTP)-based interception plugin, specifically designed to intercept structured query language (MQP) requests submitted by laboratory personnel via the front-end page.

[0035] After detecting a data submission action, the data analysis and metadata extraction unit extracts basic analytical data and associated metadata from the network request payload or database cache. Basic analytical data consists of specific physicochemical index values ​​output by laboratory instruments or manually entered. Associated metadata consists of structured system tags describing the attributes of the basic analytical data. In this embodiment, the extracted associated metadata includes not only the sample's unique identifier and component type code, but also independent sampling timestamps. With the entry timestamp Sampling timestamp A production condition benchmark reflecting physical samples, used to resolve time window alignment issues in correlation rules for different samples; timestamp entry. This reflects the physical order in which data enters the system and is the sole logical benchmark for subsequent asynchronous suspension and wake-up scheduling.

[0036] After acquiring the aforementioned data, the data analysis and metadata listening and extraction unit performs data cleaning and type conversion operations. Due to differences in the output precision of different analytical instruments, the unit calls a preset normalization algorithm to map the basic analytical data with different units of measurement to standard floating-point values. When dealing with extreme values, the unit automatically detects any outliers with zero denominators or data overflow, marking them as logically bad values ​​to prevent subsequent calculation crashes.

[0037] The data analysis and metadata listening and extraction unit further performs clock alignment processing on the multi-source data to ensure that data collected from different terminals all use a unified system standard clock. Based on the above preprocessing logic, the data analysis and metadata listening and extraction unit encapsulates the transformed basic analysis data and its corresponding associated metadata into a standard data object with a unique hash value, and transmits the standard data object to the subsequent scheduling execution unit in the data acquisition and timing scheduling module 1, thereby supporting the parallel invocation of subsequent multi-dimensional rules.

[0038] In this embodiment, the data acquisition and timing scheduling module 1 further includes an asynchronous suspension and joint wake-up scheduler. Based on the standard data objects output by the preceding analysis data and metadata listening and extraction unit, this scheduler is used to solve the problem of joint verification logic breakage caused by asynchronous cross-submission of data from different components in actual laboratory production. The core underlying architecture of the asynchronous suspension and joint wake-up scheduler is a finite state machine built based on in-memory computing. This state machine dynamically assigns an independent lifecycle state identifier to each entered data object, and the specific lifecycle states include an initial state, a suspended state, and a verifiable state.

[0039] The scheduler receives the currently entered data item Then, a query request will be sent to the underlying rule parameter library 5 to extract the set of associated identifiers corresponding to the data item. This set contains the feature identifiers of other components to be entered that have a logical dependency on the current data item. Based on these dependencies, the scheduler systematically determines the flow path of the current data. As a preferred method, the scheduler verifies... Whether or not the set is empty determines whether the interception mechanism is triggered. If the set is empty, it indicates that the data item only involves irrelevant rule calculations of a single dimension. The scheduler immediately updates its status to verifiable and pushes it directly to the basic audit rule calculation module 2 to perform subsequent mathematical calculations.

[0040] When the judgment When the data stream is a non-empty set, the system confirms that the current data stream involves major and minor components or cross-sample co-correlation rules. At this point, the scheduler extracts the entry timestamp of the current data item. and sampling timestamp The system then launches an association item retrieval engine in the system's shared suspended cache pool. To ensure strict consistency of multi-source data under physical conditions and analytical backgrounds, and to avoid erroneous coupling across batches of data, this retrieval engine performs rigorous multi-dimensional joint verification. Specifically, the retrieval engine not only compares the association identifiers contained in the data objects, but also extracts the sampling timestamp of the current data item. Sampling timestamps of historical data in the cache pool And calculate the absolute difference between the two. .

[0041] Based on this, the system determines the absolute difference. Is it less than the preset working condition alignment tolerance threshold? The alignment tolerance threshold for this operating condition. The specific value depends on the physical interval between laboratory sampling batches, and is usually set as the maximum time fluctuation deviation within the same inspection cycle. This applies if and only if the associated identifier matches and the time constraint condition is met. The search engine determines that the expected paired data item exists only when both conditions are met. If the search results indicate that the expected paired data item has not yet been submitted to the laboratory information management system, the scheduler will truncate the current data item from the system's underlying layer. The scheduler switches the current data item's state flag to suspended and moves it to the asynchronous suspended queue for dwelling. While performing the dwelling operation, the scheduler writes the absolute system time of this state transition to the database log as the suspension timestamp. .

[0042] During the continuous operation of the system, whenever a new standard data object enters the data acquisition and timing scheduling module 1, the scheduler synchronously activates the bidirectional matching protocol. Based on the dual verification logic of condition alignment and identifier matching, once the newly extracted data features are fully verified with a historical data item in the asynchronous suspended queue, the scheduler immediately generates a joint wake-up instruction. This instruction causes the state machine to simultaneously transition the data items of both pairs from the suspended state to the verifiable state. After completing the state transition, the scheduler, based on a unified association identifier, performs logical alignment of memory addresses and data packaging on the multi-source heterogeneous data in the verifiable state, and then transmits the entire dataset to the basic audit rule calculation module 2.

[0043] Regarding the aforementioned control flow, this control logic, which triggers a suspension and wait based on the first data entry and a joint wake-up based on the subsequent data entry, enables the system to autonomously complete the condition alignment and time sequence splicing of multi-dimensional data in the background without interfering with the analyst's original fragmented data entry behavior.

[0044] In this embodiment, the data acquisition and timing scheduling module 1 also integrates a dynamic timeout degradation decoupling engine. To address abnormal situations in laboratory production where associated data is permanently lost due to sample damage, missed tests, or instrument malfunctions, this engine is configured as a system-level safety boundary mechanism that asynchronously suspends and jointly wakes up the scheduler to prevent logical deadlocks in the data processing flow.

[0045] Based on the underlying thread pool architecture, the engine maintains a background daemon process that periodically polls all data items in the asynchronous suspension queue at a preset time frequency. For each data item in the queue that is in a suspended state, the engine extracts the suspension timestamp recorded in the database when it entered the suspension queue. And simultaneously obtain the standard real-time clock time of the server operating system. To avoid computational anomalies caused by time rollback due to network time protocol synchronization in a distributed server cluster, the following... and The engine uses a timer timestamp, which is monotonically increasing at the operating system's underlying level, for measurement. Based on this time variable, the engine calculates the actual dwell time of the data item in the suspension pool. The technical purpose of this calculation process is to quantify the physical time cost of waiting for related components for one-dimensional data, thereby avoiding system memory overflow or indefinite stagnation of subsequent data report generation.

[0046] After determining the actual dwell time, the dynamic timeout degradation decoupling engine will determine the dwell time. The maximum tolerance time threshold configured in the underlying rule parameter library 5 Compare this to the maximum tolerance time threshold. The specific value is set based on the laboratory's physical shift change cycle or the highest timeliness requirement for issuing test reports, and is typically configured as eight hours or twenty-four hours. To avoid business blockage caused by the system relying solely on a single time extreme value, the engine also incorporates a status linkage mechanism with the external production scheduling system in its judgment logic. Specifically, while comparing the time threshold, the engine queries the physical flow status of the missing sample through an interface. As a preferred approach, when the actual dwell time is met... If the external system reports that the physical state of the expected associated sample is damaged or obsolete, the engine will no longer execute the single extreme value waiting strategy. Instead, it will comprehensively determine that the current asynchronous suspended link has been irreversibly broken, and then immediately trigger the system-level degradation and decoupling mechanism.

[0047] After triggering the degradation decoupling mechanism, the engine forcibly takes over the flow control of the suspended data item and performs a state unbinding operation. The engine modifies the state machine attributes, directly and forcibly changing the lifecycle state of the data item from the suspended state to the verifiable state. To support the self-consistency of the operation logic when performing parallel computation of missing related data, the engine further dynamically injects a degradation verification flag into the metadata structure of the data item. This degradation verification flag, as a specific Boolean control variable, is used to issue explicit operation constraint instructions to the downstream system. When the basic audit rule calculation module 2 reads that the flag is true, its internal multiplexer will automatically maintain the opening of the non-related rule calculation channel, while forcibly blocking all related rule calculation channels involving the expected missing related items. More importantly, the basic audit rule calculation module 2 will forcibly assign the deviation output value corresponding to these blocked channels to the value of zero, thereby ensuring the completeness and logical closure of the subsequent nonlinear multiplication formula at the mathematical level, completely avoiding the underlying division by zero or abnormal error caused by input null values. While dynamically injecting the degradation verification flag, the system verifies the business level of the data object. If the data object belongs to a strongly coupled item or a key control point (A-level indicator) that only depends on the correlation rule, the engine will no longer force the zero assignment, but will directly trigger the abnormal locking of the data status and generate an incomplete work order for manual processing; the above zero assignment self-consistent logic will only be activated for regular indicators that allow independent verification.

[0048] After completing the state change and control flag injection, the engine packages the downgraded standard data object and pushes it to the basic audit rule calculation module 2 for independent mathematical verification. Based on the cache management mechanism, the engine synchronously executes memory cleanup instructions to completely release the memory address space originally occupied by the data item in the shared suspended cache pool.

[0049] In this embodiment, the basic audit rule calculation module 2 embeds a rule parameter configuration and detection limit constant truncation. This truncation, as a pre-processing core unit before data enters the specific mathematical calculation channel, is responsible for dynamically matching benchmark calculation parameters for various analytical data and executing mathematical boundary protection of the underlying architecture. Based on the standard data object input from the data acquisition and timing scheduling module 1, the truncation parses the component type code and sample unique identifier code in its associated metadata. To adapt to the heterogeneous testing environment of multiple concurrent and multiple analytical methods in the laboratory, the truncation establishes a concurrent connection with the underlying rule parameter library 5 based on the aforementioned structured tags, extracting the benchmark configuration parameter set required for the calculation of the current batch of data. This parameter set specifically includes the upper and lower limits of process control for the corresponding components, the historical sliding window data queue, the preset volatility limit, and logical expressions for various correlation judgments.

[0050] In actual laboratory operations, the test results for some trace components often show extremely low values, sometimes even being marked as zero by laboratory personnel when undetectable. Meanwhile, the raw data received by the laboratory information management system includes not only continuous physicochemical values ​​but also qualitative text descriptions. If the system indiscriminately imports all extreme values ​​or text objects directly into subsequent volatility or relative deviation ratio calculation models, it will inevitably trigger system-level errors such as division by zero exceptions or data type mismatches at the computer's underlying level. To address this algorithmic incompleteness deficiency, the truncation mechanism is configured with an underlying data type splitter and a detection limit constant truncation mechanism before data distribution.

[0051] Specifically, the truncate mechanism reads the data format tags from the standard data object. For data objects determined to be text-based, the truncate mechanism directly bypasses and forwards them to the text matching channel in the irrelevant rules, without executing subsequent numerical boundary protection logic. For each numerical basic analysis data or historical benchmark value scheduled to perform ratio calculations, the truncate mechanism calls the built-in maximum value operator to perform denominator safety protection. The truncate mechanism synchronously retrieves the instrument method detection limit constants bound to the current test component and the instrument used from the underlying rule parameter library 5. Due to the significant differences in detection accuracy among different laboratory tests, as a preferred approach, the underlying rule parameter library 5 maintains a multi-dimensional mapping data table. The cutoff dynamically obtains the constants specific to the current calculation channel through a dual-key addressing method using component type coding and instrument identification. This constant At the physical level, it characterizes the minimum concentration limit of a substance that can be reliably quantified by a specific analytical instrument or testing method; its actual value is strictly constrained by… The mathematical range of this value. The specific source of this value is the signal-to-noise ratio baseline data generated by the laboratory's periodic calibration and verification of various analytical instruments, which is pre-stored in a relational database by the system administrator.

[0052] When performing denominator boundary protection, the cutoff operator uses arithmetic logic... Complete the numerical substitution:

[0053] ;

[0054] in, Represents the raw basic analysis data currently input or the historical average baseline value calculated via a sliding window; This represents the safety calculation denominator after mathematical truncation. Based on this mathematical transformation, when the absolute value of the actual test result is lower than the instrument's physical detection limit, the system automatically calls a constant. This serves as a backup benchmark for subsequent ratio calculations. This technical processing not only completely avoids the risk of memory overflow caused by the denominator approaching zero in floating-point arithmetic at the computer level, but also gives the trace component data a reasonable tolerance for fluctuations in physical logic, preventing the system from making one-sided anomaly interception judgments based solely on microscopic extreme values.

[0055] After completing the mounting of basic configuration parameters and the safe truncation of the calculation denominator, the unit distributes the packaged and generated safe computing load object in parallel to the non-correlation rule calculation channel and the correlation rule calculation channel inside the module, thereby initiating the independent quantization process of multi-dimensional deviation.

[0056] In this embodiment, the basic audit rule calculation module 2 is internally configured with an irrelevant rule calculation unit. The system-level task of this unit is to receive the security computing load object output by the pre-truncation and perform parallel mathematical verification on the analysis data of a single sample and a single component based on an independent dimension. In order to achieve subsequent mathematical fusion of multiple rule results, the underlying design principle of this unit is to transform the hard interception rules or logical judgments under different business scenarios into deviation values ​​in the continuous numerical range [0, 1].

[0057] For continuous indicators, the non-correlation rule calculation unit performs numerical limit rule calculation. The technical purpose of this calculation logic is to quantify the severity of the deviation of the current basic analysis data from the established process specifications. The system extracts the upper limit values ​​of process control configured in the underlying rule parameter library 5. With process control lower limit The logic comparator inside the unit will analyze the current value. A parallel comparison is performed with the aforementioned limits. If the value is within the normal process range, then the requirement is met. The comparator outputs the first deviation. When a value exceeds the limit, the unit calls a linear penalty function to calculate the degree of deviation. This is to avoid the upper and lower limits of process control coinciding, which could lead to poor tolerance parameters. A value of zero can cause calculation errors, so the system introduces a system-level minimum constant into the denominator. Implement boundary protection. As a preferred method, this... The value is fixed at 10. -6 The specific calculation formula has been updated as follows:

[0058] ;

[0059] in, The specific boundary value exceeded by the current data, i.e., the direction of exceeding the limit. or ; It is a system-level minimum constant; This is the current analysis value; This is the process tolerance range parameter. The specific value of this tolerance parameter is set based on the process tolerance of the test components. Its physical meaning is the maximum absolute amplitude of data fluctuations that the production system can withstand; (built-in...) The operator is used to map the final calculation result to the maximum limit value to prevent numerical overflow in subsequent multiplication fusion operations based on the deviation.

[0060] For text-based data objects with qualitative descriptions, the irrelevant rule calculation unit activates the sensitive text matching rule calculation channel. To ensure compatibility of text entered from multiple heterogeneous terminals, the unit first calls the character parser before performing matching to convert the input text into a standard Unicode format and performs preprocessing logic to eliminate case differences and redundant spaces. It is an absolute value operator, ensuring that both positive and negative deviations will output a positive deviation penalty.

[0061] Based on the preprocessed clean text, the pre-compiled regular expression engine performs pattern matching against the sensitive dictionary embedded in the underlying rule parameter library 5. The sensitive dictionary pre-defines a set of feature strings representing abnormal states. When the matching engine captures a feature string, the unit triggers a state flag switching instruction and directly assigns the second deviation value. If no match occurs, then assign a value. This discrete mapping mechanism provides a standardized mathematical expression interface for qualitative test indicators that is consistent with quantitative indicators.

[0062] Based on the comparisons across the aforementioned independent dimensions, the system further integrates the historical timeline and activates the volatility model calculation channel to capture hidden anomalies that, while not exceeding the absolute upper and lower limits, have clearly deviated from the historical normal operating range. The unit extracts the current component's continuous value within the historical time window from the system cache. Calculate the arithmetic mean of the sliding sequence for each valid test value. with standard deviation The specific capacity of the historical time window. The sampling frequency is determined based on the production cycle, and its numerical range is typically configured to be the most recent ten to thirty sampling points. To quantify fluctuation anomalies, the unit performs calculations based on an improved standard score algorithm, the calculation formula of which is:

[0063] ;

[0064] in, This is the sensitivity adjustment coefficient, based on the statistical control chart theory of normal distribution, and its value is usually set between 2.0 and 3.0; This is the current analysis value; and These are the arithmetic mean and standard deviation of the moving series, respectively. The instrument method for detecting limit constants is passed to the pre-cutoff, and the constant is then used to detect the limit constant. As a compensation term injected into the denominator, this formula provides a natural margin of immunity to interference at the physical level for extremely low concentration components, and completely avoids the standard deviation caused by the absolute stability of historical test data at the computer level. The memory division by zero crash caused when the value is zero.

[0065] To further monitor the slow drift trend of the process system, the non-correlation rule calculation unit initiates the monotonic trend rule calculation channel in parallel. The unit extracts historical time series data, including the current test point. The trend assessment engine calculates the first-order difference values ​​between adjacent data points. If multiple consecutive difference values ​​maintain strict positive or negative sign consistency, the system determines that the test data exhibits a monotonic drift tendency. To ensure that the output results are strictly normalized to the aforementioned continuous interval, the trend deviation calculation formula is set as follows:

[0066] ;

[0067] in, This represents the cumulative number of continuously and strictly monotonically increasing or decreasing data points detected so far; This is a preset alarm tolerance step size threshold, calibrated by process experts based on the dynamic hysteresis characteristics of the production unit or storage tank, typically set to 5 to 7. When the actual number of continuous monotonic points... Reaching or exceeding the tolerance step size At that time, the operator automatically performs mathematical truncation, making The output is the highest confidence level.

[0068] Based on the four independent parallel computing channels mentioned above, the non-correlation rule calculation unit transforms the extracted single-dimensional business features into a structured set of deviation values. This set of deviation values ​​serves as the basic input parameter for subsequent comprehensive anomaly confidence fusion and is asynchronously distributed to the subsequent computing modules via the system's internal memory data bus.

[0069] In this embodiment, the basic audit rule calculation module 2 is internally configured with a product correlation rule calculation unit. Based on the data packaged object output by the asynchronous suspension and joint wake-up scheduler, the technical purpose of this unit is to mine and quantify the physicochemical contradictions between multiple heterogeneous test items originating from the same physical sample. To achieve this purpose, the system constructs a comprehensive verification space covering mass conservation, logical proportion, and neural network regression mapping, and uniformly maps multi-dimensional correlation judgments into continuous deviation values.

[0070] Before initiating specific correlation mathematical verification, the control flow gateway built into the same sample correlation rule calculation unit extracts the degradation verification flag from the metadata of the current data object. When this flag is detected as true, the system confirms that the current sample has irreversible missing correlation data. To maintain the logical closed loop of the underlying parallel computing framework, the control flow gateway forcibly blocks all correlation calculation channels within the current unit and directly outputs a value of zero to the subsequent modules as the deviation of each correlation rule. This underlying channel zeroing self-consistent logic completely eliminates system-level arithmetic anomalies caused by null input values ​​at the mathematical level. When this flag is false, the unit normally activates the multiplexed calculation channels.

[0071] For full-component analysis or assays with a clear quality loop, the sample correlation rule calculation unit activates the component summation rule calculation channel. This channel calculates the sum of the concentrations of all major and minor components in the same batch of samples to verify whether serious omissions or systematic calibration drift have occurred during the assay process. The unit extracts the associated components of the sample. Set of measured concentration values ​​of each component To ensure the physical validity of the addition operation, the unit calls the underlying dimension conversion operator before performing the accumulation. This operator parses the physical dimension identifiers attached to the test results of each component and, using mass fraction as the reference dimension, uniformly converts heterogeneous dimensions such as parts per million and volume ratio into a dimensionless mass proportion constant. After completing the dimension alignment, the unit substitutes the cleaned values ​​into the calculation formula:

[0072] ;

[0073] in, As a preferred approach, the target total constant is defined by the laws of physics. Configured to a normalized value of 100; The allowable cumulative analytical error margin is typically calibrated between 0.5 and 2.0 based on the overall uncertainty of the laboratory instrument. It is a system-level minimum constant. This represents the total number of components in the currently associated test object; Characterized by the algebraic summation of the values ​​of each washed component. This is achieved by combining the maximum operator with a constant. The system constructs a mathematical protection barrier at the underlying level, ensuring the denominator approaches zero. The calculated summation deviation... It quantifies the absolute severity of the disruption to the overall material balance.

[0074] Based on chemical reaction principles or physical inclusion relationships, some testing items exhibit strict numerical ratio constraints. To identify such logical paradoxes, the sample correlation rule calculation unit activates the physicochemical indicator ratio rule calculation channel. This channel extracts the main indicator data with inclusion relationships. With secondary and ancillary indicator data Taking the testing scenarios of total nitrogen and ammonia nitrogen as an example, the absolute value of secondary indicators cannot, in physical reality, exceed that of the primary indicator. To quantify this logical violation, the unit uses a calculation formula:

[0075] ;

[0076] The numerator portion incorporates a built-in maximum operator to filter out normal fluctuations consistent with physical logic, outputting a positive difference only when a secondary indicator illegally exceeds the primary indicator; the denominator portion utilizes... The calculation prevents memory crashes caused by the main indicator's measured value approaching zero. This ratio deviation... This provides a quantitative basis for subsequent identification of laboratory personnel who may confuse samples or fill in incorrect data.

[0077] To cover more complex nonlinear physical laws, the product correlation rule calculation unit is deployed in parallel with multivariate collaborative rule calculation channels. For the implicit coupling mapping between multidimensional physicochemical indicators, the system abandons the traditional hard-coded polynomial and instead constructs a working condition proxy model based on a multi-layer feedforward neural network. The network layers of this model specifically include an input layer responsible for receiving multidimensional test parameters, two hidden layers using linear rectified functions as activation functions, and a linear output layer for outputting predicted indicator values. Before performing inference, the unit extracts basic feature vectors such as temperature, pressure, and auxiliary component concentrations from the input dimension, and performs standard score standardization preprocessing on these feature vectors using the mean and variance parameters obtained offline to eliminate the influence of dimensional differences on the network weight gradient.

[0078] After feature preprocessing, the standardized data flows forward into the network, ultimately outputting the theoretical predicted value of the target attribute. The training process of this proxy model is conducted offline based on the historical valid test dataset accumulated over a long period of time by the laboratory information management system. The system uses physically related auxiliary component data as sample input and the corresponding actual measured values ​​of the target components as supervision labels. The training engine uses mean squared error as the loss function and iteratively updates the network weights through a gradient descent algorithm based on the objective moment estimation until the loss function curve on the validation set converges to below a preset threshold.

[0079] In obtaining theoretical prediction values The unit then compared the theoretical prediction with the actual measurement submitted by the laboratory. The deviation formula for mathematical comparison is:

[0080] ;

[0081] in, The tolerance range parameter for the multivariate prediction of this proxy model under specific operating conditions is set by the system administrator based on the confidence interval of the historical modeling residuals; As an absolute value operator, it ensures that the model's predicted values ​​can consistently deliver risk representation values ​​to subsequent stages, regardless of whether the deviation is positive or negative.

[0082] Based on the correlation verification of the above three dimensions, the product correlation rule calculation unit also completed the logical cross-recognition between different test indicators within the same physical entity. The unit encapsulates the calculated correlation deviation set and outputs it synchronously to the subsequent data fusion and comprehensive evaluation network through a shared memory stack.

[0083] In this embodiment, the basic audit rule calculation module 2 is also equipped with a different sample correlation rule calculation unit. Based on the cross-sample alignment data delivered by the asynchronous suspension and joint wake-up scheduler, the technical purpose of this unit is to mine and quantify the abnormal evolution of physicochemical logic between different physical entities during spatial location or operational process. Unlike the comparison of identical samples which focuses on internal attributes, the correlation calculation of different samples focuses more on verifying the material conservation and reaction transformation laws in the upstream and downstream of the production unit.

[0084] Before performing specific cross-sample mathematical derivations, the system must confirm the physical causality of upstream and downstream multi-source data. In continuous production processes, there is an objective physical lag time between the material flow from the inlet to the outlet. If this temporal difference is ignored and concentration comparison is performed directly, it will inevitably lead to a misalignment between the data label and the actual physical batch. Based on the above physical constraints, this unit extracts the sampling timestamp of the upstream sample. Sampling timestamps of downstream samples Based on the calibrated residence time configured for a specific production unit in the underlying rule parameter library. The system calculates the absolute deviation between the actual stay time and the theoretical stay time. This is used to determine whether two batches of samples belong to the same physical transfer batch. The deviation is only considered if it is less than a preset operating condition tolerance window. Only then will the unit activate the subsequent associated mathematical verification channel. This time tolerance window The specific value range is usually configured to be five to fifteen minutes, and is calibrated by process experts based on the severity of liquid level fluctuations in the device.

[0085] For upstream and downstream correlated samples that meet the spatiotemporal alignment conditions, the correlation rule calculation unit for different samples activates the material transformation rule calculation channel. This channel calculates the concentration change rate of the target component before and after the reactor or separation tower to verify whether the analytical data violates the objective physical removal or transformation capacity of the process unit. The unit extracts the measured value of the upstream feed concentration. Compared with the measured value of downstream discharge concentration To quantify conversion performance, the actual conversion rate is calculated per unit:

[0086] ;

[0087] Among them, the denominator introduces a system-level minimum constant. As a preferred method, this The value is fixed as This approach completely avoids the risk of memory division by zero caused by extremely low upstream feed concentration at the computer level. Based on the calculated actual conversion rate, the system compares it with the target conversion rate. The formula for calculating the deviation is as follows:

[0088] ;

[0089] in, This is the allowable conversion rate fluctuation range parameter, typically ranging from 2% to 10%, depending on the catalyst's lifespan or the unit's load conditions. This deviation... It can effectively reveal the logical paradox of an abnormally higher discharge concentration than feed concentration due to sampling confusion or testing errors.

[0090] For parallel samples or retained retest samples set up within the laboratory to verify operational repeatability, the unit deploys a parallel consistency rule calculation channel. The technical purpose of this channel is to quantify the measurement precision of the same physical batch across different test operations. The unit extracts the first test value generated from the subsampling of the same original sample. Compared with the second test value To assess its absolute differences, the unit was calculated using a formula based on laboratory quality control standards:

[0091] Quantification of execution;

[0092] in, This is the legally permissible limit for parallel differences in this laboratory test, and its value is directly derived from the national standards or industry specifications regarding the repeatability of specific testing methods. Based on the above calculation logic, when the difference between two test results exceeds the legally permissible boundary, the deviation is... This will be amplified, thus providing highly confident single-dimensional evidence for subsequent comprehensive interception.

[0093] To address the quality balance verification at complex intersections of multi-inlet and multi-outlet pipeline networks, the unit further introduces a material conservation calculation channel based on the node matrix. For those involving... Each feed branch and The unit extracts the flow weight vector of the feed branch from the confluence node of each discharge branch. Vector of feed component concentration and the flow weight vector of the discharge branch. Vector of effluent component concentration Based on the law of conservation of mass, the system constructs the mass balance equation for the merge nodes and calculates the mass residual vector:

[0094] ;

[0095] When evaluating the multi-source cooperative equilibrium state, the unit needs to solve the system residual covariance matrix. The inverse matrix. To prevent the covariance matrix from exhibiting singularity and becoming invertible due to perfect linear correlation of flow data in certain branches, the unit forcibly invokes the Tikhonov regularization mechanism before executing the matrix inversion operator. This mechanism appends a tiny perturbation scalar to the diagonal elements of the covariance matrix. (Its value is usually set to) Multiply by the identity matrix This ensures that the system maintains full matrix rank even when faced with redundant sensors or collinear inputs. Based on the regularized inverse matrix, the unit uses the Mahalanobis distance metric algorithm to calculate the overall quality deviation, and its calculation formula is as follows:

[0096] ;

[0097] in, It is the transpose of a new vector formed by the residuals of multiple nodes; The node matrix tolerance threshold is statically calibrated based on the critical value of the chi-square distribution table at a 95% confidence level. This mathematical boundary protection robustly supports the overall quality deviation of the system's output bus nodes. .

[0098] Based on the above multi-condition and cross-sample spatial correlation verification, the correlation rule calculation unit for different samples outputs a structured deviation value characterizing the contradictions in external physical flow. These cross-sample verification results will be merged into the memory data bus along with the internal deviation output by the front-end unit.

[0099] In this embodiment, the rule conflict resolution and confidence fusion module 3 is equipped with a weight matrix and amplification factor configuration unit. The technical purpose of this unit is to receive the structured deviation set output from each of the preceding rule calculation channels and, based on the relative importance differences of different business rules under the current production conditions, dynamically assign influence weights to the deviations of each dimension. Simultaneously, for extreme anomalies that severely violate physical laws or process limits, the system constructs a nonlinear mapping mechanism to overcome the drawbacks of linear weighted averaging, ensuring that high-risk features directly dominate the final comprehensive confidence assessment.

[0100] To adapt to complex and ever-changing chemical production scenarios, traditional static weight allocation often struggles to accurately represent rule priorities under real-time operating conditions. Therefore, this unit incorporates a dynamic weight inference proxy model based on a multilayer perceptron. To extract nonlinear coupling relationships from multi-source heterogeneous data, the proxy model's network hierarchy specifically includes an input layer responsible for receiving multi-dimensional environmental features, two hidden layers using linear rectified functions as activation functions, and a network terminal using a normalized exponential function as the output layer. Before inference, the system extracts corresponding production environment feature sequences from the real-time database of the distributed control system (DCS) or manufacturing execution system (MES) on-site via a pre-built OPC-UA industrial communication protocol interface or microservice data bus. The unit also collects production unit operating parameters (including reaction temperature, system pressure, and material flow rate) strictly aligned with the current sample sampling timestamp, as well as the instrument's historical baseline drift rate, constructing a dimension-based model. The system uses the preset mean and standard deviation parameters to perform standard score standardization preprocessing on the feature vector, thereby eliminating the interference of heterogeneous dimensions on the gradient of the neural network weights.

[0101] The preprocessed feature vectors are fed into the surrogate model for forward propagation, where implicit nonlinear features between the working environment and rule sensitivity are extracted through the hidden layer. To avoid memory overflow in the exponential operation of the normalized exponential function at the lower level due to excessively large logarithmic feature values ​​output by the hidden layer under extreme industrial conditions, a numerical stability truncation mechanism is forcibly introduced within the output layer of the unit. This mechanism extracts the maximum scalar value in the logarithmic feature vector and subtracts this maximum value uniformly from each feature term before performing exponentiation, thereby generating a value more sensitive to the underlying conditions. The initial dynamic weight vector of each independently effective audit rule In this vector, This represents the total number of relevant and unrelevant rules activated in the current processing channel.

[0102] The offline training process of this model is based on a sample library extracted from historical manual review logs of the laboratory information management system. The system uses historical operating condition parameters as model input. To construct a benchmark target for supervised learning, the system extracts the attention scores of senior review engineers to various anomaly rules under corresponding operating conditions and converts them into a single-dimensional dataset. probability distribution label vector The constraint is that the sum of the probabilities of all dimensions is always equal to 1. The training engine uses relative entropy (KL divergence) as the loss function to rigorously measure the difference between the predicted weight distribution and the expert experience distribution. It then uses an adaptive moment estimation optimization algorithm to iteratively update the network parameters until the loss function curve on the validation set drops below a preset convergence threshold. This mechanism enables the system to have the business deduction ability to dynamically adjust the review focus under different working conditions, similar to that of human experts.

[0103] After obtaining the initial dynamic weights, to prevent multiple low-confidence regular fluctuations from overwhelming a single high-confidence fatal anomaly during mathematical fusion, an adaptive amplification operator is further introduced into the weight matrix and amplification factor configuration unit. This operator is tailored to the rule deviation of each input. (in ), calculate its corresponding nonlinear amplification factor independently. The specific calculation formula is set as follows:

[0104] ;

[0105] in, As a preferred approach, the value of the preset penalty activation threshold is typically set between 0.75 and 0.85. The gain factor is typically configured to range from 5.0 to 10.0. This is an exponential constant, typically set to 2 to construct a quadratic nonlinear mutation curve. When the deviation... The activation physical boundary was not reached. At that time, the built-in maximum value operator outputs zero, and the amplification factor... Maintain the baseline value; once the deviation exceeds the physical boundary, the system determines that the indicator has an extremely high risk of anomaly or instrument failure, and then increases its weight in the subsequent fusion model by exponentially amplifying the factor.

[0106] Based on the aforementioned dual-track parallel parameter configuration logic, the weight matrix and amplification factor configuration unit will configure the initial dynamic weights output by the surrogate model. With nonlinear amplification factor Perform Hadamard product operations to generate effective weight values ​​for each rule. To meet the numerical closed-loop requirements of subsequent probabilistic fusion calculation channels, the weight matrix and amplification factor configuration unit invoke the global normalization operator to compress the effective weight set. The normalization calculation formula is set as follows:

[0107] ;

[0108] Among them, by embedding a system-level minimum constant at the denominator end (Its value is fixed as) This computational logic completely prevents the division-by-zero memory overflow exception caused by all initial weights approaching zero at the computer's underlying level. After completing the above normalization calculation, the weight matrix and amplification factor configuration unit are packaged and output, containing the final weights of all rules. The structured weight matrix object is asynchronously distributed to subsequent comprehensive confidence calculation nodes via a high-concurrency memory data bus.

[0109] Please see the appendix Figure 3 In this embodiment, the rule conflict resolution and confidence fusion module 3 is equipped with a comprehensive anomaly confidence nonlinear multiplication solver. Based on the data load transmitted via the system's shared memory bus, the system-level task of this solver is to receive a multi-dimensional set of independent rule deviations and their corresponding dynamically normalized weight matrices, and to calculate a unique scalar value representing the overall failure risk of the current test data through rigorous probabilistic space mapping.

[0110] To accurately characterize the synergistic effects of multi-source anomalies, traditional linear weighted averaging algorithms suffer from significant mathematical flaws. In laboratory auditing scenarios, if a key indicator deviates to an extreme degree, the sample has essentially lost its analytical representativeness. Linear additive fusion mechanisms sum and average this extreme feature with a large number of low-value features within the normal range, thus forcibly diluting the mathematical confidence of severe anomalies. Based on this physical paradox, this unit abandons the conventional additive fusion architecture and instead constructs a dual-channel fusion architecture based on multidimensional independent probability distributions and extreme value bypass direct access. This ensures boundary protection triggering mechanisms for core extreme value features while accommodating multidimensional cross-validation.

[0111] Before performing a specific comprehensive simulation, the system aligns and extracts the currently analyzed sample from the memory queue based on the globally unique sample identifier. A vector of rule deviations With the corresponding final weight vector To address the issue of numerical underflow crashes that can easily occur when multiplying a large number of floating-point numbers in the (0, 1) range at the computer's underlying level, the unit does not directly execute multiplication instructions in the original linear space. Instead, it calls the logarithmic mapping engine to transform the computation array into the natural logarithmic space for addition dimensionality reduction.

[0112] Based on the above spatial transformation, the fundamental probability joint deduction formula adopted by the unit is set as follows:

[0113] ;

[0114] in, The joint anomaly confidence score is the output of the probability domain fusion, and its scalar value is strictly convergent within the continuous interval [0, 1]. For the natural constant The exponential reduction operator with base is used to perform the inverse mapping from logarithmic space to linear space; It is the natural logarithm function operator; This is a summation traversal operator for traversing the entire set of rules. It aims to prevent issues arising when a certain item deviates from its set value. The consequences of being absolutely equal to 1 The logarithmic unbounded overflow anomaly was detected by forcibly wrapping a maximum comparison operator at the input of the logarithmic function and introducing a system-level minimum constant. Perform bottom-level boundary protection. As a preferred approach, this constant... The value is fixed as follows: .

[0115] The core physical meaning of the above formula for multiplying logarithmic fields lies in treating the verification process of each independent rule as a reliability measure that the sample has not failed. The basic probability value of a single rule determining that the sample is in a normal state is defined as follows: Incorporating a weighted index that incorporates nonlinear amplification and operating condition sensitivity characteristics After adjustment, the overall system reliability of the sample is the weighted product of all independent measures.

[0116] To compensate for the mathematical thinning flaw in the aforementioned weighted geometric average algorithm when dealing with extreme anomalies with minimal weights, a parallel extreme value bypass channel is initiated using the integrated anomaly confidence nonlinear multiplication solver. This channel directly calls the max pooling operator to extract values ​​from the input rule deviation vector. Extracting the highest confidence scalar Subsequently, the unit performs a final boundary fusion calculation on the output of the probabilistic joint deduction channel and the output of the extreme value bypass channel. The calculation formula is as follows:

[0117] ;

[0118] in, This refers to the overall anomaly confidence level output by the final system, achieved by introducing... The system mathematically guarantees that, when any single term has an extremely high deviation from 1, regardless of how small the weight assigned to that dimension is, the final overall anomaly confidence level will remain constant. All of them will directly inherit the highest risk value and be forced to approach the limit confidence level, thus perfectly matching the safety net principle in industrial control that it is better to have a false alarm than an omission.

[0119] After calculating the target scalar value, the comprehensive anomaly confidence nonlinear multiplication solver will generate the comprehensive anomaly confidence. The metadata tags of the raw analysis data are then structurally encapsulated and packaged. This encapsulated object is then asynchronously pushed into the system's high-speed communication bus, awaiting extraction and consumption by the subsequent automatic decision-making interception module 6.

[0120] In this embodiment, the automatic decision-making interception module 6 is internally configured with a threshold configuration and a dynamic comparator. The technical purpose of this unit is to receive the unique scalar confidence level output by the pre-processed comprehensive anomaly confidence level nonlinear multiplication solver, and dynamically generate an adaptive decision boundary based on the fluctuation characteristics of the current production conditions and the sample business level. By rigorously comparing the real-time confidence level with the dynamic boundary, this unit outputs the final differentiated control instruction.

[0121] In conventional laboratory information management systems, hard-coded static thresholds are typically used as data release boundaries. This fixed strategy is prone to a significant increase in false alarm rates when facing chemical plant start-up and shutdown or large load changes due to the objective existence of process background fluctuations; while during stable production periods, static thresholds are too broad, leading to the failure to intercept minor anomalies. To balance the sensitivity and robustness of the audit, this unit abandons the single static cutoff mechanism and constructs a dynamic threshold calculation channel based on the extraction of production stability features.

[0122] Before executing the comparison logic, the unit needs to extract operating condition context features and deduce the dynamic threshold. To achieve a nonlinear mapping of operating condition parameters to tolerance boundaries, the system incorporates a threshold generation surrogate model based on a multi-layer feedforward neural network. This surrogate model's network layers specifically include an input layer responsible for receiving environmental features, three cascaded hidden layers, and a linear output layer that uses a sigmoid function to limit the output range. To prevent neuron death due to gradient vanishing during deep network training, all hidden layers uniformly use a leaky linear rectified function as the activation function.

[0123] During the inference phase, the unit acquires multidimensional environmental data in real time, strictly aligned with the current sample sampling timestamp. Regarding the relevance of feature selection, the unit extracts the device production load rate to characterize the severity of macroscopic material balance, extracts the sliding time window variance of the core reactor temperature to quantify the stability of microscopic reaction kinetics, and simultaneously extracts the sample's operational level code to determine the safety redundancy of mass offset. To address the issue that discrete operational levels cannot directly participate in numerical calculations, the system calls the one-heat coding operator to map the categorical variable into a sparse binary vector, and merges it with continuous operating condition parameters to construct a multidimensional input feature vector. Based on the statistics of historical valid data, the system uses the standard score mechanism to perform standardized preprocessing on the continuous variables in the feature vector, thereby stretching the distribution of the input space to a state of zero mean and unit variance, so as to eliminate the interference of heterogeneous dimensions on the network weight update.

[0124] The training process of the aforementioned threshold generation proxy model relies on an offline-extracted historical inspection and review database. The system uses historical working condition environment vectors as sample inputs and the actual release decision boundaries of senior quality inspection experts under corresponding working conditions (normalized and mapped to the [0, 1] interval) as supervision labels. The training engine uses root mean square error as the loss function to accurately measure the deviation between the predicted threshold and the expert's experience threshold, and iteratively updates the network weight matrix using an adaptive moment estimation optimization algorithm. When the loss function curve on the validation set smoothly converges to below a preset minimum value, the model completes training and is then deployed. This training mechanism ensures that the model can accurately simulate the dynamic adjustment strategy of human experts regarding anomaly tolerance during different fluctuation stages of the device.

[0125] After completing the forward propagation derivation, the unit output network generates the baseline dynamic threshold. To prevent the neural network from outputting excessively divergent or illogical out-of-control thresholds when faced with unknown extreme anomalies (i.e., out-of-distribution data that the model has never seen), the unit invokes the boundary constraint operator to perform hard-limiting pruning on the baseline threshold. The specific pruning formula is as follows:

[0126] ;

[0127] in, The final, effective, and controlled dynamic threshold; The minimum interception threshold allowed by the system is typically calibrated between 0.60 and 0.65 based on the inherent uncertainty baseline of the laboratory instrument. The highest permissible tolerance threshold represents the limit of risk that safe production can withstand, and is typically calibrated between 0.85 and 0.90. This two-way constraint formula not only mathematically guarantees the legal range of the dynamic threshold, but also ensures the controllability and safety of the system intervention boundary in engineering practice.

[0128] After obtaining a safe and controlled dynamic threshold, the unit initiates a multi-stage comparator array to measure the comprehensive anomaly confidence level of the input from the front-end module. Discretization is performed. To implement refined hierarchical control, the system does not adopt a simple binary, one-size-fits-all logic, but rather is based on dynamic thresholds. Downward offset of the preset warning margin Early warning thresholds for early intervention were constructed. As a preferred approach, this warning margin... The value range is configured to be 0.10 to 0.15, and is statically configured by the process engineer based on the quality deviation sensitivity of the specific test item.

[0129] Based on the aforementioned dual-track boundary, the dynamic comparator executes rigorous interval routing logic. When the calculated confidence level satisfies... When the system determines that the current test data is completely reliable and within the normal fluctuation range, the comparator directly outputs a status code object containing the automatic release flag; when the confidence level falls within... When the data falls within the gray area, the system determines that the data is on the edge of a suspicious fluctuation. The comparator outputs a status code object containing a manual review identifier and simultaneously triggers a memory snapshot to capture and highlight key deviation features that cause an increase in confidence. Once detected... The system determines that the current analysis data has a high-confidence risk of logical failure or instrument malfunction, and the comparator forces the output of a status code object containing an automatic interception flag. Regarding the floating-point logic gates and interrupt triggering mechanisms upon which this comparator array relies in its underlying hardware, those skilled in the art can use standard industrial control programmable logic devices for instruction set mapping. Its clock synchronization and level-flipping logic are well-known technologies in the field and will not be elaborated upon here.

[0130] Based on the aforementioned dynamic deduction and multi-level comparison logic, the threshold configuration and dynamic comparator complete a rigorous dimensionality reduction mapping from continuous probability values ​​to discrete control instructions. The generated structured status code object is then pushed to the system's routing and distribution bus to drive subsequent business flow loops such as database storage or abnormal work order dispatch.

[0131] In this embodiment, the automatic decision-making interception module 6 is also equipped with an auxiliary prompt and visual explanation output unit. In actual business scenarios, simple system interception commands lack intuitive judgment criteria, making it difficult to guide on-site laboratory personnel or process engineers in subsequent investigations. To provide decision-making transparency, the technical purpose of this unit is to reverse-analyze and map the abstract values ​​generated by the preceding multi-dimensional matrix calculation into physical causal explanations, while constructing a visual rendering data structure to assist manual review in the grayscale range.

[0132] To quantify the specific driving force of each independent rule on the final anomaly interception command, the unit initiates a feature attribution calculation operator based on a dual-channel fusion architecture upon receiving the status code object. Because the pre-processed confidence fusion employs a dual-track mechanism of joint probability inference and extreme value direct access, the system needs to perform differentiated attribution analysis for different trigger paths. The system then determines the final comprehensive anomaly confidence level. The source channel. If the confidence level comes directly from the extreme value bypass channel (i.e. Numerically equal to the maximum single-item deviation The unit determines that a single point of failure has occurred in the current system. Based on the above physical logic, the unit forcibly assigns a contribution index of 1.0 to the target rule that triggers the maximum value, and assigns a contribution index of 0.0 to all other accompanying rules. If multiple rules have the same deviation reaching the maximum value, the system will distribute the contribution index of 1.0 equally among these rules; or, based on the rule priority dictionary preset in the underlying rule parameter library 5, it will assign 1.0 to the single rule with the highest business priority. If the final confidence level comes from the probability joint inference channel, in order to decompose the mathematical weights of each rule, the unit calls the logarithmic domain attribution mapping formula:

[0133] ;

[0134] in, For the first The normalized contribution index of each review rule represents, in physical terms, the relative responsibility for the rule's contribution to the overall decrease in the survival probability of the sample data; the absolute value operator is used to maintain the non-negative scalar property of the output. This is to address the issue of all rules deviating from their normalized values. To address the extreme case where the sum of the logarithms on both sides of the denominator is zero when both sides are absolutely equal to 0, the system forcibly incorporates an outer maximum value comparison mechanism and a system-level minimum constant into the denominator of the division operator. As a preferred method, this Fixed value This mechanism avoids the risk of program execution overflow due to division by zero during the underlying memory allocation stage.

[0135] After obtaining the contribution index array of all rules, the unit performs dimensionality reduction and structured text generation for abnormal symptoms. The system then calculates the contribution index based on these criteria. The rules are sorted in descending order of their numerical values, and those with cumulative contribution exceeding a preset interpretation threshold are excluded. The former The threshold is dynamically determined from the cumulative summation result. As a preferred method, this interpretation threshold... It is usually calibrated to 0.80, while the number of items truncated is... The result is dynamically determined by the cumulative summation. Based on the extracted core causal rule encoding, the system queries the built-in trie of the testing business domain, mapping abstract mathematical variables to business description text. For example, the internal calculation vector is mapped to specific diagnostic language such as cross-sample material balance inversion or parallel sample range exceeding the standard, and combined with the corresponding measured concentration and dynamic threshold parameters, a structured diagnostic prompt string is generated.

[0136] To support multi-dimensional graphics rendering of the front-end interface, the auxiliary prompts and visual explanation output unit further constructs a spatiotemporal distribution matrix based on temporal states. For test items identified as the core cause, the unit traces back to the historical time span based on the current sample sampling time reference point. The system uses historical baseline data from the same location. To ensure the physical validity of the historical baseline calculation, the system cannot directly perform statistical analysis on the entire historical data set. The unit extracts the current sample's production load rate and uses this as a benchmark to determine if the load deviation in the historical dataset exceeds a preset tolerance. A hard rejection is performed on 5% of the samples (typically 5%) to ensure that the remaining historical sequences are in the same steady-state physical range as the current samples. At the same time, to prevent historical extreme value noise from interfering with the baseline, the cell uses a truncated averaging algorithm to remove outliers from the first 5% and last 5% of the sequence.

[0137] Based on the cleaned and aligned valid historical dataset, the system calculates the moving average of the historical baseline sequence. with standard deviation Based on the Shewhart control chart principle, dynamic control upper and lower limits are generated, and the calculation equation is as follows:

[0138] ;

[0139] ;

[0140] in, This is a sensitivity adjustment coefficient, controlled by the current process stability, and typically ranges dynamically between 2.0 and 3.0; the time span... The value range is typically configured to be between 24 and 72 hours. After completing the baseline calculation, the unit performs tensor concatenation of the measured coordinate pairs of the current abnormal sample, the historical time series coordinate set, the calculated control boundary vector, and the color gradient mapping encoding of the corresponding rule contribution (e.g., mapping high contribution indicators to a high-saturation warning color scheme) to generate a comprehensive visualization rendering matrix. Finally, the auxiliary prompt and visualization explanation output unit will generate structured diagnostic prompt text and a rendering matrix. The data is packaged into a standardized data packet payload and sent to the client of the laboratory information management system via an asynchronous message queue.

[0141] In this embodiment, the automatic decision-making interception module 6 is equipped with an automated interception and review dispatch execution unit at its end. The technical purpose of this unit is to receive the discretized status code object output by the front-end dynamic comparator, and combine it with the visual diagnostic message generated by the auxiliary prompting unit to perform data status change and business flow control in the underlying database of the laboratory information management system.

[0142] To address the high concurrency of data interaction between systems in industrial control networks, traditional laboratory systems experience time lags when abnormal status changes occur. This can easily lead to downstream manufacturing execution systems reading abnormal data that has not yet been confirmed or intercepted. To prevent the flow of abnormal test data, this unit incorporates a database transaction locking mechanism. When a status code object is parsed to contain an automatic interception or manual review flag, the system initiates a status freeze command. This command locks the target sample's record in the relational database using an exclusive lock, overwriting its lifecycle status field from "Pending Review" to "Abnormal Locked" or "Under Review." Through this atomic submission logic, the system prevents the sample data from being exposed to the external distribution bus, ensuring the consistency of data acquired by the production scheduling system.

[0143] After data freezing is complete, for data marked for manual review, the system needs to generate an anomaly traceability work order and assign it to a specific inspection or process engineer. To avoid localized personnel overload caused by static polling allocation, this unit constructs a dynamic assignment solver based on real-time load and skill matrix matching. Before performing specific assignment calculations, the unit polls the queue of currently logged-in laboratory personnel in real time and extracts... The real-time status characteristics of each online worker are used as the input vector. This addresses situations in industrial settings where specific skill groups are unattended during shift changes or night shifts (i.e.,...). In extreme cases, the system has a built-in empty queue degradation caching mechanism. When an empty queue is detected, the solver pauses dispatching calculations, pushes the review work order into the pending waiting queue, and triggers a global audible and visual alarm in the laboratory. If the suspension time exceeds the preset tolerance period (usually 30 minutes), the system escalates the work order to the laboratory supervisor's mobile terminal, thus ensuring that any abnormal interception can form a closed management loop.

[0144] After confirming that the online queue is not empty, the solver works for each online user in the queue. (in ), independently calculate its current dynamic load equivalent index This index aims to quantify the current pressure of concurrent tasks and the urgency caused by task backlog. The specific calculation formula is set as follows:

[0145] ;

[0146] in, For personnel The total number of historical tasks currently pending or being processed; For the first The baseline man-hour consumption weight for each existing task is statically determined by the business dictionary based on the complexity of the test project, and the value range is constrained to be between 1.0 and 5.0. For the first The queuing delay time of an existing task is calculated by extracting the difference between the current system's absolute timestamp and the task's initial dispatch timestamp, and then rounding it down to a scalar value in minutes to ensure the physical validity of subsequent multiplication operations. This is the timeout penalty coefficient. As a preferred method, this penalty coefficient... The value is configured to be 0.05. The built-in maximum value comparison operator is used to ensure that the non-linear expansion of the load equivalent is triggered only when existing tasks are queued.

[0147] After calculating the dynamic load equivalent of each online worker, the solver incorporates the worker's professional skill proficiency and calculates a comprehensive fit score. This determines the final work order dispatch target. The fit mapping formula is:

[0148] ;

[0149] in, For personnel The skill matching coefficient for the analytical method of the current anomalous sample is constrained to a continuous range of 0.1 to 1.0, and is typically automatically calibrated and updated periodically by the system based on personnel's historical training records and operational qualification rates. To meet the numerical robustness requirements of industrial software under extreme concurrent environments, a system-level minimum constant is explicitly embedded in the denominator of the division operator. As a preferred approach, this constant is fixed at... By introducing this minimal deviation value, the system completely prevents situations where an online user is in a completely idle state (i.e., existing load) at the underlying level. This is a division-by-zero overflow exception that occurs when the value is equal to 0.

[0150] After completing parallel computation of the entire queue, the solver is dynamically dispatched to call the outer sorting operator to extract the overall fitness score. The target personnel who reach the global maximum value are designated as the optimal dispatch node. To address the dispatch deadlock problem caused by multiple personnel having the same highest fitness score (e.g., multiple personnel are completely idle and have the same skill coefficient), the sorting operator incorporates a secondary balance check mechanism. When a secondary balance check mechanism is detected... When there are tied maximum values, the system retrieves the total number of work orders completed by relevant personnel over the past 24 hours and prioritizes assigning the current work order to the personnel with the lowest total completion count. Subsequently, the unit encapsulates the unique global identifier of the abnormal sample, the visualization rendering matrix generated by the previous unit, and the structured diagnostic prompt text to construct a standardized review work order object. This work order object is asynchronously pushed to the target personnel's front-end browser or mobile terminal device via a message middleware, and triggers a hardware-level interrupt event to provide pop-up window or warning sound reminders.

[0151] If the status code object parsed earlier contains an automatic release flag, the system bypasses the aforementioned work order dispatch logic. The unit calls the database's non-blocking write interface to update the sample lifecycle status to "approved" and simultaneously wakes up the system's built-in callback service engine. This service engine is responsible for pushing a message containing complete analysis results and an automatic release timestamp to the data consumption interfaces of the advanced process control system and the enterprise resource planning system.

[0152] Specific application examples:

[0153] In this application embodiment, the influent and effluent water quality of the wastewater treatment device (such as chemical oxygen demand (COD), total nitrogen (TN), and ammonia nitrogen (NH3-N)) are subject to strict logical and physical transformation constraints. At the 15th hour of system operation (corresponding to...) Figure 4 Runtime The data acquisition and timing scheduling module 1 detects the input of influent sample data. At this time, the unit is in a period of large load fluctuation, with the production process load parameter climbing to approximately 8696 (deviating from the 50% stability baseline). The scheduler extracts the set of associated identifiers corresponding to the currently input data. After confirming the existence of upstream and downstream material conversion dependencies, the influent data status is set to suspended and an absolute suspended timestamp is written. When the effluent data is subsequently entered and meets the operating condition alignment tolerance threshold... At that time, the system accurately triggers the joint wake-up, packages the integrated upstream and downstream data and sends it to the basic audit rule calculation module 2. The detection limit cutoff device confirms that all test data are higher than the instrument detection limit. Subsequently, multi-dimensional calculation channels were initiated in parallel. Due to the objective influence of the drastic fluctuations in operating conditions at the time, the influent and effluent concentrations exhibited reasonable and significant fluctuations, resulting in deviation values ​​being output by the non-correlation rules and material conversion rules. The comprehensive anomaly confidence nonlinear multiplication solver was based on the joint probability inference formula. The overall anomaly confidence level of this batch of samples was calculated. The threshold reaches 0.82. In a traditional LIMS system, the hard-coded static threshold is fixed at 0.70. A score of 0.82 would trigger automatic interception, causing false alarms where normal process fluctuations are misjudged as laboratory anomalies. However, this invention uses a dynamic threshold configuration and dynamic comparator, employing a neural network proxy model, to adaptively calculate the current dynamic interception threshold. System comparison and judgment It accurately identifies that the data group belongs to the normal fluctuation under the working condition, and thus generates a status code to allow safe passage.

[0154] When the system has been running smoothly for 35 hours (corresponding to...) Figure 4 Runtime The process load dropped back to the 50% baseline. The laboratory personnel entered data for another batch of water samples, in which the measured total nitrogen in the influent was 45 mg / L, but the measured ammonia nitrogen, a byproduct, was incorrectly reported as 52 mg / L. The correlation rule calculation unit for the same product in the basic audit rule calculation module 2 immediately detected this logical paradox that defied common sense and invoked the physicochemical index ratio rule calculation formula. Accurately calculate the deviation of each item. At this point, the weight matrix and amplification factor configuration unit detected a risk that the data might exceed the mass conservation threshold and immediately invoked the adaptive amplification formula. Its weights are exponentially amplified, and an extreme value bypass is triggered to directly extract the amplified highest confidence scalar. Subsequently, the system uses the boundary fusion formula. Force output of the final overall anomaly confidence level Under the current stable operating conditions, the dynamic threshold of this invention, derived by the dynamic comparator, is... The value was only 0.65, according to system comparison. This immediately triggered a high-level warning.

[0155] Upon receiving the interception status code, the automated interception and review dispatch execution unit immediately changes the lifecycle of the batch of data records to an abnormal lock in the underlying relational database using an exclusive lock, completely preventing the exposure of dirty data to the downstream process control system. Simultaneously, the system generates a visual rendering matrix containing structured diagnostic text containing the logical paradox of influent ammonia nitrogen > total nitrogen, and pushes it to the mobile terminal of the online laboratory supervisor. Just as... Figure 5 The revealed replay verification results, spanning 6 months and covering 15,000 samples, show that traditional auditing mechanisms, due to their rigid 0.70 static boundary and lack of extreme value protection channels, have a false alarm rate as high as 12.4% and a false negative rate of 6.8%. In contrast, the intelligent auditing system of this invention, through the aforementioned dual-track fusion calculation and adaptive boundary inference, significantly reduces the false alarm rate to 2.1% and simultaneously narrows the false negative rate to a maximum of 0.5%, thus completely solving the auditing blind spot of complex heterogeneous data in the laboratory from a mathematical perspective.

Claims

1. An intelligent verification system based on sample analysis data from a LIMS system, characterized in that, include: The underlying rule parameter library (5) stores the parameters required for logical verification, including rule dependencies, conditions and limits, rule baseline weights and amplification factors, and alarm classification thresholds; The data acquisition and timing scheduling module (1) acquires sample analysis data and metadata, determines the verification timing according to the rule dependency relationship and performs suspension or wake-up operations, and outputs integrated data; The data acquisition and timing scheduling module (1) integrates the following: The data analysis and metadata listening and extraction unit acquires the sample analysis data and extracts the metadata containing sampling and recording timestamps; An asynchronous suspension and joint wake-up scheduler allocates initial state, suspension state and verifiable state to the sample analysis data according to the rule dependency relationship, so as to perform suspension operation or wake-up operation; The dynamic timeout degradation decoupling engine updates the sample analysis data that has exceeded the limit and is in the suspended state to the verifiable state; The basic audit rule calculation module (2) calculates the integrated data according to the conditions and limits, and outputs the deviation dataset. The rule conflict resolution and confidence fusion module (3) combines the rule benchmark weight and amplification factor to perform a nonlinear multiplication on the deviation dataset to generate a comprehensive anomaly confidence. The graded alarm and closed-loop intervention module (4) compares the comprehensive anomaly confidence level with the alarm graded threshold and sends a task instruction; The automatic decision interception module (6) compares the comprehensive anomaly confidence level with the dynamic threshold generated based on production stability, generates a status code object and locks the status of the sample analysis data, outputs the audit conclusion or triggers the review workflow; The asynchronous suspension and joint wake-up scheduler extracts the set of associated identifiers corresponding to the sample analysis data. When it determines that the set of associated identifiers is empty, it updates the status of the sample analysis data to the verifiable state. When the set of associated identifiers is determined to be non-empty, the suspension operation is executed, historical data containing sampling timestamps in the cache pool is obtained, the absolute difference between the sampling timestamp of the sample analysis data and the sampling timestamp of the historical data is calculated, and when the set of associated identifiers matches and the absolute difference is less than the working condition alignment tolerance threshold, a joint wake-up instruction is triggered, and the wake-up operation is executed to update the status of the sample analysis data to the verifiable state. The data acquisition and timing scheduling module (1) extracts the sample analysis data and the corresponding metadata in the verifiable state, splices and encapsulates them, and outputs the integrated data.

2. The intelligent review system based on LIMS system sample analysis data according to claim 1, characterized in that, The basic audit rule calculation module (2) integrates the following: The rule parameter configuration and detection limit constant truncation are used to receive the integrated data, parse the component type code and sample unique identifier code in the metadata contained in the integrated data, and extract the corresponding benchmark according to the conditions and limits. Configure the parameter set, call the built-in maximum value operator to perform denominator boundary protection, and generate and distribute safe computing load objects; The irrelevant rule calculation unit is used to receive the security computing load object, perform independent dimension digital comparison calculation, generate and output individual deviation values; The same product correlation rule calculation unit is used to receive the security calculation load object, activate the component summation rule calculation channel and the physicochemical index ratio rule calculation channel to perform calculations, and output the total concentration deviation of components and the deviation of primary and secondary auxiliary indicators in the same batch of samples. The correlation rule calculation unit for different samples is used to extract the sampling timestamps of the upstream samples and the downstream samples in the safe calculation load object, calculate the absolute deviation between the actual residence time and the theoretical residence time, and when the absolute deviation is less than the working condition tolerance window, start the material conversion rule calculation channel to perform the calculation and output the conversion rate deviation. The basic audit rule calculation module (2) merges and encapsulates the individual deviation values, the total concentration deviation, the primary and secondary auxiliary indicator deviations, and the conversion rate deviation, and outputs the deviation dataset.

3. The intelligent review system based on LIMS system sample analysis data according to claim 2, characterized in that, When calculating the total concentration deviation, the same product correlation rule calculation unit extracts multiple component values ​​to be added from the safe calculation load object. When it determines that any component value to be added carries a missing identifier, it calls the default substitution constant in the benchmark configuration parameter set to perform numerical interpolation and then performs a summation operation to generate the total concentration deviation. When extracting the measured values ​​from the secure computing load object, the non-correlation rule calculation unit determines whether the measured values ​​fall within the closed interval formed by the upper and lower control limits defined by the baseline configuration parameter set, and generates the individual deviation value containing the direction feature code based on the interval crossing result and the difference distance.

4. The intelligent review system based on LIMS system sample analysis data according to claim 1, characterized in that, The rule conflict resolution and confidence fusion module (3) integrates the following: The weight matrix and amplification factor configuration unit extracts the rule benchmark weight and the amplification factor from the underlying rule parameter library (5), deduces the initial dynamic weight based on the rule benchmark weight through the built-in multilayer perceptron model, performs Hadamard product operation on the initial dynamic weight and the amplification factor to generate an effective rule benchmark weight value, calls the global normalization operator to perform data compression on the effective rule benchmark weight value and outputs a structured weight matrix object; The comprehensive anomaly confidence nonlinear multiplication solver receives the deviation dataset and the structured weight matrix object, constructs an operational array, and performs fusion calculations through a parallel dual-channel architecture. The first channel extracts the maximum deviation scalar from the deviation dataset and performs extreme value bypass extraction; The second channel calls the logarithmic mapping engine to transform the computation array into the natural logarithm space, and realizes the nonlinear multiplication through addition dimensionality reduction calculation to generate joint anomaly confidence. The maximum deviation scalar is numerically compared with the joint anomaly confidence score, and the maximum value of the two is extracted as the final comprehensive anomaly confidence score.

5. The intelligent review system based on LIMS system sample analysis data according to claim 1, characterized in that, The underlying rule parameter library (5) contains the following: A multidimensional parameter mapping table, with the sample inspection method standard number, product brand number and device tag number as the joint primary key, is used to map and store the parameters required for the logical verification. The parameters required for the logical verification include the corresponding rule dependency relationship, the conditions and limits and the alarm classification threshold. The version control and hot update component is used to listen to the submission payload of the external rule configuration interface. The version control and hot update component parses the submission payload and generates an incremental update script. During the verification time interval of the data acquisition and timing scheduling module (1), the lock-free reload operation of the multidimensional parameter mapping table is executed, and the incremental update script is written into the multidimensional parameter mapping table so as to realize the dynamic effect of the parameters required for the logical verification without interrupting the current review process of the system.

6. The intelligent review system based on LIMS system sample analysis data according to claim 2, characterized in that, The automatic decision-making interception module (6) integrates the following: Threshold configuration and dynamic comparator are used to extract production stability features. A baseline dynamic threshold is generated based on the production stability features through a feedforward neural network. A boundary constraint operator is called to limit and prune the baseline dynamic threshold and output the dynamic threshold. The comprehensive anomaly confidence is compared with the dynamic threshold, and the status code object is generated based on the comparison result. The automated interception and review dispatch execution unit has a built-in database transaction lock mechanism. When it receives the status code object and parses the status code object to find that it contains an automatic interception or manual review identifier, it initiates a status freeze command to lock the status of the sample analysis data. After locking the status, it starts a dynamic dispatch solver to poll the online personnel queue to extract real-time status features, calculates the dynamic load equivalent index and the comprehensive fitness score, and extracts the online personnel node whose comprehensive fitness score reaches the maximum value to dispatch a review work order object, thereby triggering the review workflow. The auxiliary prompt and visual explanation output unit is used to map the cause rule encoding to business description text based on the status code object, concatenate them to generate a structured diagnostic prompt string, and then output the audit conclusion.

7. The intelligent review system based on LIMS system sample analysis data according to claim 6, characterized in that, The auxiliary prompt and visualization explanation output unit extracts the same-site measured dataset from the historical timeline based on the sampling timestamp of the sample. After removing samples with deviations from the operating conditions and outliers, it calculates the moving average and standard deviation of the historical baseline sequence to generate dynamic baseline control upper and lower limits. The dynamic baseline control upper and lower limits are packaged with the same-site measured dataset to generate a comprehensive visualization rendering matrix message. The comprehensive visualization rendering matrix message is fused with the structured diagnostic prompt string and output as the review conclusion containing graphical traceability explanation, providing a decision basis for manual review.

8. The intelligent review system based on LIMS system sample analysis data according to claim 1, characterized in that, The hierarchical alarm and closed-loop intervention module (4) integrates the following: The alarm level mapping unit is used to receive the comprehensive anomaly confidence, extract the alarm classification thresholds in the underlying rule parameter library (5) to construct a step threshold interval matrix, compare the comprehensive anomaly confidence with each alarm classification threshold in the step threshold interval matrix, determine the target interval in which the comprehensive anomaly confidence falls, and output the corresponding discrete alarm level scalar. A closed-loop action assembly engine is used to extract response strategy templates bound to the discrete alarm level scalar. The response strategy templates include message push channel identifiers and approval node flow blocking identifiers. The engine generates the task instructions based on the response strategy templates and pushes the task instructions into the system task scheduling queue for distribution, so as to complete the sending of the task instructions and execute the hierarchical alarm and closed-loop intervention.

Images