A production line state monitoring and mes integration notification method and system

Abstract

This invention relates to the field of production line status monitoring, specifically providing a method and system for integrating production line status monitoring with MES (Manufacturing Execution System) notifications. The method includes: collecting multimodal data and constructing a unified state context object; performing spatiotemporal alignment and dynamic weight fusion based on process stages on multi-source asynchronous data to obtain feature vectors; inferring the semantic current state of the production line through a time-series pattern recognition model; generating an enhanced data set based on the state context and the current production line state, and then outputting structured state events through anomaly detection and root cause analysis; combining MES business information and outputting the optimal response strategy through predefined decision rules; encapsulating the strategy into operation instructions recognizable by MES and issuing them to drive MES to execute work order creation, Kanban update, or message notification. This invention achieves semantic diagnosis, global decision-making, and automated response of production line status, thereby improving production transparency, decision accuracy, and the level of operational automation.

Description

Technical Field

[0001] This invention relates to the field of production line status monitoring, and specifically to a method and system for integrating production line status monitoring with MES (Manufacturing Execution System) notification. Background Technology

[0002] While current automated production lines are equipped with a large number of intelligent devices, the phenomenon of "information silos" is prevalent due to the heterogeneity of these devices and the lack of unified data protocols. The integration of existing monitoring systems with MES (Manufacturing Execution System) is often low, with the following main problems: 1) Device status, alarms, and quality information are usually independent of the MES production management process, failing to automatically drive core business processes in MES such as work order creation, material scheduling, and quality traceability. Status monitoring is disconnected from MES business execution, impacting operational efficiency; 2) MES dashboards typically display a large number of alarms in a simple list format, lacking intelligent aggregation, root cause analysis, and semantic extraction of multi-source alarm information. Key information is easily buried, making it difficult for managers to quickly locate core problems and make decisions; 3) Notifications of production anomalies largely rely on personnel checking on MES terminals, failing to achieve proactive and accurate push notifications based on event severity and business logic; 4) MES systems typically lack the ability to integrate and understand real-time multimodal data from the production line, and also struggle to dynamically correlate real-time status with business information such as production orders and delivery cycles, making it difficult to support forward-looking decision-making. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a production line status monitoring and MES integrated notification method and system, which enables semantic diagnosis, global decision-making, and automated response of production line status, thereby improving production transparency, decision accuracy, and the level of operational automation.

[0004] In a first aspect, the technical solution of the present invention provides a production line status monitoring and MES integration notification method, comprising the following steps: Multimodal data, including equipment parameters, process quality and material information, is collected. After preprocessing, the multimodal data is associated with and encapsulated with the unique identifier and timestamp of the production unit in the production line to construct a unified state context object. This object contains the original state information, data vector and metadata reported by each device. Spatiotemporal alignment is performed on the multi-source asynchronous data contained in the state context object, and dynamic weights are assigned to the data vectors from different devices in the state context object based on the current process stage for data fusion to obtain the fused feature vector; The feature vector is input into a pre-trained temporal pattern recognition model to infer the current state of the production line with semantics; the current state of the production line refers to a semantic description that includes at least one of the following: equipment operation stage, process completion status, or quality defect type. The state context object and the current state of the production line are linked through data association and denoising algorithms to generate an enhanced dataset; and the enhanced dataset is used as input, and an anomaly detection model and a root cause analysis algorithm based on multi-source information association are used to output state events containing anomaly type, severity level and root cause inference results. The state events and business information obtained from the MES system are used as inputs, and the optimal response strategy is output through predefined decision rules. The optimal response strategy is taken as input and encapsulated according to the structured event object format recognizable by the MES system to generate operation instructions that drive the MES workflow. The operation instructions are sent to the MES system through the API interface to drive the MES system to execute at least one workflow, such as creating a work order, updating the Kanban board, or sending a notification.

[0005] Secondly, the technical solution of the present invention provides a production line status monitoring and MES integrated notification system, including: The State Context Object Construction Module is used to collect multimodal data, including equipment parameters, process quality and material information. After preprocessing the multimodal data, it associates and encapsulates it with the unique identifier and timestamp of the production unit in the production line to construct a unified State Context Object. This object contains the original state information, data vector and metadata reported by each device. The data fusion module is used to perform spatiotemporal alignment on the multi-source asynchronous data contained in the state context object, and to assign dynamic weights to the data vectors from different devices in the state context object based on the current process stage to perform data fusion and obtain the fused feature vector. The production line current state reasoning module is used to input the feature vector into a pre-trained temporal pattern recognition model to reason out the semantic current state of the production line; the current state of the production line refers to a semantic description that includes at least one of the following: equipment operation stage, process completion status, or quality defect type. The state event inference module is used to generate an enhanced dataset by associating the state context object with the current state of the production line through data association and denoising algorithms; and using the enhanced dataset as input, it outputs state events containing anomaly type, severity level and root cause inference results through an anomaly detection model and a root cause analysis algorithm based on multi-source information association. The optimal response strategy output module is used to take the status event and the business information obtained from the MES system as input, and output the optimal response strategy through predefined decision rules. The MES operation instruction generation and distribution module is used to take the optimal response strategy as input, encapsulate it according to the structured event object format that the MES system can recognize, and generate operation instructions to drive the MES workflow; and send the operation instructions to the MES system through the API interface to drive the MES system to execute at least one of the workflows of creating work orders, updating Kanban, or sending notifications.

[0006] As can be seen from the above technical solutions, this application has the following advantages: 1) By unifying the state perception to generate and issue MES operation instructions, the monitoring layer and the execution layer are connected, so that changes in equipment state can automatically and accurately trigger business workflows such as maintenance, quality, and materials in MES, thereby improving production response speed and the level of operation and maintenance automation. 2) By integrating multi-source data, identifying time-series patterns, and analyzing root causes, the massive and chaotic equipment data at the bottom layer is refined into high-level "current production line status" and "status events" with clear semantics and root causes, solving the problem of MES alarm information overload and providing managers with clear and structured information that can be directly used for decision-making. 3) It can automatically generate and issue instructions to drive the MES operation based on the diagnosis and decision results. At the same time, the systematic driving of the entire process provides a foundation for the full tracking and result feedback of anomaly handling, forming a management closed loop. 4) By constructing a unified state context and performing data fusion, real-time monitoring of the overall production line status is achieved. During the decision-making process, physical state events are associated with business information in the MES, and optimal response strategies that comprehensively consider efficiency, cost, risk, and other dimensions are output through predefined decision rules, thereby improving the accuracy and overall effectiveness of production decisions. Attached Figure Description

[0007] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of a production line status monitoring and MES integration notification method provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the production line status monitoring and MES integration notification architecture.

[0010] Figure 3 This is a schematic block diagram of a production line status monitoring and MES integrated notification system provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0013] Figure 1 This is a schematic diagram of a production line status monitoring and MES integration notification method provided in an embodiment of the present invention. Figure 1 The executing entity can be a production line status monitoring and MES integrated notification system. The production line status monitoring and MES integrated notification method provided in this embodiment of the invention is executed by a computer device; correspondingly, the production line status monitoring and MES integrated notification system runs on the computer device. Depending on different requirements, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0014] like Figure 1 As shown, the method includes the following steps.

[0015] S1. Collect multimodal data, including equipment parameters, process quality and material information, and after preprocessing the multimodal data, associate and encapsulate it with the unique identifier and timestamp of the production unit of the production line to construct a unified state context object; this object contains the original state information, data vector and metadata reported by each device.

[0016] S2, perform spatiotemporal alignment on the multi-source asynchronous data contained in the state context object, and assign dynamic weights to the data vectors from different devices in the state context object based on the current process stage to perform data fusion and obtain the fused feature vector.

[0017] S3, input the feature vector into a pre-trained temporal pattern recognition model to infer the current state of the production line with semantics; the current state of the production line refers to a semantic description that includes at least one of the following: equipment operation stage, process completion status, or quality defect type.

[0018] S4 generates an enhanced dataset by associating the state context object with the current state of the production line through data association and denoising algorithms; and takes the enhanced dataset as input, outputs state events containing anomaly type, severity level and root cause inference results through an anomaly detection model and a root cause analysis algorithm based on multi-source information association.

[0019] S5, taking the status event and the business information obtained from the MES system as input, and outputting the optimal response strategy through predefined decision rules.

[0020] S6. The optimal response strategy is taken as input and encapsulated according to the structured event object format that the MES system can recognize to generate operation instructions that drive the MES workflow. The operation instructions are sent to the MES system through the API interface to drive the MES system to execute at least one of the workflows of creating work orders, updating Kanban, or sending notifications.

[0021] Furthermore, as a refinement and extension of the specific implementation methods described above, and to fully illustrate the specific implementation process in this embodiment, another method for production line status monitoring and MES integration notification is provided. This method is based on... Figure 2 The diagram illustrating the production line status monitoring and MES integration notification architecture includes a multi-source status perception layer, an intelligent collaborative decision-making layer, and an MES integration and execution layer. The multi-source status perception layer collects multimodal data and preprocesses it. The intelligent collaborative decision-making layer comprises an adaptive process flow engine and a central collaborative decision-making kernel. The adaptive process flow engine, acting as the system's "sensory center," is responsible for spatiotemporal alignment and dynamic weight fusion of the multi-source asynchronous data streams from the perception layer, generating a unified, structured production line status context. It also understands this data and outputs a semantically meaningful production line status context through dynamic weight fusion and status inference. The central collaborative decision-making kernel embeds multiple role-based intelligent agents that collaborate based on a shared status context. The core intelligent agents are shown in Table 1 below. Table 1: Core Intelligent Agents

[0022] To achieve efficient collaboration among agents in a multi-agent system, the system adopts a unified structured data exchange protocol, namely JSON format based on predefined fields, as shown in Table 2.

[0023] Table 2: Structured Data Exchange Protocols

[0024] MES Integration and Execution Layer: Deeply integrated with the MES system through standard interfaces such as RESTful APIs. It receives structured events from the decision-making layer and drives the MES to execute predefined workflows, such as creating maintenance / quality work orders, updating production dashboards, and sending internal messages.

[0025] Based on the above production line status monitoring and MES integrated notification architecture, this specific embodiment includes the following steps.

[0026] S101, collects multimodal data.

[0027] The collected multimodal data includes equipment parameters, process quality, and material information. After preprocessing, the multimodal data is associated with and encapsulated with the unique identifier and timestamp of the production unit on the production line to construct a unified state context object. This object contains the original state information, data vectors, and metadata reported by each device.

[0028] Specifically, multimodal data is collected in real time via data acquisition adapters deployed on key process equipment throughout the production line. Taking a PCBA board production line as an example, key process equipment includes, but is not limited to: laser engraving machines for circuit marking, SPI (solder paste inspection) machines for solder paste printing quality inspection, pick-and-place machines for electronic component mounting, AOI (automatic optical inspection) machines for post-soldering quality inspection, and reflow ovens and wave soldering ovens. The collected multimodal data specifically includes: Equipment parameters: such as the vacuum value and placement pressure of the placement head of the pick-and-place machine, the set temperature and actual temperature of each temperature zone of the reflow oven, and the speed of the conveyor chain, etc. Process quality data: such as solder paste thickness, area, and volume detected by SPI, and component defects such as missing components, misaligned components, offsets, and bridging detected by AOI, as well as their coordinate locations; Material information: such as the remaining quantity of components in the pick-and-place machine feeder, the material batch number, and the unique serial number of the PCB board entering the production line obtained by the scanner.

[0029] The collected raw multimodal data underwent preliminary cleaning and standardization to eliminate the influence of dimensions, laying the foundation for subsequent data fusion and comparison. Specifically, at least one method from Min-Max normalization and Z-Score standardization was used.

[0030] The preprocessed, standardized data is associated with and encapsulated with the core business logic of the production line to construct a unified state context object C(t). This object is a complete snapshot of the production line state at timestamp t; it is a structured data volume that can be represented as follows:

[0031] Among them, PCBA is the unique identifier for the manufactured PCB board, Timestamp is the timestamp, and State is the production line status. Given a set of data vectors from n devices, A collection of m metadata or event tags.

[0032] It should be noted that the production line state here refers to the raw, local, single-event-based state code or flag reported by the equipment. It is the output of the perception layer and a part of the input data. For example, the pick-and-place machine may report the state as Running, the board mounter as Waiting, and the AOI as Testing, which are state facts. The current production line state inferred in the subsequent step S103 is a conclusion reached by the adaptive process flow engine after comprehensive reasoning about the entire C(t) object. It is the output of the intelligent collaborative decision-making layer and is a semantic result that has been understood and refined.

[0033] S102, Multi-source data fusion generates fused feature vectors.

[0034] This step performs spatiotemporal alignment on the multi-source asynchronous data contained in the state context object, and assigns dynamic weights to the data vectors from different devices in the state context object based on the current process stage to perform data fusion, resulting in a fused feature vector. Specifically, it includes the following steps.

[0035] S1021 uses the production line's global clock or a unified timing reference to interpolate or resample all data points, aligning the data points to a unified, equally spaced time series to achieve time alignment.

[0036] Receives the original data vector set from the state context object C(t). Because these data originate from heterogeneous devices with different sampling frequencies and communication delays, they are not synchronized in time and may not be precisely matched in space (referring to their association with production units). Therefore, alignment processing is performed first, including time alignment and spatial alignment.

[0037] Time alignment references the production line's global clock or a unified timing baseline, interpolating or resampling all data points to align them to a uniform, equally spaced time series. The unified timing baseline can be the timeline of the PCB board entering the production line. Specifically, for a target time point... The first is calculated using linear interpolation. The values ​​of each data source at that moment:

[0038] In the formula, and It is a data source exist The nearest original data timestamp before and after.

[0039] S1022 uses the production unit currently being processed on the production line as the spatial reference, and through the unique identifier of the production unit, filters and associates all equipment data related to the production unit within the same time window to achieve spatial alignment.

[0040] This step is for spatial alignment, ensuring that all data in subsequent fusion processing points to the same object. A production unit can be a specific PCB board.

[0041] S1023, Perform dynamic weighted fusion based on process stage on the spatiotemporally aligned data, including: determining the current process stage, assigning dynamic weight values ​​to each data source based on the current process stage, and fusing the target state features or indicators to be fused using the following formula to obtain the fused feature vector. ,

[0042] In the formula, This indicates the total number of data sources participating in the integration. For the first Each data source is at the current process stage. Dynamic weights under, For the first Data quality scores for each data source For the first The data values ​​from the data source are preprocessed (referring to the preprocessing in step S101) and spatiotemporally aligned.

[0043] First, based on production line scheduling information or identification of equipment status sequences, determine the specific process stage of the current production line, such as "solder paste printing," "chip mounting," or "reflow soldering." Second, assign a dynamic weight value to each data source, i.e., the data vector from different devices. This weight This is the current stage of the process. The function assigns values ​​based on a predefined weight mapping table, which defines the importance of each data source to the overall state reasoning at different process stages. The basis for this is that data from equipment that plays a key role in the current process stage should be assigned higher weights. Table 3 below provides an example of the weighting strategy.

[0044] Table 3: Examples of Weighting Strategies

[0045] By performing the above calculations on all relevant feature dimensions, a fused feature vector that comprehensively reflects the global state of the production line at the current process stage is finally formed. This vector will be used as the input of a lightweight temporal pattern recognition model.

[0046] In one optional implementation, the first Data quality score of each data source The data quality score is calculated through the following steps: the completeness score is obtained by comparing the actual number of data points received with the expected number of data points received; the freshness score is obtained by using an indicator function based on the difference between the current time and the data timestamp; the reasonableness score is calculated based on the predefined reasonable range and tolerance range of the data values; and the data quality score is obtained by weighted summation of the completeness score, freshness score, and reasonableness score.

[0047] No. Data quality score of each data source The expression is,

[0048] In the formula, The completeness score is calculated as follows: = Actual number of data points received / Expected number of data points received. This value will decrease if data is lost during transmission.

[0049] The freshness score is calculated as follows: ,in, It is the difference between the current time and the data timestamp. This is the attenuation coefficient. The older the data, the lower the score for this item.

[0050] The reasonableness score is calculated based on a predefined reasonableness range [Min, Max]. If... Within the scope, =1; if it is outside the range but within the fault tolerance range. It decreases linearly from 1 to 0; otherwise it is 0.

[0051] For example, , where Mid is the midpoint of the range.

[0052] , , These are the weighting coefficients for completeness, freshness, and reasonableness scores, with the sum of the three being 1. These weights can be adjusted according to the characteristics of different data sources and the importance of the business.

[0053] S103, infer the current state of the production line based on the time sequence pattern recognition model.

[0054] The feature vector obtained in the previous step is input into a pre-trained temporal pattern recognition model to infer the semantic current state of the production line. The current state of the production line refers to a semantic description that includes at least one of the following: equipment operation stage, process completion status, or quality defect type.

[0055] In this embodiment, the temporal pattern recognition model is an encoder-decoder architecture combined with a state semantic mapping layer, including an input layer, a feature encoder, a state decoder, and a state semantic mapping layer.

[0056] Input layer: Receives a fixed-dimensional sequence of fused feature vectors generated by the adaptive process flow engine, denoted as... ,in For the current moment, This represents the size of the historical time window.

[0057] Feature encoder: Consists of a multi-layer long short-term memory network. Its function is to encode the temporal dependencies of the input sequence, extract high-level temporal features, and ultimately condense the information of the entire sequence into a fixed-dimensional context vector.

[0058] State decoder: Consists of a fully connected neural network. It takes the context vector output by the encoder as input and its task is to learn the mapping from temporal features to an abstract state space. The decoder's output is a state probability distribution vector. , where m is the number of all possible states. The model indicates that the current production line is in the first stage. The confidence level of a state.

[0059] State semantic mapping layer: A non-parametric, predefined mapping table. It receives the state probability distribution vector output by the state decoder. The state with the highest probability is mapped to the corresponding semantic state description.

[0060] The training process of the temporal pattern recognition model includes the following steps: collecting historical multimodal data of the production line under normal and various typical abnormal states to form a training dataset; performing the same preprocessing, spatiotemporal alignment, and dynamic weight fusion operations as online processing on the training dataset to generate corresponding fused feature vector sequences and labeled real state labels; using the fused feature vector sequences as input and the corresponding real state labels as supervision signals to perform supervised training of the temporal pattern recognition model; during training, using the cross-entropy loss function to calculate the difference between the state probability distribution output by the model and the real labels, and optimizing the model parameters through the backpropagation algorithm until the model converges; and solidifying the trained model parameters into a pre-trained temporal pattern recognition model for online inference.

[0061] Based on a pre-trained temporal pattern recognition model, the current state of the production line with semantics is inferred, specifically including the following steps.

[0062] S1031, construct a fused feature vector sequence based on the fused feature vector, and input the fused feature vector sequence into a pre-trained temporal pattern recognition model, which is then received by the input layer of the model.

[0063] S1032, the input layer outputs the fused feature vector sequence to the feature encoder, which learns the long-term and short-term dependencies in the sequence through a gating mechanism and outputs a context vector representing the production line operation mode within the entire time window; the feature encoder is a multi-layer LSTM network.

[0064] S1033, The context vector is input to the state decoder, which calculates and outputs the state probability distribution vector through forward propagation. , This indicates that the production line is in the first stage. The confidence probability of a predefined state; where the state decoder is a fully connected neural network.

[0065] S1034, the state probability distribution vector is input to the state semantic mapping layer. This layer determines whether the maximum probability in the state probability distribution vector is not less than the confidence threshold. If so, the state ID J = argmax(P) with the highest confidence is selected as the output. Then, by querying the predefined state-semantic mapping table, the state ID J is mapped to the corresponding semantic current state of the production line. Otherwise, "state unknown" is output or a manual review process is triggered. The state-semantic mapping table defines the correspondence between different states and natural language descriptions.

[0066] For example, status ID 5 can be mapped to the semantics "Reflow oven temperature fluctuations caused bridging in the QFN chip area". The current status of the production line can also be "Printing solder paste on side B", "U5 component mounting completed", "Bridging detected in QFN chip area", "23 stations have less than 30 minutes remaining".

[0067] S104, Generate an enhanced data set based on the state context object and the current state of the production line, and output state events based on the enhanced data set.

[0068] First, the state context object and the current state of the production line are correlated using data association and denoising algorithms to generate an enhanced dataset. Then, using the enhanced dataset as input, an anomaly detection model and a root cause analysis algorithm based on multi-source information association are used to output state events containing anomaly type, severity level, and root cause inference results.

[0069] The enhanced dataset is generated using the following steps.

[0070] S1041, based on the semantics of the current state of the production line, extract the relevant subset of equipment parameter data from the state context object from the predefined state-data association mapping library.

[0071] Receive the current status of the production line, for example: "QFN chip area bridging". Based on this semantic state, query a predefined state-data association mapping library. This mapping library defines which devices (data sources) and which parameters are most relevant to diagnosing the system in a specific semantic state.

[0072] Based on the query results, from the data vector set of the state context object C(t) In this process, specific devices and their parameter data marked as "relevant" by the mapping library are extracted, while unmarked irrelevant data is ignored, thus achieving data dimensionality reduction and focus.

[0073] S1042, using the unique identifier of the production unit in the state context object as the key, the data subset is associated to realize entity association.

[0074] Using the unique identifier of the production unit carried in the state context object C(t) as the primary key, entity-level associations are performed on the subset of data indexed by S1041. This step ensures that all processed data belongs to the same monitored physical object, such as a specific PCB board, thus resolving the issue of data consistency across spatial dimensions.

[0075] S1043, based on the process topology knowledge of the production line, logically links data from different devices in the associated data subset that correspond to the same physical location or process step to achieve logical association.

[0076] Based on the completed entity association, further logical links are made between data from different equipment that describe the same physical location or the same process step, according to the process topology knowledge graph of the production line.

[0077] For example, physical location association includes associating a defect event detected by the AOI equipment as "bridging exists at coordinates (X=15.2, Y=20.5)" with component information recorded by the pick-and-place machine as "a U5 chip was mounted at coordinates (X=15.2, Y=20.5)". Process stage association includes associating the temperature profile data of the reflow oven's "zone 5" with the monitoring readings of the thermocouple "TC_05".

[0078] S1044 uses the physical knowledge model of equipment or process to verify the physical rationality of the data, identifies outliers through dynamic thresholding, and performs noise reduction and correction on the logically associated data based on the verification and identification results.

[0079] Physical model-based rationality verification includes: using a physical knowledge model of the equipment or process to verify the physical rationality of the data. The physical knowledge model can be a thermodynamic model. For example, for a reflow oven data point displayed as room temperature, if the time points before and after it are both at high process temperatures and the equipment status is "running", then this point is determined to be transient noise and replaced with a model-based predicted value or time series interpolation.

[0080] Outlier suppression based on statistical rules includes applying statistical process control methods; in this embodiment, a dynamic thresholding method is used to identify and process outliers. For a key parameter sequence, the dynamic threshold is calculated based on the mean μ and standard deviation σ of the parameter within a recent sliding time window, for example, set to [μ-3σ, μ+3σ]. Data points outside this range are considered statistical outliers and can be smoothed or removed.

[0081] S1045 encapsulates the processed data and its associated metadata, and outputs an enhanced dataset.

[0082] The data, after being indexed, associated, and cleaned as described above, is encapsulated into a new structured object, namely, an enhanced data set. This set not only contains the cleaned numerical data but also includes the entity and logical relationships established in steps S1042 and S1043 as metadata.

[0083] Then, taking enhanced dataset collaboration as input, the anomaly detection model and the root cause analysis algorithm based on multi-source information association are used to output state events containing anomaly type, severity level and root cause inference results, specifically including the following steps.

[0084] S1046, the target parameter vector in the augmented dataset is input into a pre-trained unsupervised anomaly detection model. The model calculates the reconstruction error of the input vector and compares the reconstruction error with a preset threshold. If the reconstruction error exceeds the preset threshold, the production line status is determined to be abnormal, and an anomaly score and an anomaly dimension identifier are output. The anomaly dimension identifier indicates which parameter or group of parameters has become abnormal. The unsupervised anomaly detection model is an unsupervised learning model based on an autoencoder.

[0085] The anomaly detection model is a hybrid detection system that combines unsupervised and supervised learning. It does not rely on a large number of labeled anomaly samples for training, but instead learns and memorizes normal production conditions to identify deviations.

[0086] The anomaly detection model employs an unsupervised learning model based on an autoencoder or a type of support vector machine. During the training phase, the model is trained only on a large amount of historical data where the production line is in a known normal state. The autoencoder learns how to compress and reconstruct normal data with minimal reconstruction error, while the type of support vector machine learns a compact boundary in the feature space that can encompass all normal data points. During the inference phase, the key parameter vectors from the real-time augmented data set are input into the trained model. For the autoencoder, its reconstruction error, such as mean squared error, is calculated. If the error exceeds a preset threshold, it is judged as an anomaly. For the type of support vector machine, it is determined whether the data point falls outside the training-derived decision boundary; if it falls outside, it is judged as an anomaly. The direct output of this model is an anomaly score and a preliminary anomaly label. The anomaly score can be the reconstruction error value or the distance to the decision boundary, and the anomaly label can be "abnormal temperature pattern" or "abnormal vibration signal."

[0087] S1047, The abnormal dimension identifier and the logically related metadata contained in the enhanced data set are used as evidence input into the production line knowledge graph. Starting from the node corresponding to the evidence, the upstream traversal is performed along the causal edge to find all possible source nodes and form a set of root cause hypotheses.

[0088] S1048, For each hypothesis in the root cause hypothesis set, perform a confidence assessment, sort all root cause hypotheses by confidence based on the assessment results, and select one or more hypotheses with the highest confidence as the root cause inference results; wherein, the confidence assessment is based on at least one of the following factors: the path distance and relationship weight between the hypothesis node and the evidence node in the knowledge graph, the co-occurrence probability of the hypothesis and the evidence in historical data, and the magnitude of the current anomaly score.

[0089] S1049: Determine the severity level of the event based on the anomaly score, determine the anomaly type based on the anomaly dimension identifier, and encapsulate the event severity level, anomaly type, and root cause inference results into a structured state event object.

[0090] For example, severity levels include warning, critical, and fatal; anomaly types can be abnormal process parameters or abnormal equipment mechanics.

[0091] The root cause analysis algorithm based on multi-source information association is a knowledge-driven system based on graph reasoning and causal discovery. It uses the data associations established in the previous steps to connect isolated outliers into a path leading to the root cause.

[0092] The root cause analysis algorithm comprises a knowledge base and an analysis engine. The knowledge base maintains a production line knowledge graph constructed using an entity-relationship model. Nodes include equipment, parameters, process steps, components, etc., and edges represent causal, topological, or logical relationships between them, such as "reflow oven temperature affects solder joint quality" or "the feeder of the placement machine supplies materials to the placement head." The analysis engine's execution process includes: injecting anomaly labels and scores output by the anomaly detection model, along with logically related anomaly data points from the augmented dataset, as "evidence nodes" into the knowledge graph; starting from these evidence nodes, traversing backward (upstream) along the causal edges in the knowledge graph to identify all possible root cause nodes that could lead to these anomalies, forming a series of root cause hypotheses; for each root cause hypothesis, based on its connection strength with evidence nodes in the graph, historical co-occurrence frequency, and the anomaly score of the current data, using mechanisms such as Bayesian networks or weighted scoring, calculating its probability or confidence level as a true root cause. Finally, all hypotheses are sorted by confidence level.

[0093] S105: Select the optimal response strategy based on status events and business information.

[0094] The status events obtained in step S104 and the business information obtained from the MES system are used as inputs. The optimal response strategy is output through predefined decision rules, specifically including the following steps.

[0095] S1051 encapsulates the state time and business information obtained from the MES system together to construct a complete decision context object.

[0096] The system receives status events from upstream sources, each event containing at least the exception type, severity level, and root cause inference results. Simultaneously, it acquires business information relevant to the current production context in real time through the MES system interface. This business information includes, but is not limited to: the priority of current production orders, order delivery dates, batch information of work-in-process on the production line, inventory status of alternative materials, and the workload status of relevant maintenance personnel. The status events and business information are then encapsulated to construct a complete decision context object.

[0097] S1052, Input the decision context object into the strategy decision engine. The engine predefines several decision dimensions and stores a strategy rule base. Each rule is in the form of "IF<condition>THEN<candidate strategy>".

[0098] The decision context object is input into the strategy decision engine. The engine predefines multiple decision dimensions, including production efficiency, quality cost, equipment risk, and order delivery. The engine stores a strategy rule base, with each rule in the form of "IF <condition> THEN <candidate strategy>". The engine matches the decision context against the conditions in the rule base, activating all matching candidate strategies to form an initial strategy candidate set.

[0099] S1053, match the decision context with the conditions in the policy rule base, activate all candidate policies that meet the conditions, and form an initial policy candidate set.

[0100] S1054, For each candidate strategy in the initial strategy candidate set, calculate the comprehensive utility value based on the utility impact value of each decision dimension, and select the candidate strategy with the highest comprehensive utility value as the optimal response strategy; the optimal response strategy is a structured instruction object, which includes the action goal, the execution object, the required resources, and the constraints.

[0101] Specifically, the decision dimensions include order delivery, quality cost, equipment risk, and production efficiency. A comprehensive utility value is calculated based on the utility impact of each decision dimension. Specifically, this includes: calculating the utility impact of the order delivery dimension based on the expected production delay time caused by the strategy; calculating the utility impact of the quality cost dimension based on the strategy's effect on suppressing defect escape rate; calculating the utility impact of the equipment risk dimension based on the strategy's degree of mitigation of the equipment risk index; calculating the utility impact of the production efficiency dimension based on the changes in capacity utilization caused by the strategy; and finally, a weighted sum of the utility impact values ​​of each dimension is taken to obtain the comprehensive utility value.

[0102] The formula for calculating the overall utility value is expressed as follows:

[0103] In the formula, The weights for each dimension are as follows: order delivery weight is 0.4, quality cost weight is 0.3, equipment risk weight is 0.2, and production efficiency weight is 0.1. This is the impact quantification value of the strategy in this dimension. The impact quantification value is the utility impact value, which means that for a specific candidate strategy, it is a numerical measure of the positive or negative effect it can produce in a specific decision dimension. This is a predicted value of the result after the strategy is executed, calculated based on the current decision context.

[0104] The impact of quantification strategies on on-time order delivery is calculated based on the expected production delay time caused by the strategy, and is expressed as follows:

[0105] In the formula, The anticipated production delays resulting from implementing this strategy, This is the maximum tolerable delay threshold for the order, which can be dynamically set based on the order's delivery margin.

[0106] Expected production delays The calculation is based on the estimated time required for strategy execution and the degree of disruption it will cause to the production line process. The specific calculation method is as follows:

[0107] In the formula, The standard operating time required to execute this strategy, such as the estimated time for equipment adjustment, material change or quality re-inspection, is derived from the standard time library or historical operation data. This refers to the waiting time caused by strategy execution, such as waiting for maintenance personnel to arrive or waiting for specific materials to be delivered. This value can be estimated based on the resource scheduling status. It refers to the use of a waiting time estimation model based on rules and status queries. The core of this model is to use the real-time updated resource status data in the MES system for logical reasoning. This refers to the operation time that can be carried out in parallel with normal production, i.e., the operation portion that does not directly occupy core production line resources. If strategy execution and production do not conflict at all, then this term equals... .

[0108] When the delay is 0, =1, no negative impact; when delays reach or exceed the threshold. =0, resulting in the most severe delivery risk.

[0109] The impact of quantification strategies on quality risk costs is calculated based on the strategy's ability to suppress defect escape rates, and is expressed as follows:

[0110] In the formula, The expected cost of defect escape when no strategy is adopted or when the baseline strategy is adopted. The expected defect escape cost when implementing this candidate strategy.

[0111] Escape costs refer to all economic losses incurred when defective products, which fail to be detected and intercepted in a timely manner under a specific strategy, flow into subsequent processes or are delivered to customers.

[0112] Its calculation comprehensively considers factors such as rework, scrap, and reputational damage. The specific formula is as follows:

[0113] In the formula, The estimated number of defect escapes is based on the current abnormal quality defect rate, the detection rate of the testing equipment, and the detection or control intensity corresponding to the strategy. The average rework cost for a single defective product, including labor, materials, and equipment depreciation; The average cost of scrapping a single defective product because it cannot be reworked; The cost of customer claims or contract penalties resulting from the delivery of defective products is averaged down to the cost of a single product. This is an estimate of the brand reputation loss caused by a quality incident, allocated to the cost of a single product.

[0114] The impact of quantitative strategies on the long-term health and safety risks of equipment is calculated based on the degree to which the strategy mitigates the equipment risk index, and is expressed as follows:

[0115] In the formula, This is the current equipment risk index inferred from the root causes of state events. The predicted equipment risk index after implementing this strategy.

[0116] The severity coefficient is the product of the severity coefficient and the failure probability. The severity coefficient is a predefined constant based on historical maintenance data, representing the severity of the consequences should the failure mode inferred from the current root cause occur. The failure probability represents the probability that the equipment will fail within a future planned cycle under the current root cause conditions. This value is obtained by multiplying the root cause confidence output by the root cause analysis algorithm by the historical failure conversion rate of that root cause.

[0117] The effect of intervention on device status based on candidate strategies is predicted, denoted as,

[0118] In the formula, The strategy effectiveness coefficient is a predefined coefficient in the strategy rule base for each (strategy, root cause) pair, representing the degree to which the strategy is effective in eliminating or mitigating this specific root cause. For example, the strategy "adjust furnace temperature" is effective for the root cause "furnace temperature fluctuation". It can be set to 0.9; while for the root cause "mechanical wear", its The value is 0.1.

[0119] This is the basic risk value of the equipment, a constant greater than 0, representing the inherent risk that the equipment carries even in a healthy state, and is used to prevent the risk from being overly optimistically assessed as zero.

[0120] The impact of quantification on overall production efficiency is calculated based on changes in capacity utilization caused by the strategy, and is expressed as follows:

[0121] In the formula, To determine the projected output for the next planning cycle when implementing this strategy, The planned output under ideal, undisturbed conditions.

[0122] Based on the design cycle time and planned running time of the production line, it is calculated as the product of the ideal cycle time of the production line and the planned running time of the next planning cycle. The ideal cycle time of the production line is the maximum number of products that can be theoretically produced per unit time.

[0123] This is based on the ideal output, minus the capacity loss caused by the implementation of the strategy. The capacity loss is the product of the ideal production line cycle time and the expected effective downtime caused by the strategy, plus the sum of the additional quality loss caused by the strategy. The expected effective downtime caused by the strategy is the one calculated in step S1054. This represents the core production time consumed by the strategy execution itself. The additional quality loss caused by the strategy is an optional correction. Some strategies may improve the pass rate, thereby reducing quality loss; in this case, the additional quality loss caused by the strategy is negative, indicating an increase in output. Its calculation can be based on the estimated impact of the strategy on the defect rate and the value per unit.

[0124] S106, Generate MES operation instructions and issue them.

[0125] The optimal response strategy is taken as input and encapsulated according to the structured event object format that the MES system can recognize to generate operation instructions that drive the MES workflow. The operation instructions are sent to the MES system through the API interface to drive the MES system to execute at least one workflow, such as creating a work order, updating the Kanban board, or sending a notification. The specific steps include the following steps.

[0126] S1061, Based on the action objective in the optimal response strategy, query the strategy-MES operation mapping table to determine at least one MES atomic operation.

[0127] The MES atomic operations include at least one of creating work orders, updating production dashboards, reserving materials, and sending notifications.

[0128] Specifically, upon receiving the optimal response strategy, and based on the action objectives specified in the strategy, such as "creating a quality isolation work order" or "assigning maintenance tasks," the system queries a predefined strategy-MES operation mapping table to determine one or more atomic operations that need to be initiated in the MES system. Atomic operations include, but are not limited to: create_work_order (creating a work order), update_kanban (updating the Kanban board), reserve_material (reserving materials), and send_notification (sending a notification).

[0129] S1062 loads parameter templates for MES atomic operations and extracts information from the optimal response strategy, state events, and state context objects to populate the templates.

[0130] For each atomic operation, the required data structure is loaded from its corresponding parameter template library. Subsequently, specific values ​​are extracted from the execution object, required resources, and constraints of the optimal response strategy, as well as from the upstream state events and state context object C(t), to populate the template.

[0131] For example, if the atomic operation is create_work_order, the fields work_order_type (work order type), title (title), description (description), priority (priority), and assigned_role (assigned role) are parsed from its parameter template and populated with specific information such as "quality anomaly handling" in the strategy, "QFN bridge defect handling" in the status event, description in the root cause analysis, "highest priority" in the decision, and "quality engineer".

[0132] S1063 encapsulates the filled parameters with the corresponding atomic operation identifiers to generate a structured event object that conforms to the MES interface specification.

[0133] The atomic operations with pre-configured parameters are encapsulated into one or more structured event objects, following the integration interface specifications provided by the MES system (which can be in the request body format of a RESTful API). Each event object is a standard, machine-readable data structure, preferably in JSON or XML format. Each event object explicitly contains the target operation (mes_operation) and operation parameters (operation_parameters).

[0134] S1064, based on the logical dependencies between MES atomic operations, the structured event objects are sorted or grouped to generate the operation instructions; when there are multiple structured event objects, their execution order is determined according to predefined transaction logic to generate an ordered sequence of operation instructions.

[0135] Determine the logical dependencies between multiple atomic operations. If dependencies exist between operations, for example, "reserving materials" must precede "creating a material requisition order," then according to predefined transaction logic, these structured event objects are sorted to generate an ordered sequence of operation instructions; if there are no dependencies, a concurrent set of operation instructions is generated. This sequence or set is the final generated set of operation instructions that can drive the MES workflow.

[0136] S1065, the operation instruction is sent to the MES system through the API interface to drive the MES system to execute at least one of the following workflows: creating a work order, updating the Kanban board, or sending a notification.

[0137] Specifically, the operation instructions (i.e., structured, MES-recognizable event objects) generated in step S106 are sent to the MES system via standard interfaces such as RESTful APIs or Web Services. Upon receiving the operation instructions, the MES system parses their content and drives the automatic execution of its internally predefined workflow. The workflow includes, but is not limited to: Create a work order: For example, create a quality exception handling work order to handle "QFN bridging defect" and automatically assign it to the "Quality Engineer" role; Update the production dashboard: Update the production line status in real time on the MES electronic dashboard, for example, mark the current status as "quality anomaly, in progress" and provide visual alerts to relevant positions; Send notifications: Based on the notification rules embedded in the instructions, send precise alarm and processing notifications to designated personnel through channels such as in-site messages, emails, or mobile applications.

[0138] The foregoing has described in detail an embodiment of a production line status monitoring and MES integration notification method. Based on the production line status monitoring and MES integration notification method described in the above embodiment, this invention also provides a production line status monitoring and MES integration notification system corresponding to the method.

[0139] Figure 3 This is a schematic block diagram of a production line status monitoring and MES integrated notification system provided in an embodiment of the present invention. In this embodiment, the production line status monitoring and MES integrated notification system 300 can be divided into multiple functional modules according to the functions it performs, such as... Figure 3 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and which are stored in memory.

[0140] The state context object construction module 310 is used to collect multimodal data, including equipment parameters, process quality and material information, and after preprocessing the multimodal data, associate and encapsulate it with the unique identifier and timestamp of the production unit of the production line to construct a unified state context object; the object contains the original state information, data vector and metadata reported by each device.

[0141] The data fusion module 320 is used to perform spatiotemporal alignment on the multi-source asynchronous data contained in the state context object, and to assign dynamic weights to the data vectors from different devices in the state context object based on the current process stage to perform data fusion and obtain the fused feature vector.

[0142] The production line current state reasoning module 330 is used to input the feature vector into a pre-trained temporal pattern recognition model to reason out the semantic current state of the production line; the current state of the production line refers to a semantic description that includes at least one of the following: equipment operation stage, process completion status, or quality defect type.

[0143] The state event inference module 340 is used to generate an enhanced dataset by associating the state context object with the current state of the production line through data association and denoising algorithms; and using the enhanced dataset as input, it outputs state events containing anomaly type, severity level and root cause inference results through an anomaly detection model and a root cause analysis algorithm based on multi-source information association.

[0144] The optimal response strategy output module 350 is used to take the status event and the business information obtained from the MES system as input, and output the optimal response strategy through predefined decision rules.

[0145] The MES operation instruction generation and distribution module 360 ​​is used to take the optimal response strategy as input, encapsulate it according to the structured event object format that the MES system can recognize, and generate operation instructions to drive the MES workflow; and send the operation instructions to the MES system through the API interface to drive the MES system to execute at least one of the workflows of creating work orders, updating Kanban, or sending notifications.

[0146] The production line status monitoring and MES integration notification system of this embodiment is used to implement the aforementioned production line status monitoring and MES integration notification method. Therefore, the specific implementation of this system can be found in the embodiment section of the production line status monitoring and MES integration notification method above. Thus, its specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0147] Furthermore, since the production line status monitoring and MES integration notification system in this embodiment is used to implement the aforementioned production line status monitoring and MES integration notification method, its function corresponds to the function of the above method, and will not be repeated here.

[0148] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for production line status monitoring and MES integration notification, characterized in that, Includes the following steps: Multimodal data, including equipment parameters, process quality and material information, is collected. After preprocessing, the multimodal data is associated with and encapsulated with the unique identifier and timestamp of the production unit in the production line to construct a unified state context object. This object contains the original state information, data vector and metadata reported by each device. Spatiotemporal alignment is performed on the multi-source asynchronous data contained in the state context object, and dynamic weights are assigned to the data vectors from different devices in the state context object based on the current process stage for data fusion to obtain the fused feature vector; The feature vector is input into a pre-trained temporal pattern recognition model to infer the current state of the production line with semantics; the current state of the production line refers to a semantic description that includes at least one of the following: equipment operation stage, process completion status, or quality defect type. The state context object and the current state of the production line are linked through data association and denoising algorithms to generate an enhanced data set; With enhanced dataset collaboration as input, the system uses an anomaly detection model and a root cause analysis algorithm based on multi-source information association to output state events containing anomaly type, severity level, and root cause inference results. The state events and business information obtained from the MES system are used as inputs, and the optimal response strategy is output through predefined decision rules. The optimal response strategy is taken as input and encapsulated in a structured event object format recognizable by the MES system to generate operation instructions that drive the MES workflow. The operation instructions are sent to the MES system via the API interface to drive the MES system to execute at least one workflow, such as creating a work order, updating the Kanban board, or sending a notification. Specifically, the process involves spatiotemporal alignment of the multi-source asynchronous data contained in the state context object, and dynamic weighting of data vectors from different devices within the state context object based on the current process stage for data fusion, resulting in a fused feature vector. This includes: Using the production line's global clock or a unified timing reference as a reference, all data points are interpolated or resampled to align the data points to a unified, equally spaced time series to achieve time alignment. Using the production unit currently being processed on the production line as the spatial reference, spatial alignment is achieved by filtering and associating all equipment data related to that production unit within the same time window through the unique identifier of the production unit. For the spatio-temporally aligned data, process stage-based dynamic weighting fusion is performed, including: determining a current process stage, assigning a dynamic weight value to each data source based on the current process stage, and for a target state feature or index that needs to be fused, performing fusion through the following formula to obtain a fused feature vector , wherein, denotes the total number of data sources participating in the fusion, is the dynamic weight of the th data source at the current process stage, is the data value of the th data source after pre-processing and spatio-temporal alignment, is the data quality score of the th data source, th data source after pre-processing and spatio-temporal alignment. The pre-trained temporal pattern recognition model includes an input layer, a feature encoder, a state decoder, and a state semantic mapping layer. Specifically, the feature vector is input into a pre-trained temporal pattern recognition model to infer the semantically meaningful current state of the production line, including: A fused feature vector sequence is constructed based on the fused feature vectors, and this fused feature vector sequence is input into a pre-trained temporal pattern recognition model, where it is received by the model's input layer. The input layer outputs the fused feature vector sequence to the feature encoder, which learns the long-term and short-term dependencies in the sequence through a gating mechanism and outputs a context vector representing the production line operation mode within the entire time window; the feature encoder is a multi-layer LSTM network. The context vector is input to the state decoder, which calculates and outputs the state probability distribution vector through forward propagation. , Indicates that the production line is in the first stage. The confidence probability of a predefined state; where the state decoder is a fully connected neural network; The state probability distribution vector is input to the state semantic mapping layer. This layer determines whether the maximum probability in the state probability distribution vector is not less than the confidence threshold. If so, the state ID J = argmax(P) with the highest confidence is selected as the output. Then, by querying the predefined state-semantic mapping table, the state ID J is mapped to the corresponding semantic current state of the production line. Otherwise, "state unknown" is output or a manual review process is triggered. The state-semantic mapping table defines the correspondence between different states and natural language descriptions.

2. The line status monitoring and MES integration notification method of claim 1, wherein, The data quality score of the first data source is calculated by the following steps: The completeness score is obtained by comparing the actual number of data points received with the expected number of data points received. A freshness score is obtained using an indicator function based on the difference between the current time and the data timestamp. Calculate a reasonableness score based on a predefined reasonable range and tolerance range of data values; The data quality score is obtained by weighting and summing the integrity score, freshness score, and reasonableness score.

3. The method of line status monitoring and MES integration notification of claim 1, wherein, The state context object is linked to the current production line state using data association and denoising algorithms to generate an enhanced dataset, specifically including: Based on the semantics of the current state of the production line, extract the relevant subset of equipment parameter data from the state context object from the predefined state-data association mapping library; Entity association is achieved by associating the data subset using the unique identifier of the production unit in the state context object as the key; Based on the process topology knowledge of the production line, logical links are used to connect data from different equipment in the associated data subset that correspond to the same physical location or process step to achieve logical association. The physical rationality of the data is verified by using the physical knowledge model of the equipment or process, and outliers are identified by the dynamic threshold method. Based on the verification and identification results, the logically correlated data is denoised and corrected. The processed data and its associated metadata are encapsulated to output an enhanced dataset.

4. The line status monitoring and MES integration notification method of claim 3, wherein, Taking enhanced dataset collaboration as input, the algorithm combines an anomaly detection model with a root cause analysis algorithm based on multi-source information association to output state events that include anomaly type, severity level, and root cause inference results. Specifically, these include: The target parameter vector in the augmented dataset is input into a pre-trained unsupervised anomaly detection model. The model calculates the reconstruction error of the input vector and compares the reconstruction error with a preset threshold. If the reconstruction error exceeds the preset threshold, the production line status is determined to be abnormal, and an anomaly score and an anomaly dimension identifier are output. The anomaly dimension identifier indicates which parameter or group of parameters has become abnormal. The unsupervised anomaly detection model is an unsupervised learning model based on an autoencoder. The abnormal dimension identifier and the logically related metadata contained in the enhanced data set are used as evidence input into the production line knowledge graph. Starting from the node corresponding to the evidence, the upstream traversal is carried out along the causal edge to find all possible source nodes and form a set of root cause hypotheses. For each hypothesis in the root cause hypothesis set, a confidence level assessment is performed. Based on the assessment results, all root cause hypotheses are sorted by confidence level, and one or more hypotheses with the highest confidence level are selected as the root cause inference results. The confidence level assessment is based on at least one of the following factors: the path distance and relationship weight between the hypothesis node and the evidence node in the knowledge graph, the co-occurrence probability of the hypothesis and the evidence in historical data, and the magnitude of the current anomaly score. The severity level of an event is determined based on the anomaly score, and the anomaly type is determined based on the anomaly dimension identifier. The severity level, anomaly type, and root cause inference results of the event are then encapsulated into a structured state event object.

5. The line status monitoring and MES integration notification method of claim 4, wherein, Taking the state events and business information obtained from the MES system as input, and using predefined decision rules, the optimal response strategy is output, specifically including: The state time is encapsulated together with the business information obtained from the MES system to construct a complete decision context object; the business information includes the priority of the current production order, the order delivery date, the batch information of work-in-process on the production line, the inventory status of alternative materials, and the workload status of relevant maintenance personnel; Input the decision context object into the strategy decision engine, which predefines several decision dimensions and stores a strategy rule base. Each rule is in the form of "IF<condition>THEN<candidate strategy>". The decision context is matched with the conditions in the policy rule base, and all candidate policies that meet the conditions are activated to form an initial policy candidate set. For each candidate strategy in the initial strategy candidate set, a comprehensive utility value is calculated based on the utility impact value of each decision dimension. The candidate strategy with the highest comprehensive utility value is selected as the optimal response strategy. The optimal response strategy is a structured instruction object that includes the action objective, the execution object, the required resources, and the constraints.

6. The line status monitoring and MES integration notification method of claim 5, wherein, Decision-making dimensions include order delivery, quality costs, equipment risks, and production efficiency; The overall utility value is calculated based on the utility impact value of each decision dimension, specifically including: Calculate the utility impact value for order delivery based on the expected production delay time caused by the strategy; Calculate the utility impact value of the quality cost dimension based on the effect of the strategy on the defect escape rate; The utility impact value of the equipment risk dimension is calculated based on the degree to which the strategy mitigates the equipment risk index. Calculate the utility impact value in the production efficiency dimension based on the changes in capacity utilization caused by the strategy; The overall utility value is obtained by weighting and summing the utility impact values ​​of each dimension.

7. The line status monitoring and MES integration notification method of claim 5, wherein, The optimal response strategy is taken as input and encapsulated according to the structured event object format recognizable by the MES system to generate operation instructions that drive the MES workflow, specifically including: Based on the action objective in the optimal response strategy, query the strategy-MES operation mapping table to determine at least one MES atomic operation; the MES atomic operation includes at least one of creating a work order, updating the production Kanban, reserving materials, and sending a notification. Load parameter templates for MES atomic operations and populate the templates with information extracted from the optimal response strategy, state events, and state context objects; The filled parameters are encapsulated with the corresponding atomic operation identifiers to generate a structured event object that conforms to the MES interface specification; Based on the logical dependencies between MES atomic operations, structured event objects are sorted or grouped to generate the operation instructions; when there are multiple structured event objects, their execution order is determined according to predefined transaction logic to generate an ordered sequence of operation instructions.

8. A production line status monitoring and MES integration notification system, characterized by, The system is used to perform the method of claim 1, comprising: The State Context Object Construction Module is used to collect multimodal data, including equipment parameters, process quality and material information. After preprocessing the multimodal data, it associates and encapsulates it with the unique identifier and timestamp of the production unit in the production line to construct a unified State Context Object. This object contains the original state information, data vector and metadata reported by each device. The data fusion module is used to perform spatiotemporal alignment on the multi-source asynchronous data contained in the state context object, and to assign dynamic weights to the data vectors from different devices in the state context object based on the current process stage to perform data fusion and obtain the fused feature vector. The production line current state reasoning module is used to input the feature vector into a pre-trained temporal pattern recognition model to reason out the semantic current state of the production line; the current state of the production line refers to a semantic description that includes at least one of the following: equipment operation stage, process completion status, or quality defect type. The state event inference module is used to generate an enhanced dataset by associating the state context object with the current state of the production line through data association and denoising algorithms; and using the enhanced dataset as input, it outputs state events containing anomaly type, severity level and root cause inference results through an anomaly detection model and a root cause analysis algorithm based on multi-source information association. The optimal response strategy output module is used to take the status event and the business information obtained from the MES system as input, and output the optimal response strategy through predefined decision rules. The MES operation instruction generation and distribution module is used to take the optimal response strategy as input, encapsulate it according to the structured event object format that the MES system can recognize, and generate operation instructions to drive the MES workflow; and send the operation instructions to the MES system through the API interface to drive the MES system to execute at least one of the workflows of creating work orders, updating Kanban, or sending notifications.

Images