AI-based production process management method and system of intelligent electric meter shell

Abstract

The invention discloses an AI-based production process management method and system for an intelligent electric meter shell, and relates to the technical field of intelligent manufacturing, and the method comprises the steps: carrying out the online quality estimation through the deployment of a plurality of sensors, collection and preprocessing of process data, combination with an offline quality label of an MES end, and through an artificial neural network regression device; calculating a health degree score according to the quality estimation, and obtaining a quality health degree flow by using the health degree score; based on the quality health degree flow, predicting the quality health degree through an LSTM model, calculating the gradient of the health degree to a set value in combination with a calculus method, solving an optimal set value by using a reinforcement learning algorithm, generating a real-time quality prediction and optimization set value based on the optimal set value, and packaging the real-time quality prediction and optimization set value into a standard message format; a single message is read, the set value is issued to the PLC and the MES at the same time through the RPA, and the write-in operation is completed. According to the invention, the quality and efficiency collaborative optimization capability in the production process of the intelligent electric meter shell is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a production process management method and system for smart meter housings based on AI. Background Technology

[0002] Against the backdrop of deep integration between intelligent manufacturing and industry, the production process of smart meter housings, as key structural components of power metering equipment, places extremely high demands on accuracy, consistency, and traceability. The collaborative capabilities of manufacturing execution systems (MES) and process control systems (such as PLCs) have been significantly enhanced, driving key processes such as injection molding and assembly towards data-driven and closed-loop optimization. In particular, in terms of soft sensor technology, online quality estimation methods based on data-driven models such as artificial neural networks (ANNs) have gradually replaced traditional offline sampling inspection modes, achieving real-time mapping of difficult-to-measure quality indicators such as melt index and wall thickness deviation. At the same time, time-series modeling methods such as LSTM have been introduced into process prediction, combined with reinforcement learning for parameter self-optimization, initially forming an intelligent control closed loop of "perception-prediction-decision-execution". Currently, most solutions are still limited to local optimization of a single link, lacking a unified modeling and dynamic control mechanism for the quality and health of the entire process. Several key bottlenecks remain in the practical application of existing technologies: First, quality assessment often relies on static thresholds or empirical rules, failing to construct a dynamic health index system with physical meaning and quantifiability, making it difficult to accurately reflect the comprehensive quality status under multi-variable coupling. Second, setpoint optimization often ignores the process structure characteristics in the process path (such as parallel and branch logic), leading to a disconnect between the optimization strategy and actual production line behavior, and even causing unexpected disturbances. Third, although some systems introduce AI for parameter recommendation, they lack a reliable integration mechanism with PLC / MES, and the ability to confirm execution, track effects, and roll back anomalies after setpoint issuance is weak, making it difficult to meet industrial-grade SLA (Service Level Agreement) requirements. Fourth, existing reinforcement learning strategies are mostly based on simple reward functions and do not integrate root cause diagnostic information (such as the potential defect causes output by SVM), resulting in a lack of interpretability and robustness in action selection. The lack of a collaborative evaluation mechanism for production efficiency and quality efficiency makes it difficult for the system to automatically identify "high-quality steady state" and trigger continuous optimization under dynamic operating conditions. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides an AI-based method and system for managing the production process of smart meter casings, addressing the problems of crude quality and health modeling, insufficient process structure perception, and weak multi-objective collaborative optimization capabilities in smart meter casing production process management technology.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an AI-based method for managing the production process of smart meter casings, which includes: collecting and preprocessing process data by deploying multiple sensors, combining offline quality labels from the MES terminal, performing online quality estimation through an artificial neural network regressor, calculating a health score based on the quality estimation, and obtaining a quality health stream using the health score; Based on the quality and health stream, the quality and health level is predicted by the LSTM model, the gradient of the health level with respect to the set value is calculated by the calculus method, the optimal set value is solved by the reinforcement learning algorithm, and real-time quality prediction and optimized set value are generated based on the optimal set value and packaged into a standard message format. Read a single message, send the set value to both PLC and MES simultaneously via RPA and complete the write operation, perform readback confirmation and execution effect evaluation within the set SLA time limit, and summarize and record the execution status based on the effect evaluation. Based on the aggregated records, an instant reward signal is constructed, and the expected cumulative reward value is calculated based on the instant reward signal using the Q-learning algorithm. The pre-trained SVM model outputs enhanced feature sequences, which are then combined with the expected cumulative reward value and reward signal to automatically adjust and execute the set values. A quality report is generated based on the execution results, and production efficiency and quality efficiency values ​​are calculated based on the quality report to obtain a comprehensive evaluation score. The current working condition is automatically judged based on the comprehensive evaluation score.

[0006] As a preferred embodiment of the AI-based smart meter casing production process management method described in this invention, the method involves: deploying multiple types of hardware sensors and uniformly connecting them to an edge gateway for data sampling; writing the timestamped sampled data into the message queue of the edge gateway; and simultaneously extracting offline quality labels for each piece and batch at the MES quality inspection terminal. The message queue of the edge gateway is preprocessed. Based on soft sensor technology, a data-driven online soft sensor is selected, and an artificial neural network regressor model is trained using a supervised learning method. The trained artificial neural network regressor model is then embedded into the edge gateway, and the real-time collected feature sequences are processed. Input an artificial neural network regressor model and output real-time quality estimation variables. Calculate single indicator scores The total score is calculated by aggregating the scores of individual indicators using a weighted geometric mean. Obtain a quality health stream sorted by timestamp.

[0007] As a preferred embodiment of the AI-based smart meter casing production process management method of the present invention, the following steps are taken: Based on the quality health stream, the quality health is predicted by an LSTM model, the gradient of the health with respect to a set value is calculated by calculus, the optimal set value is solved by reinforcement learning algorithm, real-time quality prediction and optimized set value are generated based on the optimal set value, and encapsulated into a standard message format including the quality health stream with work order ID and timestamp, as well as synchronized equipment events and process feature data. The time series is sliced ​​according to each work order to generate an event log, and the event log is analyzed by process mining algorithm to automatically discover the temporal flow relationship between activities and construct a direct follow graph. The sliced ​​event logs are input into the Inductive Miner algorithm to form a bottleneck list B. Each trajectory in the event logs is compared with the process model M to generate a clean sample domain Ω, which contains time series triples. The process characteristics at the current moment and setting value splicing into joint input Input an LSTM network and output a quality prediction for the next time window. Quality prediction Substituting into the health mapping function, we obtain the predicted health score. Calculate the gradient value of health quality with respect to the set value. Through optimization algorithms in reinforcement learning, based on the objective function Solve for the optimal next time step setpoint within the current time window. Obtain optimized settings Then, maintain the current process characteristics. Unchanged, and splice as new input The data is fed into a trained LSTM model for forward inference, and the model outputs the corresponding quality prediction. The predicted health score is recalculated using the health score mapping formula. Forming a triplet It is then encapsulated into a standard message format.

[0008] As a preferred embodiment of the AI-based smart meter casing production process management method of the present invention, the following steps are described: Reading a single message, simultaneously sending the set value to the PLC and MES via RPA and completing the write operation, performing readback confirmation and execution effect evaluation within the set SLA time limit, summarizing and recording the execution status based on the effect evaluation, and reading the single message in the order of arrival, and then recommending the set value... Simultaneously, the data is sent to the PLC and MES. The writing is completed by simulating manual operation through RPA. The orchestrator triggers two controlled RPA tasks at the same time, including the MES-side RPA and the PLC-side RPA to log in to SCADA. The setpoints are written one by one according to the address table. The operation status and timestamp are recorded after each task is completed. Set up a service level agreement. If any task times out, fails to write, or if subsequent monitoring finds that the quality health does not improve as expected, immediately trigger a rollback plan, restore to the settings snapshot before the change, and enter the exception handling process to ensure production safety and continuity. After the setpoints are issued and the SLA readback confirmation is completed, the execution effect evaluation begins, including the actual quality health within the observation window. The predicted target was not achieved. If the set value is not met, the execution will be marked as "poor performance", and the actual set value and special sequence information issued this time will be packaged into a structured experience sample. If health statistics If the expected goal is achieved or exceeded, it is marked as "Execution Completed". The complete status package of this execution is summarized and written to the experience queue, with the format: {Work Order ID, Timestamp, Actual Issuance Setting u, Readback Status, Health Statistics in the Observation Window}. , Execution conclusion.

[0009] As a preferred embodiment of the AI-based smart meter casing production process management method of the present invention, wherein: the step of constructing an instant reward signal based on the summary record, and calculating the expected cumulative reward value based on the instant reward signal using the Q-learning algorithm, refers to initializing the Q-value table for each state and corresponding action in the production process, with all Q values ​​initially set to zero; Construct an instant reward signal, awarding positive rewards for positive deviations and negative rewards for negative deviations. Using the Q-learning algorithm based on immediate rewards Get updated As a preferred embodiment of the AI-based smart meter casing production process management method of the present invention, the following steps are described: The enhanced feature sequence is output through a pre-trained SVM model; the setpoint is automatically adjusted and executed by combining the expected cumulative reward value and the reward signal; a quality report is generated based on the execution result; the target setpoint and real-time collected production data are received and converted into feature vectors; these vectors are input into the trained SVM model; the potential root causes of the current quality deviation are output; and the SVM diagnostic results are encoded as additional feature dimensions to form the enhanced feature sequence. Give the master control policy based on Q-learning, and let Q-learning determine the current state. renew Combined with reward signals Employ strategies to select from the action space Biggest movement The process is executed, and the results are continuously monitored. If the expected outcome is not achieved, Q-learning triggers a new round of exploration or regression, incorporating the explanations provided by SVM and the actions performed by Q-learning. The actual changes in health status are integrated to generate a structured quality report.

[0010] As a preferred embodiment of the AI-based smart meter casing production process management method of the present invention, the following steps are included: calculating production efficiency and quality efficiency values ​​based on quality reports, obtaining a comprehensive evaluation score, and automatically determining and adjusting production parameters and health status based on the comprehensive evaluation score. Using real-time feedback data and historical production data, production efficiency is calculated to obtain a production efficiency value. Calculate the mass efficiency value Construct a multi-objective comprehensive evaluation score If the current comprehensive evaluation score If the value is greater than or equal to threshold A, it is considered a "high-quality steady state". The current parameters are confirmed to be valid and continuously monitored. If the value is less than threshold A, then the collaborative optimization mechanism is triggered, and the current value of Y is adjusted accordingly. The parameters and bottleneck list information are packaged into experience samples for updating the Q-learning strategy.

[0011] Secondly, this invention provides an AI-based smart meter casing production process management system, including: The data acquisition and online quality estimation module is used to realize online quality index estimation and quality health stream generation based on real-time process data through multi-sensor fusion and ANN soft sensor model; The Quality and Health Prediction and Setpoint Optimization Module is used to predict health using LSTM and combine gradient and reinforcement learning algorithms to dynamically solve and output the optimal equipment setpoints that satisfy process constraints. The setpoint issuance and execution verification module is used to synchronously issue setpoints to PLC and MES via RPA, complete readback confirmation, effect evaluation and abnormal rollback within the SLA time limit, and record the structured execution status. The reinforcement learning reward modeling and strategy update module is used to construct immediate reward signals based on performance, and adopts Q-learning to continuously optimize the set value adjustment strategy to maximize long-term quality health benefits. Root Cause Analysis and Enhanced Decision Module: Utilizes pre-trained SVM to diagnose the root causes of quality deviations, generates enhanced features to assist Q-learning in making interpretable setpoint adjustments, and outputs a structured quality report.

[0012] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the production process management method for AI-based smart meter housings as described in the first aspect of the present invention.

[0013] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the production process management method for AI-based smart meter housings as described in the first aspect of the present invention.

[0014] The beneficial effects of this invention are as follows: By deploying multi-dimensional sensors at key process points such as injection molding, cooling, and assembly, and uniformly connecting them to an edge gateway, high-efficiency and high-consistency process data acquisition is achieved; by combining MES offline quality inspection labels to construct an ANN soft sensor model of "easy-to-measure input → difficult-to-measure output", real-time online quality estimation is generated; based on the dynamic health mapping mechanism of process specifications, a quantifiable and traceable quality health flow is formed; process mining technology is used to automatically discover the real production line process structure from event logs, accurately identify bottleneck activities and deviation windows, and select clean samples to train an LSTM prediction model; the gradient of health with respect to the set value is calculated using calculus, and the optimal process set value is solved by combining a reinforcement learning objective function with stationarity constraints and safety boundaries; and readback verification and effect evaluation are completed within the SLA time limit, thereby realizing a leap from "passive detection" to "active prediction and precise optimization" across the entire chain. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the AI-based smart meter casing production process management method in Example 1.

[0017] Figure 2 This is a schematic diagram of the production process management system for the AI-based smart meter casing in Example 1. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0021] Example 1, referring to Figures 1 to 2 This is the first embodiment of the present invention, which provides an AI-based method for managing the production process of smart meter casings, including the following steps: S1. By deploying multiple sensors, process data is collected and preprocessed. Combined with offline quality labels from the MES terminal, online quality estimation is performed using an artificial neural network regressor. Based on the quality estimation, a health score is calculated, and the health score is used to obtain the quality health stream. Based on the quality and health stream, the quality and health level is predicted by the LSTM model, the gradient of the health level with respect to the set value is calculated by the calculus method, the optimal set value is solved by the reinforcement learning algorithm, and real-time quality prediction and optimized set value are generated based on the optimal set value and packaged into a standard message format. Specifically, by deploying multiple sensors, process data is collected and preprocessed. Combined with offline quality tags at the MES end, online quality estimates are calculated using an artificial neural network regressor. Based on the quality estimates, a health score is calculated, and the health score is used to obtain a quality health index. Hardware sensors for temperature T, mold cavity pressure P, screw speed N, mold temperature water flow F, ambient humidity H, and vibration V are deployed at specific points in processes such as injection molding, cooling, and assembly. These sensors are then uniformly connected to an edge gateway (OPC-UA / Modbus→MQTT) for data sampling, and the timestamped sampled data is written to the message queue of the edge gateway. Simultaneously extract offline quality labels for each item and batch at the MES quality inspection terminal. (e.g., melt index MI, shell purity Pur, wall thickness deviation Th) and aligned with process data by combining work order ID with time window; The message queue of the edge gateway is preprocessed, including clock calibration and hysteresis compensation, noise reduction, anomaly removal, missing data imputation, and scaling, to obtain a standardized feature sequence. Based on "Easy-to-Test Input" To address the "difficult-to-measure output" of existing soft sensor technologies, this paper selects data-driven online soft sensors and trains an artificial neural network (ANN) regressor model using supervised learning methods (such as mean squared error loss + Adam optimizer). Specifically, it uses standardized multivariate process feature sequences from historical aligned data. As input, the corresponding MES offline quality inspection label As the true target output, input-output sample pairs are constructed, and mean squared error (MSE) is used as the loss function to measure the model's predicted values. With real labels The deviation between them is calculated, and the weight parameters of the ANN are iteratively updated through the Adam optimizer to minimize the overall loss until the model converges, thus obtaining an artificial neural network regressor model that can accurately map from easily measurable process variables to difficult-to-measurable quality indicators. The trained artificial neural network regressor model is embedded into the edge gateway, and the real-time collected feature sequences are stored therein. Input an artificial neural network regressor model and output real-time quality estimation variables. Obtain online quality estimation Extract the pre-processed process feature sequence from the edge side. Online estimation with soft sensors Alignment is performed using "work order ID + deterministic time window (piece-first, batch-later rollback)" and resampling is performed to a unified timestamp. Upper and lower limits of process specifications are set for each quality indicator. Standards Center Half For each time point and each quality indicator (e.g., MI, Pur, Th), determine the reference value based on whether there is a label. (with tags) Unlabeled revert to Then, the single-indicator score is calculated using the continuous health mapping given within the same segment, using the following formula: in, This refers to limiting the optimized and updated settings to preset upper and lower bounds, such as the [0,1] interval. The target value of a quality indicator is the ideal set value of the indicator (e.g., the target value of melt index MI is 2.5 g / 10min), used to measure the degree of deviation between the current value and the target. The total score is calculated using a weighted geometric mean. The formula is: in, It is the first The weighting coefficients of each quality indicator (such as MI, Pur, Th) represent their relative importance in the overall health assessment, and satisfy the following requirements: It is used to reflect the priority differences of different indicators when aggregating multiple indicators, and is obtained by an optimization learning method based on historical data; Obtain the quality health stream sorted by timestamp, and record it as follows. By deploying multi-dimensional sensors for temperature, pressure, screw speed, mold temperature, water flow rate, ambient humidity, and vibration at key workstations, and combining them with edge gateways, high-efficiency data acquisition and standardized preprocessing are achieved. An innovative online soft sensor model based on an ANN regressor is constructed to accurately map easily measurable process variables to difficult-to-measure quality indicators such as melt index, shell purity, and wall thickness deviation. A dynamic health mapping mechanism integrating process specification upper and lower limits with target values ​​is introduced. Based on a strategy of "using actual measurement for labeled data and reverting to prediction for unlabeled data," single-indicator health scores are generated and aggregated using a weighted geometric average to form a comprehensive health score with physical meaning and priority distinction, ultimately forming a quality health score stream indexed by work order ID and timestamp. This method not only achieves real-time, continuous, and interpretable assessment of quality status, providing a highly reliable data foundation for subsequent predictive optimization and closed-loop control, but also overcomes the core bottlenecks of "quality black box" and "assessment lag" in traditional manufacturing systems.

[0022] Furthermore, based on the quality and health stream, the quality and health level is predicted using an LSTM model. The gradient of the health level with respect to the setpoint is calculated using calculus methods. The optimal setpoint is solved using a reinforcement learning algorithm. Based on the optimal setpoint, real-time quality prediction and optimized setpoint are generated and encapsulated into a standard message format, which includes the quality and health stream with work order ID and timestamp, as well as synchronized device events and process feature data. The time series is sliced ​​for each work order to generate an event log (each log includes the activity name, start / end time, and average quality and health level of the interval). The time series of the same work order ID is sliced ​​according to process thresholds (such as mold closing / injection start / holding pressure end / cooling completion / demolding) to generate event logs containing activities, start and end times, and quality health. Process mining algorithms (such as DFG) are applied to analyze the event logs to automatically discover the temporal flow relationships between activities and construct a direct follow graph (representing the actual process path). The sliced ​​event logs are converted to the XES standard format and input into the Inductive Miner algorithm to mine a hierarchical process model M (such as Petri net) to more accurately express the parallel operations and non-trivial branch logic in the actual process. Then, based on the path replay on M, the performance indicators (average duration, dwell time, average health) of each node and edge are calculated. Combined with the performance data, bottleneck activities (the key process links with the longest average duration and the greatest impact on the overall quality in the production process) are identified to form a bottleneck list B. Each trajectory in the event log (referring to the sequence of activities a work order experiences in chronological order during production (e.g., "injection molding → holding pressure → cooling → demolding")) is compared with the process model M. A fitness check method based on playback is used to identify parts of the actual trajectory that cannot be correctly reproduced by the model. For mismatched parts, the start and end timestamps are recorded and marked as deviation windows D. During the modeling data selection phase, time periods included in the bottleneck list B but not within any deviation window D are retained as clean sample domains Ω, containing time series triples. in, This is the process feature vector of the previous window (such as the standardized measured values ​​of temperature, pressure, and screw speed). This is a vector of settings for the current window (such as target settings for barrel temperature, mold temperature, and injection speed). For the window's quality health score (normalized to) ); The process characteristics at the current moment and setting value splicing into joint input The input is fed into an LSTM network, which combines the hidden state from the previous time step with the current input. By fusing historical and current information through a gating mechanism (forget gate, input gate, candidate cell state, cell state update, output gate, and hidden state update), the current hidden state is generated. This is then mapped through a fully connected layer to predict the quality of the next time window. (Such as outputs from soft sensors for melt viscosity, shrinkage rate, etc.); Quality prediction Substituting into the health mapping function, we obtain the predicted health score. The formula is: in, Δ represents the target center value (such as the ideal process point), and Δ represents the tolerance range (half of the upper and lower limits). The gradient value of health quality with respect to a set value is calculated using calculus. The formula is: in, For the partial derivative of health status with respect to quality estimation, The partial derivative of the quality prediction output with respect to the setpoint; Through optimization algorithms in reinforcement learning (RL), based on the objective function Solve for the optimal next time step setpoint within the current time window. The formula is: in, To maximize the predicted quality of health (i.e., minimize the degree of deviation from the target). To punish drastic changes in the set value, To adjust the weights, Ensuring control stability and energy consumption constraints (referring to the sequence of activities a work order undergoes in chronological order during production, such as "injection molding → holding pressure → cooling → demolding")) This includes the process feasibility domain (e.g., the temperature cannot be lower than the material's melting point) and equipment safety boundaries (e.g., the upper limit of motor speed). Obtain optimized settings Then, maintain the current process characteristics. Unchanged, and splice as new input The data is fed into a trained LSTM model for forward inference, and the model outputs the corresponding quality prediction. The predicted health score is recalculated using the health score mapping formula. Forming a triplet Output-based triples Combined with the work order ID and timestamp from the original input And the identified bottleneck activities and corresponding time window information, will set values Convert to setpoint key-value pairs with key names. The predicted health status is packaged together into a standard message format. Based on quality and health metrics and equipment event logs, a hierarchical process model reflecting real parallel and branching logic is constructed using algorithms such as Inductive Miner. Fitness checks are used to accurately identify bottleneck activities and abnormal deviation windows, thereby selecting high-confidence "clean sample domains" for model training. An LSTM network is employed to jointly encode process features and current setpoints, enabling accurate prediction of future quality and health. Calculus methods are used to explicitly calculate the sensitivity gradient of health to setpoints, providing interpretable directional guidance for optimization. A multi-objective reinforcement learning objective function is then designed, integrating health maximization, setpoint smoothness, and gradient stability, to solve for the optimal setpoints under strict constraints of the process feasible domain and equipment safety boundaries. Finally, the optimization results, along with work order context and bottleneck information, are encapsulated into a standard message format, laying a structured foundation for subsequent RPA automatic deployment and closed-loop verification. This overcomes the problem of traditional "black box" parameter tuning being disconnected from actual production line behavior, achieving a leap from "experience-driven" to an intelligent optimization paradigm of "model-driven, process-aware, gradient-guided, and safety-constrained."

[0023] S2. Read a single message, send the set value to the PLC and MES simultaneously via RPA and complete the write operation, perform readback confirmation and execution effect evaluation within the set SLA time limit, and summarize and record the execution status based on the effect evaluation. Based on the aggregated records, an instant reward signal is constructed, and the expected cumulative reward value is calculated based on the instant reward signal using the Q-learning algorithm. Specifically, a single message is read, and the set value is simultaneously sent to the PLC and MES via RPA to complete the write operation. Within the set SLA time limit, a readback confirmation and execution effect evaluation are performed. Based on the effect evaluation, the execution status is summarized and recorded. The process involves reading single messages in the order of arrival, verifying field integrity, calculating the message fingerprint and comparing it with the most recent execution record, eliminating duplicate messages, and verifying that the timestamp is not earlier than the current system clock. ( If the device communication maximum delay is reached, otherwise the request will be rejected and the reason recorded. Recommended settings Simultaneously, the data is sent to both the PLC (control layer) and MES (management layer). The write operation is simulated using RPA to complete the process. Specifically, the write address table is created for the PLC, specifying the variable name, data type, and corresponding target value for each control loop. The work order change record (including fields, old value, new value, and reason) is created for the MES. The orchestrator simultaneously triggers two controlled RPA tasks, including the MES-side RPA: automatically logging into the MES system, opening the work order change transaction, and entering the recommended settings. Related bottleneck activity names, expected quality health Obtain the transaction number after submission for tracking purposes; The PLC-side RPA logs into the SCADA or edge gateway and writes the settings point item by item according to the address table. After each item is completed, the operation status (success / failure) and timestamp are recorded. Set a Service Level Agreement (SLA) that all write operations must be completed within 30 seconds and readback confirmation must be completed within 60 seconds (to verify that the settings are effective and the process is stable). If any task times out, fails to write, or if subsequent monitoring finds that the quality health does not improve as expected, the rollback plan will be triggered immediately to restore the settings snapshot before the change and enter the exception handling process to ensure production safety and continuity. The health statistics are recalculated using the same health mapping formula within the observation time window after the set value is executed, based on real-time acquired process data and soft sensor outputs (or MES quality inspection labels). (e.g., mean, variance, increase rate).

[0024] After the setpoints are issued and the SLA readback confirmation is completed, the execution effect evaluation begins, including the actual quality health within the observation window. The predicted target was not achieved. If the set value (e.g., 80%, i.e., the minimum effective improvement is not achieved) is not met, the execution will be marked as "poor performance", and the actual set value u and feature sequence will be sent this time. The information is packaged into a structured experience sample; If health statistics If the expected goal is met or exceeded, it is marked as "Execution Achieved"; Summarize the complete status packet of this execution and write it to the experience queue. The format is: {work order ID, timestamp, actual issuance setting u, readback status, health statistics in the observation window}. ,Execution conclusion (achieved / unsatisfactory / reverted)}.

[0025] By rigorously verifying message integrity, timeliness, and uniqueness, the validity of instructions is ensured. An orchestrator is used to trigger RPA tasks for PLCs and MES in parallel, enabling atomic and traceable writing of setpoints into the equipment control and production management system. Two-way tracking is achieved through transaction numbers and operation logs. A Service Level Agreement (SLA) is introduced, and execution effectiveness is dynamically evaluated in conjunction with real-time health monitoring. Ultimately, the entire execution process is structured and recorded as an experience sample containing setpoints, observed health status, readback status, and conclusions. This provides high-quality feedback data for subsequent reinforcement learning strategy iterations, effectively bridging the gap between "intelligent decision-making" and "industrial execution." It addresses the core pain points of traditional systems—"calculable, but not applicable, unverifiable, and unrecoverable"—significantly improving the engineering implementation capability and operational robustness of AI-driven process optimization.

[0026] Furthermore, based on the summary records, an instant reward signal is constructed. Based on the instant reward signal, the expected cumulative reward value is calculated using the Q-learning algorithm. This initializes the Q-value table for each state (such as temperature, pressure, etc.) and corresponding action (such as adjusting temperature, pressure, etc.) in the production process, with all Q values ​​initially set to zero. Construct an instant reward signal, awarding a positive reward for positive deviations and a negative reward for negative deviations, using the following formula: The sign function determines the direction of the deviation (positive / negative), and the absolute value measures the degree of deviation, thus forming a signed reward signal. Using the Q-learning algorithm based on immediate rewards Get updated The updated formula is: in, For current features Execute action (such as adjusting temperature and pressure) The value represents the expected cumulative reward for the "state-action" pair. For the next state The largest of all possible actions value, For learning rate, As a discount factor, in an offline simulation environment, a state-action-reward sequence is constructed using historical execution data, and different discount factors are set. and Q-learning training is performed, and the cumulative reward, convergence speed, or the improvement in the health of the final policy on the validation set is used as evaluation metrics to select the best-performing set of parameters. In industrial scenarios, domain knowledge is often incorporated into the initial design. Then, fine-tune through online A / B testing to obtain the final result. and To address the shortcomings of existing intelligent manufacturing systems, such as the lack of effective feedback mechanisms for optimization strategies, reliance on trial-and-error experience for parameter adjustments, and difficulty in achieving continuous self-learning, this invention constructs a physically meaningful instant reward signal based on structured summary records of execution results. This signal truly reflects the effectiveness of actions (such as temperature and pressure adjustments) in improving quality. The invention employs a Q-learning algorithm to iteratively update the expected cumulative reward of "state-action" pairs. Combined with offline simulation and online A / B testing, the learning rate α and discount factor γ are collaboratively optimized to ensure a balance between convergence speed and long-term benefits. This transforms the actual effects of each process intervention into learnable experience, enabling the system to autonomously extract optimal control strategies from historical executions. This effectively overcomes the limitations of traditional methods, such as "no feedback in optimization, no memory in adjustment, and no evolution in strategy," providing a solid reinforcement learning foundation for continuous adaptive optimization of the smart meter casing production process.

[0027] S3. The enhanced feature sequence is output through the pre-trained SVM model. The set value is automatically adjusted and executed by combining the expected cumulative reward value and the reward signal. A quality report is generated based on the execution result. The production efficiency and quality efficiency values ​​are calculated based on the quality report, and a comprehensive evaluation score is obtained. The current working condition is automatically judged based on the comprehensive evaluation score. Specifically, the enhanced feature sequence output by the pre-trained SVM model is combined with the expected cumulative reward value and reward signal to automatically adjust and execute the set value. Based on the execution result, a quality report is generated. This report receives target set values ​​(such as temperature, pressure, and injection speed) and real-time collected production data (such as actual temperature, pressure, and humidity), converts them into feature vectors, inputs them into the trained SVM model, and outputs the potential root cause of the current quality deviation (such as "excessive temperature causing surface roughness"). The SVM diagnostic result is encoded as an additional feature dimension (such as 0 / 1 flags or normalized deviation) to form the enhanced feature sequence. Given a master control policy based on Q-learning, Q-learning determines the policy based on the current state. renew Combined with reward signals Adopt strategies (such as) (or softmax) select from the action space to make Biggest movement And execute (e.g., adjust temperature settings), continuously monitor the effect after execution, and if the expected result is not achieved, Q-learning will trigger a new round of exploration or rollback, combining the explanation of the cause provided by SVM with the actions executed by Q-learning. The actual changes in health status are integrated to generate a structured quality report, which is then pushed to the management interface.

[0028] A pre-trained SVM model is introduced to diagnose and analyze real-time production data, accurately identify the potential causes of quality deviations, and encode the root cause information into structured additional features to construct an enhanced state representation, thereby improving the semantic understanding ability of reinforcement learning strategies for process anomalies. The master control strategy, based on enhanced feature sequences and immediate reward signals, dynamically selects and executes the optimal action, achieving a shift from "blind trial and error" to "root cause-guided exploration"; the system integrates SVM diagnostic conclusions, The selected action and the actual health status change are integrated to automatically generate a structured quality report with causal logic, which is then pushed to the management interface. This not only supports operators in quickly tracing the source of the problem, but also provides high-value monitoring signals for subsequent strategy iterations, effectively bridging the gap between data-driven optimization and process knowledge.

[0029] Furthermore, the production efficiency and quality efficiency values ​​are calculated based on the quality report, and a comprehensive evaluation score is obtained. Based on the comprehensive evaluation score, the current working conditions are automatically assessed, and adjusted production parameters (such as temperature and pressure) and health status are evaluated. As real-time feedback data and historical production data, it is used to evaluate various key performance indicators (KPIs) in the production process, such as production efficiency, quality fluctuations, and resource consumption, and to calculate production efficiency. in, For the efficiency of the current production process, This is an error adjustment coefficient used to evaluate the deviation between the actual output and the target setting. Data on the process cycle and actual output deviation under different setting values ​​are collected to establish a statistical relationship model between error and efficiency. The optimal model is then fitted using the least squares method. value, Indicates the first Adjustment values ​​for the cycle time of each process stage (such as injection molding, holding pressure, and cooling). For process index, The parameter indicates that it is the "adjusted" value after optimization or feedback correction, reflecting the deviation between the actual production rhythm and the target. K is the total number of process stages (e.g., K=5 means there are 5 stages in total), and k is the loop variable. Calculate the mass efficiency value The formula is used to assess whether the quality meets expectations: Constructing a multi-objective comprehensive evaluation score : Among them, weight and Set by process experts based on production line targets; Collect the score values ​​of all "execution achieved" work orders from the past 6 months, and take the quartile (P25) or the mean minus one standard deviation as the dynamic threshold A to ensure that only high-quality and efficient work conditions are retained. If the current comprehensive evaluation score If the value is greater than or equal to threshold A, it is considered a "high-quality steady state". The current parameters are confirmed to be valid and continuously monitored. If the value is less than threshold A, then the collaborative optimization mechanism is triggered, and the current value of Y is adjusted accordingly. The parameters and bottleneck list information are packaged into experience samples to update the Q-learning strategy, achieving a joint closed-loop optimization of efficiency and quality.

[0030] By introducing a production efficiency model based on process stage cycle deviations and a quality and efficiency indicator centered on health achievement, and combining expert weights to generate a unified comprehensive evaluation score, the system achieves a quantitative characterization of the current "high output and stable quality" level. The system dynamically calculates thresholds based on historical "execution achievement" work orders, avoiding the mismatch problem of fixed thresholds under different production conditions, thereby achieving automatic identification of "high-quality steady state". Efficiency, quality, bottlenecks and set parameters are packaged into structured experience samples to feed back into the Q-learning strategy update, effectively solving the one-sided optimization dilemma of "emphasizing output over quality" or "quality over efficiency" in traditional systems, and realizing a joint closed loop and adaptive evolution of efficiency and quality.

[0031] This embodiment also provides an AI-based smart meter casing production process management system, including: The data acquisition and online quality estimation module is used to realize online quality index estimation and quality health stream generation based on real-time process data through multi-sensor fusion and ANN soft sensor model; The Quality and Health Prediction and Setpoint Optimization Module is used to predict health using LSTM and combine gradient and reinforcement learning algorithms to dynamically solve and output the optimal equipment setpoints that satisfy process constraints. The setpoint issuance and execution verification module is used to synchronously issue setpoints to PLC and MES via RPA, complete readback confirmation, effect evaluation and abnormal rollback within the SLA time limit, and record the structured execution status. The reinforcement learning reward modeling and strategy update module is used to construct immediate reward signals based on performance, and adopts Q-learning to continuously optimize the set value adjustment strategy to maximize long-term quality health benefits. Root Cause Analysis and Enhanced Decision Module: Utilizes pre-trained SVM to diagnose the root causes of quality deviations, generates enhanced features to assist Q-learning in making interpretable setpoint adjustments, and outputs a structured quality report.

[0032] This embodiment also provides a computer device applicable to the production process management method for AI-based smart meter casings, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI-based smart meter casing production process management method proposed in the above embodiment.

[0033] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0034] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the production process management method and system for AI-based smart meter casings as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0035] In summary, this invention achieves highly timely and consistent process data acquisition by deploying multi-dimensional sensors at key points in injection molding, cooling, and assembly processes and uniformly connecting them to an edge gateway. It constructs an ANN soft sensor model with "easy-to-measure input → difficult-to-measure output" using MES offline quality inspection labels, generating real-time online quality estimates. A dynamic health mapping mechanism based on process specifications forms a quantifiable and traceable quality health stream. Process mining technology automatically discovers the real production line process structure from event logs, accurately identifying bottleneck activities and deviation windows, and using this to select clean samples to train an LSTM prediction model. The gradient of health with respect to a set value is calculated using calculus, and the optimal process set value is solved by combining a reinforcement learning objective function with stationarity constraints and safety boundaries. Readback verification and effect evaluation are completed within the SLA timeframe, thus achieving a leap from "passive detection" to "active prediction and precise optimization" across the entire process chain.

[0036] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A production process management method for smart meter casings based on AI, characterized in that: include, By deploying multiple sensors, process data is collected and preprocessed. Combined with offline quality labels from the MES terminal, online quality estimation is performed using an artificial neural network regressor. Based on the quality estimation, a health score is calculated, and the health score is used to obtain a quality health stream. Based on the quality and health stream, the quality and health level is predicted by the LSTM model, the gradient of the health level with respect to the set value is calculated by the calculus method, the optimal set value is solved by the reinforcement learning algorithm, and real-time quality prediction and optimized set value are generated based on the optimal set value and packaged into a standard message format. Read a single message, send the set value to both PLC and MES simultaneously via RPA and complete the write operation, perform readback confirmation and execution effect evaluation within the set SLA time limit, and summarize and record the execution status based on the effect evaluation. Based on the aggregated records, an instant reward signal is constructed, and the expected cumulative reward value is calculated based on the instant reward signal using the Q-learning algorithm. The pre-trained SVM model outputs enhanced feature sequences, which are then combined with the expected cumulative reward value and reward signal to automatically adjust and execute the set values. A quality report is generated based on the execution results, and production efficiency and quality efficiency values ​​are calculated based on the quality report to obtain a comprehensive evaluation score. The current working condition is automatically judged based on the comprehensive evaluation score.

2. The production process management method for AI-based smart meter casings as described in claim 1, characterized in that: The process involves deploying multiple sensors to collect and preprocess process data. Combined with offline quality tags from the MES (Manufacturing Execution System) terminal, an artificial neural network regressor calculates online quality estimates. Based on these estimates, a health score is calculated, and a quality health index is derived. Multiple types of hardware sensors are deployed and uniformly connected to an edge gateway for data sampling. Timestamped sampled data is written to the edge gateway's message queue, and offline quality tags for each item and batch are simultaneously extracted at the MES quality inspection terminal. The message queue of the edge gateway is preprocessed. Based on soft sensor technology, a data-driven online soft sensor is selected, and an artificial neural network regressor model is trained using a supervised learning method. The trained artificial neural network regressor model is then embedded into the edge gateway, and the real-time collected feature sequences are processed. Input an artificial neural network regressor model and output real-time quality estimation variables. Calculate single indicator scores The total score is calculated by aggregating the scores of individual indicators using a weighted geometric mean. Obtain a quality health stream sorted by timestamp.

3. The production process management method for AI-based smart meter casings as described in claim 2, characterized in that: The process involves predicting quality and health based on a quality and health stream using an LSTM model, calculating the gradient of health with respect to a set value using calculus, solving for the optimal set value using a reinforcement learning algorithm, generating real-time quality predictions and optimized set values ​​based on the optimal set value, and encapsulating them into a standard message format containing a quality and health stream with a work order ID and timestamp, along with synchronized device events and process feature data. Event logs are generated by slicing the time series for each work order, and process mining algorithms are applied to analyze the event logs, automatically discovering the temporal flow relationships between activities and constructing a direct follow-up graph. The sliced ​​event logs are input into the Inductive Miner algorithm to form a bottleneck list B. Each trajectory in the event logs is compared with the process model M to generate a clean sample domain Ω, which contains time series triples. ; The process characteristics at the current moment and setting value splicing into joint input Input an LSTM network and output a quality prediction for the next time window. Quality prediction Substituting into the health mapping function, we obtain the predicted health score. ; Calculate the gradient value of health quality with respect to the set value. Through optimization algorithms in reinforcement learning, based on the objective function Solve for the optimal next time step setpoint within the current time window. ; Obtain optimized settings Then, maintain the current process characteristics. Unchanged, and splice as new input The data is fed into a trained LSTM model for forward inference, and the model outputs the corresponding quality prediction. The predicted health score is recalculated using the health score mapping formula. Forming a triplet It is then encapsulated into a standard message format.

4. The production process management method for AI-based smart meter casings as described in claim 3, characterized in that: The process of reading a single message involves simultaneously sending the set value to the PLC and MES via RPA and completing the write operation. Within the set SLA time limit, a readback confirmation and execution effect evaluation are performed. Based on the effect evaluation, the execution status is summarized and recorded, indicating that the single message is read in the order of arrival, and the recommended set value is... Simultaneously, the data is sent to the PLC and MES. The writing is completed by simulating manual operation through RPA. The orchestrator triggers two controlled RPA tasks at the same time, including the MES-side RPA and the PLC-side RPA to log in to SCADA. The setpoints are written one by one according to the address table. The operation status and timestamp are recorded after each task is completed. Set up a service level agreement. If any task times out, fails to write, or if subsequent monitoring finds that the quality health does not improve as expected, immediately trigger a rollback plan, restore to the settings snapshot before the change, and enter the exception handling process to ensure production safety and continuity. After the setpoints are issued and the SLA readback confirmation is completed, the execution effect evaluation begins, including the actual quality health within the observation window. The predicted target was not achieved. If the set value is not met, the execution will be marked as "poor performance", and the actual set value and special sequence information issued this time will be packaged into a structured experience sample. If health statistics If the expected goal is achieved or exceeded, it is marked as "Execution Completed". A complete status package of this execution is summarized and written to the experience queue, in the format: {Work Order ID, Timestamp, Actual Issuance Setting u, Readback Status, Health Statistics in the Observation Window}. , Execution conclusion.

5. The production process management method for AI-based smart meter casings as described in claim 4, characterized in that: The step of constructing an instant reward signal based on the summary record, and calculating the expected cumulative reward value based on the instant reward signal using the Q-learning algorithm refers to initializing the Q-value table for each state and corresponding action in the production process, with all Q values ​​initially set to zero. Construct an instant reward signal, awarding positive rewards for positive deviations and negative rewards for negative deviations. Using the Q-learning algorithm based on immediate rewards Get updated .

6. The production process management method for AI-based smart meter casings as described in claim 5, characterized in that: The process involves outputting an enhanced feature sequence from a pre-trained SVM model, automatically adjusting and executing the setpoint by combining the expected cumulative reward value and the reward signal, and generating a quality report based on the execution result. This report receives the target setpoint and real-time collected production data, converts them into feature vectors, inputs them into the trained SVM model, outputs the potential root causes of the current quality deviation, and encodes the SVM diagnostic results as additional feature dimensions to form an enhanced feature sequence. Given a master control policy based on Q-learning, Q-learning determines the policy based on the current state. renew Combined with reward signals Employ strategies to select from the action space Biggest movement The process is executed, and the results are continuously monitored. If the expected outcome is not achieved, Q-learning triggers a new round of exploration or regression, incorporating the explanations provided by SVM and the actions performed by Q-learning. The actual changes in health status are integrated to generate a structured quality report.

7. The production process management method for AI-based smart meter casings as described in claim 6, characterized in that: The process involves calculating production efficiency and quality efficiency values ​​based on the quality report, obtaining a comprehensive evaluation score, and then automatically determining and adjusting production parameters and health status based on the comprehensive evaluation score for the current operating conditions. Using real-time feedback data and historical production data, production efficiency is calculated to obtain a production efficiency value. Calculate the mass efficiency value Construct a multi-objective comprehensive evaluation score If the current comprehensive evaluation score If the value is greater than or equal to threshold A, it is considered a "high-quality steady state". The current parameters are confirmed to be valid and continuously monitored. If the value is less than threshold A, then the collaborative optimization mechanism is triggered, and the current value of Y is adjusted accordingly. The parameters and bottleneck list information are packaged into experience samples for updating the Q-learning strategy.

8. A production process management system for AI-based smart meter casings, based on the production process management method for AI-based smart meter casings according to any one of claims 1 to 7, characterized in that: include, The data acquisition and online quality estimation module is used to realize online quality index estimation and quality health stream generation based on real-time process data through multi-sensor fusion and ANN soft sensor model; The Quality and Health Prediction and Setpoint Optimization Module is used to predict health using LSTM and combine gradient and reinforcement learning algorithms to dynamically solve and output the optimal equipment setpoints that satisfy process constraints. The setpoint issuance and execution verification module is used to synchronously issue setpoints to PLC and MES via RPA, complete readback confirmation, effect evaluation and abnormal rollback within the SLA time limit, and record the structured execution status. The reinforcement learning reward modeling and strategy update module is used to construct immediate reward signals based on performance, and adopts Q-learning to continuously optimize the set value adjustment strategy to maximize long-term quality health benefits. Root Cause Analysis and Enhanced Decision Module: Utilizes pre-trained SVM to diagnose the root causes of quality deviations, generates enhanced features to assist Q-learning in making interpretable setpoint adjustments, and outputs a structured quality report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the production process management method for the smart meter casing based on AI as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the production process management method for the smart meter casing based on AI as described in any one of claims 1 to 7.

Images