Self-optimization method and system for quality system based on PDCA cycle

Abstract

This invention belongs to the field of quality management and discloses a self-optimization method and system for a quality system based on the PDCA cycle. The method includes: collecting multi-source heterogeneous quality data to construct a quality knowledge graph containing six types of nodes: personnel, equipment, materials, methods, environment, and measurement data; using a causal inference algorithm to construct a causal graph and calculate the average causal effect of each causal variable on quality indicators to locate the root cause variable of quality fluctuations; using the root cause variable, system parameters, and resource status as the state space, and making decisions through a reinforcement learning agent; setting up a causal effect prediction network, calculating causal guidance rewards based on the network output, and updating the agent by weighting and combining the reward with the actual reward from environmental feedback; and deploying the optimal solution to the production environment after verifying its safety in shadow execution mode. This achieves a leap from fine-tuning of execution parameters to reshaping management logic in the quality system, possessing a global perspective and deep self-evolution capabilities.

Description

Technical Field

[0001] This invention relates to the field of quality management, and more specifically to a self-optimization method and system for quality systems based on the PDCA cycle. Background Technology

[0002] As the cornerstone of Total Quality Management (TQM) theory, the PDCA (Plan-Do-Check-Act) cycle provides a rigorous logical framework for organizational process optimization through a cyclical, spiraling upward process. With the deep integration of the Industrial Internet and big data technologies, how to transform this classic closed-loop management logic from a traditional manually driven model into a highly adaptive and self-evolving intelligent management system has become a cutting-edge research topic in the field of quality engineering.

[0003] Chinese patent application CN118411052A discloses a data governance method based on the PDCA cycle. This method constructs an initial governance strategy and dynamically evaluates the execution results in real time, achieving continuous optimization of data quality to a certain extent. Specifically, this solution adjusts the strategy parameters of data governance through a closed-loop feedback mechanism, significantly improving the consistency and accuracy of data assets.

[0004] The invention patent with authorization announcement number CN113935084B discloses an intelligent construction system based on a BIM model. Its core logic lies in using Building Information Modeling (BIM) as a digital foundation, automatically collecting construction quality data through sensors and transmitting it back to the model in real time, thereby correcting the construction plan. This method successfully achieves a local closed loop from monitoring to correction in highly standardized engineering scenarios, providing strong support for precise control of construction quality.

[0005] Existing technologies exhibit clear characteristics of "verticalization" and "fragmentation." For example, data governance solutions focus solely on the single dimension of data quality, while engineering construction solutions are deeply coupled with pre-defined BIM physical models. This results in these solutions lacking a holistic review of the organization's overall quality architecture (involving complex unstructured elements such as cross-departmental collaboration, dynamic resource scheduling, and the evolution of standards and specifications) at the underlying principle level. When faced with optimizations of general management logic that is neither purely engineering-related nor purely data-driven, existing solutions often exhibit insufficient compatibility due to the lack of a universal technological foundation.

[0006] Existing technologies correct execution parameters for specific tasks by monitoring deviations, but this only addresses the superficial imbalances and does not address the inherent self-evolution of the quality management system. In complex business scenarios, quality bottlenecks often stem from logical redundancy in management process nodes, misallocation of resource allocation weights, or a disconnect between quality standards and market demands. Existing PDCA implementations based on specific models or single strategies are limited to optimization paths within pre-defined algorithmic boundaries, failing to autonomously identify and reconstruct the underlying management logic and decision-making models through deep correlation analysis of multi-dimensional quality indicators. Summary of the Invention

[0007] The purpose of this invention is to provide a self-optimization method and system for quality systems based on the PDCA cycle, which solves the technical problems of existing quality management solutions, such as strong domain limitations, inability to achieve deep optimization at the global logic layer, and lack of real-time self-evolution capabilities.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A quality system self-optimization method and system based on the PDCA cycle includes the following steps:

[0010] Step 1: Collect multi-source heterogeneous quality data and construct a quality knowledge graph containing quality element nodes and their logical association edges. The quality element nodes include personnel nodes, equipment nodes, material nodes, method nodes, environment nodes, and measurement data nodes.

[0011] Step 2: Based on the quality knowledge graph, a causal inference algorithm is used to construct a causal graph and calculate the average causal effect of each causal variable on the quality index to locate the root cause variable of quality fluctuation;

[0012] Step 3: Using the real-time monitoring values ​​of the root dependent variable, system parameters, and resource status as the state space, and the disposal schemes in the policy library as the action space, a reinforcement learning agent makes decisions. The reinforcement learning agent is a neural network model built on a deep deterministic policy gradient framework, including a policy network and a value network. The policy network takes the state space vector as input and outputs continuous control parameters in the action space. The value network takes the state-action pair as input and outputs an estimate of the expected cumulative reward.

[0013] Step 4: Set up a causal effect prediction network, take the current state and candidate actions as input, output the predicted causal effect vector of the quality index, and calculate the causal guidance reward based on the inverse of the predicted causal effect vector.

[0014] Step 5: The actual reward from environmental feedback is weighted and combined with the causal guided reward to obtain a total reward, which is used to update the value network and policy network of the reinforcement learning agent; wherein, the update gradient of the policy network includes a causal guided term, which is determined according to the gradient of the predicted causal effect vector with respect to the action; the value network updates parameters by minimizing the temporal difference error, and the policy network updates parameters by policy gradient ascent, and during the update, historical transfer samples are sampled from the experience replay pool for offline training;

[0015] Step 6: After verifying the security of the optimal solution output by the reinforcement learning agent in shadow execution mode, deploy it to the production environment for execution.

[0016] Furthermore, the collection of multi-source heterogeneous quality data in step 1 includes:

[0017] Access to the operating status parameters of production equipment, process log files of business management system, and structured reports of quality inspection process is achieved through a distributed data acquisition gateway.

[0018] For unstructured document quality standards and regulations, a natural language processing engine is used for feature extraction. The natural language processing engine uses a bidirectional long short-term memory network combined with an attention mechanism to transform the input text into structured knowledge containing entities and their semantic attributes.

[0019] The extracted entities are used as nodes in the quality knowledge graph, and the relationships between entities are used as directed edges. Each node is associated with a node feature vector, which is formed by concatenating a semantic feature vector and a structured attribute vector.

[0020] Furthermore, the quality-related entities extracted by the natural language processing engine include personnel, equipment, materials, methods, environment, and measurement data entity types, and the extracted relationship types include logical associations of "specification", "requirement", "responsibility", and "cause".

[0021] The semantic feature vectors are obtained by encoding the contextual semantics of entities through a bidirectional long short-term memory network, with the dimension uniformly set to 512.

[0022] If a node has no corresponding text description, the semantic feature vector is filled with all zeros or a default vector is generated by a pre-trained model.

[0023] The structured attribute vector is obtained by normalizing or embedding the structured attributes of the nodes, and its dimension is preset to a fixed value; the structured attributes are stored in the nodes in the form of key-value pairs for easy independent querying.

[0024] Furthermore, the causal inference algorithm in step 2 includes:

[0025] A fast causal inference algorithm or PC algorithm is used to learn and construct a causal graph containing all relevant variables from historical data. The causal graph is a directed acyclic graph, where nodes are quality elements and directed edges represent causal relationships.

[0026] For suspected causal variables, the do-calculus method is used to calculate the intervention distribution, and the minimum set of adjustment variables is identified using the backdoor or frontdoor criterion. The average causal effect of each causal variable on the quality index is then calculated.

[0027]

[0028] in, Indicates the average causal effect; As a causal variable; For quality indicators; Indicates an abnormal state; Indicates the standard state; This indicates an intervention operation, which is about to begin. Forced to be set as And cut off other variables The impact; Indicates the expected value;

[0029] When calculating the average causal effect, the backdoor criterion or the frontdoor criterion is used to estimate the causal effect from the observed data, quantify the contribution rate of each causal variable to the quality fluctuation, and determine the causal variable with the highest contribution rate as the root cause variable.

[0030] Furthermore, in step 4, the causal effect prediction network is a three-layer fully connected neural network structure, with the input layer dimension being the sum of the state dimension and the action dimension, the number of hidden layer neurons being 256 and 128 respectively, the activation function being ReLU, and the output layer dimension being the number of quality indicators.

[0031] The training process of the causal effect prediction network includes: using historical intervention datasets Network parameters are trained by minimizing the mean squared error loss function. :

[0032]

[0033] in, The loss function; The parameters of the causal effect prediction network are: [Parameters of the network]; [Historical intervention dataset] For the first The state of each sample; For the first The actions of each sample; For the first A vector of changes in quality indicators corresponding to each sample; For predicting causal effects, the network pairs The predicted value; The Euclidean norm of a vector; Indicates the number of samples in the historical intervention dataset;

[0034] The historical intervention data is obtained through the following methods: extracting naturally occurring intervention events from the historical logs of the production system, generating simulated intervention data in a digital twin simulation environment, or collecting real intervention data through a security exploration strategy in the early stages of system deployment.

[0035] Furthermore, the formula for calculating the total reward in step 5 is as follows:

[0036]

[0037] in, Total reward; A tangible reward for environmental feedback; This is the causal guidance coefficient, used to balance the weights of actual rewards and causal guidance rewards; The reward is causally guided; the actual reward. The increment of the quality objective function is defined as:

[0038] in, To enhance the number of sequential iterations in learning interaction The corresponding comprehensive quality objective function value is based on the number of time-series iteration steps. The calculation is based on real-time monitoring data within a preset time window that is initially slid forward. This is the comprehensive quality objective function value corresponding to the previous interaction time-series iteration step number;

[0039] The causal guided reward The output calculation based on the causal effect prediction network is defined as:

[0040]

[0041] in, The quantity of quality indicators; For the first The preset weights of each quality indicator satisfy ; The first prediction of the network output for causal effects Each quality indicator in the number of time-series iteration steps The predicted change.

[0042] Furthermore, in step 5, the reinforcement learning agent adopts a deep deterministic policy gradient framework, which includes a policy network and a value network.

[0043] The value network Used to assess the state Take action below The expected cumulative reward is output as a scalar value; the policy gradient calculation formula for the policy network is:

[0044] in, For policy gradient; For policy network parameters; This refers to the number of samples in a small batch. Output for value networks; For the policy network in the 1st Individual sample status The action to be output; This indicates that the value network output corresponds to the action. The gradient; The policy network output is related to the parameters. The gradient; To guide the exploration of causal factors; Predict network output for causal effects; The square of the output modulus of the causal effect prediction network is represented by the action. The gradient;

[0045] The value network is updated by minimizing the temporal difference error.

[0046] Furthermore, the shadow execution mode in step 6 specifically refers to:

[0047] The optimized management strategy is run in a simulation environment built on digital twin technology. The digital twin environment simulates the key dynamic responses of the real production line with a prediction error of less than a preset percentage of the actual production fluctuations by fitting a physical engine with historical big data.

[0048] When the strategy to be optimized runs continuously for a preset number of cycles in the simulation environment and passes the convergence verification, and all risk indicators are within the safe range, the optimization strategy is switched to the real production environment through blue-green deployment technology; the risk indicators include equipment load peak and cycle time delay.

[0049] Furthermore, the present invention also includes a strategy distillation step:

[0050] A teacher-student model architecture is used to distill the high-dimensional policy network generated by reinforcement learning into a lightweight decision model. The teacher model is a trained high-dimensional reinforcement learning neural network, and the student model is a gradient boosting decision tree model or a decision tree model.

[0051] During the distillation process, the action probability distribution output by the teacher model is used as a soft label, and the student model takes the state as input and the soft label as the learning target, and is trained by minimizing the cross-entropy loss.

[0052] The trained student models are deployed on edge computing nodes to execute quality control decisions in real time.

[0053] Furthermore, after step 1 and before step 2, a quality planning logic generation step is also included:

[0054] Define a quality objective function, use a genetic algorithm to encode and optimize the process node configuration scheme, and generate the optimal process topology structure that maximizes the quality objective function;

[0055] The quality objective function is defined as follows:

[0056]

[0057] in, This represents the quality objective function value, used to quantify the merits of process configuration schemes; , The weighting coefficients, denoted by 1 and 2, are preset based on business scenarios, representing the relative importance of compliance, efficiency, and stability, respectively, and satisfying the following conditions: ; Represents the number of time series iterations Process compliance score; Represents the number of time series iterations Resource utilization efficiency; Represents the number of time series iterations Quality output stability; Process compliance score Defined as:

[0058]

[0059] in, For the number of time-series iteration steps The number of violations that occurred during the monitoring period; The total number of inspection items during the monitoring period; the resource utilization efficiency. Defined as:

[0060] in, To monitor the actual output during the monitoring period; The manpower and hours invested during this period; The equipment uptime within this period; the quality output stability Defined as:

[0061]

[0062] in, The variance of key quality characteristic values ​​within the monitoring period; This represents the average value of key quality characteristics during the same period. For the natural constant An exponential function with base 0; the calculation results of all the above indicators are naturally in a state of... The intervals can be directly summed using weighted averages. Furthermore, in the genetic algorithm, each possible set of process node configuration schemes is encoded as a chromosome, which consists of two parts: node parameter genes and topology structure genes.

[0063] The node parameter gene encodes key parameters for each process node, including sampling frequency, control threshold, and resource allocation weight. The topology gene uses a binary encoding of an adjacency matrix to represent the connection relationships and directions between nodes, where the matrix elements... Represents a node There is a directed connection to the node. This indicates no connection. After each genetic operation, a cycle detection is performed on each newly generated chromosome. A topological sort or depth-first search is used to determine if a cycle exists in the directed graph. If a cycle exists, one edge in the cycle is randomly deleted, i.e., the corresponding adjacency matrix element is set to... Repeat this process until the graph is a directed acyclic graph;

[0064] The fitness function is the quality objective function. The value, after a preset number of iterations, generates a process topology diagram including the distribution of key control points, sampling strategies, and resource scheduling sequences. Furthermore, this invention also includes a multi-model fusion anomaly detection step:

[0065] The system calculates the residual vector between real-time monitoring data and standard benchmark values. When an element in the residual vector exceeds a preset confidence interval threshold, the system directly determines it as a quality deviation.

[0066] Meanwhile, an anomaly detection model based on residual neural networks, a classifier based on support vector machines, and an anomaly detector based on isolated forests are integrated. The detection results of the three algorithms are used to obtain the final anomaly judgment through a weighted voting mechanism. The weights of the weighted voting are dynamically adjusted according to the F1 score of each model or fixed empirical weights are adopted.

[0067] Furthermore, it also includes blockchain quality traceability steps:

[0068] The core decision parameters, monitoring data summaries, and optimization instructions of each stage in the quality self-optimization process are used to generate a storage snapshot through a hash algorithm, wherein the hash algorithm is the SHA-256 hash algorithm.

[0069] The evidence snapshot is encapsulated in a block with timestamp information and synchronized to each peer node of the consortium blockchain through a consensus algorithm to ensure the immutability and auditability of the quality self-optimization trajectory.

[0070] In addition, this invention also discloses a quality system self-optimization system based on the PDCA cycle, used in the PDCA cycle-based quality system self-optimization method described above, comprising:

[0071] The all-element perception module is configured to collect multi-source heterogeneous quality data and construct a quality knowledge graph containing quality element nodes and their logical association edges. The quality element nodes include personnel nodes, equipment nodes, material nodes, method nodes, environment nodes, and measurement data nodes.

[0072] The causal diagnosis module is configured to construct a causal graph based on the quality knowledge graph and use a causal inference algorithm to calculate the average causal effect of each causal variable on the quality index, thereby locating the root cause variable of quality fluctuation; the causal effect prediction network adopts a three-layer fully connected neural network structure, takes the current state and candidate actions as input, and outputs a predicted causal effect vector for the quality index.

[0073] The reinforcement learning agent is configured to use the root cause variable, system parameters and resource status as the state space and the disposal schemes in the policy library as the action space for decision-making.

[0074] The reinforcement learning agent is a complete intelligent decision-making system built on a deep deterministic policy gradient framework, specifically including:

[0075] The policy network module is used to output the optimal action vector based on the current state;

[0076] The value network module is used to evaluate the expected cumulative reward of state-action pairs;

[0077] The experience replay pool module is used to store historical transfer samples for offline training;

[0078] The target network module is used for value estimation during the stable training process.

[0079] The above modules work together to enable the intelligent agent to have three core functions: real-time decision-making, parameter updates, and continuous optimization.

[0080] The total reward function of the reinforcement learning agent is a weighted combination of the actual reward from environmental feedback and the causal guided reward, wherein the causal guided reward is calculated based on the output of the causal effect prediction network.

[0081] The shadow execution module is configured to verify the security of optimization strategies in a simulation environment built on digital twin technology, and deploy the optimization strategies to the production environment after successful verification.

[0082] Furthermore, the all-element perception module includes: a distributed data acquisition gateway, the physical interface layer of which includes an Ethernet interface conforming to the IEEE 802.3 standard, a CAN bus interface conforming to the ISO 11898 standard, and a communication interface conforming to the RS-485 standard, for accessing the operating status parameters of production equipment, process log files of the business management system, and structured reports of the quality inspection process;

[0083] The natural language processing engine uses a bidirectional long short-term memory network combined with an attention mechanism to extract features from unstructured document-type quality standards and regulations, transforming the input text into structured knowledge containing entities and their semantic attributes.

[0084] The edge sensing sensor network, including pressure sensors, photoelectric encoders, temperature and humidity sensors, and industrial cameras, is connected to the edge computing node via an isolated industrial bus.

[0085] Furthermore, in the causal diagnosis module, the formula for calculating the average causal effect is:

[0086] in, Indicates the average causal effect; As a causal variable; For quality indicators; indicating abnormal conditions; Indicates the standard state; This indicates an intervention action, which will force the setting to... And cut off other variables The impact; This represents the expected value.

[0087] Furthermore, the reinforcement learning agent employs a deep deterministic policy gradient framework, which includes a policy network and a value network.

[0088] The formula for calculating the policy gradient of the policy network is as follows:

[0089] in, For policy gradient; For policy network parameters; This refers to the number of samples in a small batch. Output for value networks; For the policy network in the 1st Individual sample status The action to be output; This indicates that the value network output corresponds to the action. The gradient; The policy network output is related to the parameters. The gradient; To guide the exploration of causal factors; Predict network output for causal effects; The square of the output modulus of the causal effect prediction network is represented by the action. The gradient guides the policy to update in the direction of improving prediction quality.

[0090] Furthermore, it also includes:

[0091] The policy distillation unit uses a teacher-student model architecture to distill the high-dimensional policy network generated by reinforcement learning into a lightweight decision model. The student model is a gradient boosting decision tree model or a decision tree model, which is deployed on edge computing nodes. The logic planning module is configured to define a quality objective function and uses a genetic algorithm to encode and optimize the configuration scheme of process nodes to generate the optimal process topology structure that maximizes the quality objective function.

[0092] Furthermore, it also includes:

[0093] The real-time computing engine uses sliding window technology to perform real-time statistical analysis on key quality characteristic values ​​and calculates the mean, variance, and skewness statistics.

[0094] The blockchain quality traceability module is configured to generate a snapshot of the core decision parameters, monitoring data summaries, and optimization instructions for each stage of the quality self-optimization process using the SHA-256 hash algorithm, encapsulate it into a block with timestamp information, and synchronize it to each peer node of the consortium blockchain through a consensus algorithm.

[0095] Furthermore, in step 6, when the deployment is executed in the production environment, the optimization strategy is received and executed through the intelligent execution terminal. The intelligent execution terminal is equipped with a microprocessor with a main frequency of not less than 2.0 GHz and a running memory of not less than 8 GB. The quality control logic is encapsulated into a microservice image through containerization technology to achieve rapid deployment.

[0096] Compared with the prior art, the present invention has the following beneficial effects:

[0097] This invention decouples quality management from fragmented and vertical business scenarios by constructing a quality knowledge graph covering six elements: personnel, equipment, materials, methods, environment, and measurement data. It establishes a unified modeling architecture that can be used across multiple domains. This architecture comprehensively depicts the logical connections and causal constraints between quality elements using nodes and directed edges, providing a computable digital foundation for global optimization and breaking through the compatibility bottlenecks caused by domain limitations in existing technologies.

[0098] This invention introduces a structural causal model and a quantitative calculation method for average causal effects, achieving a leap from "correlation identification" to "causal inference." By constructing a causal graph containing latent variables and using the backdoor criterion to estimate the intervention distribution unbiasedly from the observed data, the system can accurately locate the root cause variables of quality fluctuations and quantify the contribution rate of each factor, overcoming the inherent limitation of existing technologies that can only detect superficial deviations and cannot reach the underlying causes.

[0099] This invention deeply integrates causal inference results into a reinforcement learning decision-making framework. It generates causal-guided rewards through a causal effect prediction network and explicitly adds causal gradient terms to the policy network's gradient update, forming a tightly coupled closed loop of "diagnosis—prediction—decision—verification." This mechanism enables agents to explore efficiently based on physical mechanisms, significantly reducing trial-and-error costs and achieving a leap from "fine-tuning execution parameters" to "reshaping management logic," endowing the quality system with true deep self-evolutionary capabilities.

[0100] This invention sets up a shadow execution mode and a strategy distillation unit. Through pre-verification in a digital twin environment and deployment of a lightweight model, it sinks complex optimization decision-making capabilities to edge computing nodes while ensuring system security and real-time performance, forming a complete implementation path that balances intelligence depth and engineering reliability. Attached Figure Description

[0101] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0102] Figure 1 This is an overall flowchart of the method described in this invention.

[0103] Figure 2 This is a simplified flowchart of the overall process of this invention.

[0104] Figure 3 This invention describes the process for multi-source data acquisition and quality knowledge graph construction.

[0105] Figure 4This invention provides a process for causal diagnosis and intelligent decision-making deployment.

[0106] Figure 5 This is a diagram of the system operation interface described in this invention. Detailed Implementation

[0107] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0108] The following is in conjunction with the appendix Figures 1-5 The embodiments of the present invention will be described in detail below.

[0109] This invention discloses a self-optimization method for a quality system based on the PDCA cycle. The underlying logic is built on the industrial internet architecture, and a global quality closed loop is achieved through an integrated quality management cloud platform and distributed edge sensing terminals.

[0110] In practical implementation, the system connects to the entire business chain of enterprise production and management through a distributed data acquisition gateway. The data acquisition gateway integrates multiple industry standard interfaces at the physical layer, including but not limited to a Gigabit Ethernet interface conforming to the IEEE 802.3 standard, a CAN high-speed bus interface conforming to the ISO 11898 standard, and an RS-485 interface with surge protection.

[0111] In actual deployment, the data acquisition gateway runs a real-time task operating system, responsible for synchronously sampling and preprocessing signals from the edge sensing sensor network. The edge sensing sensor network includes pressure sensors deployed at key workstations, high-precision photoelectric encoders with a measurement accuracy of 0.01mm, and integrated temperature and humidity sensors for environmental parameter monitoring.

[0112] The data acquisition gateway utilizes a pre-defined message parsing plugin to unpack and normalize raw data streams of different protocol formats. For PLC control data of production equipment, the system extracts key performance indicators such as operating speed, load rate, and real-time energy consumption of the equipment through OPC UA or Modbus TCP protocols. For process log files in business management systems, such as production scheduling documents in ERP systems or process reports in MES systems, the system performs timed polling through database middleware or achieves second-level retrieval through message queues (such as Kafka).

[0113] For the widespread use of unstructured scanned paper documents or PDF quality manuals within enterprises, a deep learning-based natural language processing engine has been integrated. This engine employs a bidirectional long short-term memory network as a feature encoder and introduces a multi-head attention mechanism to capture long-distance semantic relationships in the text.

[0114] In practice, the natural language processing engine first performs fine-grained word segmentation on the text, and then maps the words into low-dimensional dense vectors through the embedding layer. The number of hidden units in the bidirectional long short-term memory network is set to 256, and a contextual representation containing contextual information is generated by concatenating the features from the forward and backward directions. The attention mechanism layer automatically focuses on entity words with quality constraint attributes, such as "tolerance range," "inspection frequency," and "control strategy," based on the weight matrix. Finally, all unstructured information is transformed into feature tensors in a high-dimensional vector space, with its dimension uniformly set to 512. This standardized feature representation provides a unified mathematical foundation for subsequent cross-source data fusion. Each quality element node is associated with an attribute vector, which is composed of two parts: the first part is a 512-dimensional semantic feature vector obtained by the natural language processing engine (if the node has no corresponding text description, this part of the vector is filled with all zeros or a default vector is generated by a pre-trained model); the second part is a low-dimensional vector obtained by normalizing or embedding structured attributes (such as personnel skill level, equipment mean time between failures, material chemical composition deviation, etc.), the dimension of which is related to the number of specific attributes.

[0115] To ensure dimensionality consistency across all node feature vectors, the system predefines a set of standard structured attribute types (e.g., N attributes such as personnel skill level, equipment mean time between failures, and material chemical composition deviation). For each node, its corresponding attribute value is extracted, and if an attribute is missing, it is filled with 0. Alternatively, an embedding layer is used to map the structured attribute set of each node into a fixed-dimensional (e.g., 64-dimensional) vector. This fixed-dimensional vector is concatenated with a 512-dimensional semantic feature vector to form a node feature representation with uniform dimensionality (e.g., 576-dimensional), which is then input into the subsequent graph neural network. Simultaneously, structured attributes are also stored in nodes as key-value pairs for easy independent lookup.

[0116] After completing the full-element perception, this invention enters the quality planning logic generation stage based on a quality knowledge graph, corresponding to the planning stage in the PDCA cycle. With the support of a cloud computing cluster, the logic planning module constructs a dynamically evolving, multi-level quality knowledge graph.

[0117] The quality knowledge graph uses a graph database as its underlying storage engine. Nodes are defined as quality elements, and edges are defined as logical constraints or causal relationships between elements. Node types cover personnel (job skill level, historical pass rate), equipment (mean time between failures, positioning accuracy, wear compensation value), materials (batch number, chemical composition deviation), methods (process flow, detection algorithm parameters), environment (cleanliness level, environmental vibration frequency), and measurement data (mean, standard deviation). Edge attributes include association strength coefficient and time delay parameters.

[0118] To uncover potential quality risks, the system employs a link prediction algorithm based on graph convolutional neural networks. The graph convolutional neural network model calculates the potential connection probability between nodes by aggregating the first-order and higher-order neighbor features of the target node.

[0119] In a specific example, if the model identifies a strong implicit correlation between supplier changes at the material node and fluctuations in the hardness index at the finished product node, the system will automatically establish a causal edge with high weight in the quality knowledge graph.

[0120] Furthermore, by calculating the quality objective function To generate the optimal process topology. The quality objective function... Defined as:

[0121] in:

[0122] This represents the quality objective function value, used to quantify the merits of process configuration schemes;

[0123] , , To represent the relative importance of compliance, efficiency, and stability based on pre-defined weighting coefficients according to business scenarios, and to satisfy... ;

[0124] Represents the number of time series iterations Process compliance score;

[0125] Represents the number of time series iterations Resource utilization efficiency;

[0126] Represents the number of time series iterations Stability of quality output;

[0127] This indicates the number of timing iteration steps.

[0128] The specific calculation methods for each indicator are as follows:

[0129] Process compliance score The degree of matching between real-time monitoring data and preset standards (such as ISO 9001 clauses or internal company specifications) is defined as:

[0130]

[0131] in: The number of violations that occurred during the monitoring period;

[0132] This represents the total number of inspection items during the monitoring period.

[0133] Resource utilization efficiency Defined as the ratio of effective output to resource consumption, i.e.:

[0134]

[0135] in:

[0136] To monitor the actual output during the monitoring period;

[0137] The manpower and hours invested during this period;

[0138] This refers to the equipment's operating time within this cycle.

[0139] Quality output stability Calculated using statistical indicators of key quality characteristics, using:

[0140] in:

[0141] The variance of key quality characteristic values ​​within the monitoring period; This represents the average value of key quality characteristics during the same period. For the natural constant An exponential function with base 0.5, which ensures Values ​​in Within the range, the smaller the fluctuation (smaller the variance), the better. The closer it is to 1.

[0142] The calculation results of the above indicators are all naturally in a state of... For intervals, a weighted sum can be directly calculated. The system uses a genetic algorithm to search within the solution space to make... The goal is to maximize the configuration of process nodes to complete the planning process within the PDCA cycle. Each possible configuration is encoded as a chromosome. The chromosome consists of two gene sequences: the first part is the node parameter gene, where key parameters for each process node (such as procedures or inspection points) are encoded as real numbers or binary strings, such as sampling frequency. (Value range 1-100 Hz), control threshold (Value range is 0-5% relative to the standard value), resource allocation weight (Value range 0-1), etc.; the second part is the topological structure gene, which uses the binary encoding of the adjacency matrix to represent the connection relationship and direction between nodes, where the matrix elements Represents a node arrive There is a directed connection.

[0143] This indicates no connections and ensures no cycles. Specifically, after each genetic operation (crossover, mutation), the system performs cycle detection on each newly generated chromosome: it determines whether a cycle exists in the directed graph by topological sorting or depth-first search. If a cycle exists, it randomly deletes one edge in the cycle (i.e., sets the corresponding adjacency matrix element to 0), and repeats this process until the graph is a directed acyclic graph.

[0144] In genetic algorithms, the initial population is randomly generated, and the quality objective function of the process configuration scheme corresponding to each chromosome is calculated based on historical data or simulation environment. The fitness value is used as the basis for selection, crossover (exchanging partial gene fragments), and mutation (randomly changing gene values). After more than 500 iterations, the algorithm converges to the optimal solution, generating a process topology diagram including the distribution of key control points, sampling strategies, and resource scheduling sequences. The logic planning module converts the optimized process topology into a structured configuration file. This configuration file, in JSON format, defines the task queues, control parameter thresholds, and triggering logic between each intelligent execution terminal. The dynamic execution module parses this configuration file and sends task instructions to the corresponding intelligent execution terminals via API interfaces.

[0145] During the dynamic scheduling and real-time monitoring phase of process execution—the execution stage of the PDCA cycle—the dynamic execution module is responsible for transforming the planned logic into business instructions in real time. These instructions are then distributed via a 10-gigabit fiber optic network to edge computing nodes and intelligent execution terminals deployed on-site. Each intelligent execution terminal employs an embedded system with industrial-grade rugged design, featuring a multi-core microprocessor with a clock speed of 2.0 GHz or higher and 8 GB of DDR4 RAM.

[0146] The system employs Docker containerization technology, encapsulating the quality control logic into independent microservice images, enabling minute-level deployment and hot updates of the control algorithm. During execution, the real-time computing engine processes the collected business flow data online. This real-time computing engine utilizes sliding window technology, with the window size set to... Step size set to Within each sliding window, the system performs real-time statistical inference on key quality characteristic values, calculating their moving average, sliding variance, and skewness and kurtosis. When the system detects that the sliding variance of a certain indicator shows a monotonically increasing trend for three consecutive windows, it will trigger an early warning interruption and push correction parameters to the workstation terminal in real time to ensure that the execution process does not deviate from the preset quality baseline.

[0147] Entering the multidimensional deviation analysis and quality system assessment stage, which is the inspection phase in the PDCA cycle, the deviation analysis module adopts an anomaly detection strategy based on multi-model fusion.

[0148] First, calculate the residual vector between the real-time monitoring data and the standard benchmark value. ,in yes Dimensional real-time monitoring data vector, yes dimensional standard reference vector, yes The residual vector is 3D, where each component represents the absolute deviation between the actual value and the standard value of the corresponding monitoring indicator. When an element in the residual vector exceeds a preset confidence interval threshold, the system directly determines it as a quality deviation.

[0149] Meanwhile, the system integrates an anomaly detection model based on residual neural networks, a classifier based on support vector machines, and an anomaly detector based on isolated forests to address different data characteristics and operating conditions. The specific fusion strategy is as follows: when the real-time data volume is large and nonlinear features are significant, the residual neural network model is prioritized; when the data dimensionality is high and rapid outlier identification is required, isolated forests are used; and for complex classification tasks with small samples, support vector machines are used. The detection results of the three algorithms are weighted and voted to obtain the final anomaly judgment.

[0150] In practical implementation, weight settings can be dynamically adjusted: the system periodically evaluates the F1 score of each model using a labeled validation dataset, using the F1 score ratio as the weight; or fixed empirical weights can be used, such as 0.5 for residual neural networks, 0.3 for isolated forests, and 0.2 for support vector machines. Through weighted voting, the outputs of each model are combined to obtain the final anomaly detection result, thus maintaining high accuracy under different operating conditions.

[0151] To further identify the underlying causes of anomalies, this invention introduces a causal inference algorithm based on a structural causal model. Utilizing historical data from the all-factor perception module, a causal graph containing all relevant variables is constructed using a constraint learning algorithm (such as the PC algorithm). The PC algorithm steps are as follows: First, calculate the partial correlation coefficients between all pairs of variables and then perform a Fisher-Z test (significance level...). To determine conditional independence, an undirected graph skeleton is constructed. The direction of edges is determined using the V-structure (i.e., two variables share a common child node and have no direct edges). Finally, Meek rules (such as the direction propagation rule) are applied to further determine the direction of the remaining edges, resulting in a directed acyclic graph, i.e., a causal graph. For complex scenarios with potential latent variables, a fast causal inference algorithm can be used, which infers causal relationships by detecting potential confounding factors. Nodes in the graph represent quality elements, and directed edges represent causal relationships. For nodes suspected of being anomalous, do-calculus is used to calculate their intervention distribution. Compared with observed distribution The difference, or directly calculate each node For output Average causal effect:

[0152] in: This indicates the average causal effect.

[0153] The causal variables to be analyzed (such as equipment parameters, operational variables, etc.);

[0154] For outcome variables (such as quality indicators);

[0155] and They represent Two possible value states (e.g.) This is an abnormal state. (This is the normal state)

[0156] This indicates an intervention operation, which is about to begin. Forced to be set as And cut off other variables The impact;

[0157] This represents the expected value.

[0158] In practical calculations, the backdoor criterion or instrumental variable method is used to estimate causal effects from observed data, thereby quantifying the contribution rate of each process node to quality fluctuations. By constructing a causal graph containing hundreds of variables, the system can eliminate superficial spurious correlations. For example, when analyzing coating defects in automotive parts, the system can eliminate the surface influence of environmental humidity fluctuations and accurately pinpoint the torque fluctuation of the third-axis servo motor of the painting robot as the root cause of uneven paint film thickness. This quantitative root cause diagnosis provides precise physical evidence for subsequent decision-making, enabling the inspection phase to move beyond simply "finding problems" to "interpreting problems."

[0159] During the autonomous correction and evolution phase of the system logic, i.e. the handling stage in the PDCA cycle, the logic evolution module implements a deep self-optimization strategy. The system has a built-in initial strategy library, which contains successful handling solutions extracted from historical quality events (stored in the form of rules or parameters), as well as standard handling procedures preset by domain experts.

[0160] Based on the root cause analysis results from the inspection phase, the system retrieves the optimal treatment plan from the policy library using a reinforcement learning algorithm. The state space construction process of the reinforcement learning agent is as follows: First, the root cause analysis results output by the deviation analysis module (e.g., "micro-crack in the press-fit head") are mapped into a quantifiable feature vector. This vector includes the root cause type encoding, real-time sensor feature values ​​related to the root cause (e.g., energy values ​​of specific frequency bands in the vibration spectrum), and the causal effect strength value of the root cause. This mapping process is implemented through a pre-defined "root cause-feature mapping table," which is predefined by domain experts and dynamically updated during system operation.

[0161] In one specific embodiment of the present invention, the reinforcement learning agent is implemented using a deep deterministic policy gradient framework, and its state space is defined as follows: A 3D real-valued vector, where each dimension corresponds to a state variable. Taking the assembly scenario of a precision reducer as an example, the state dimensions are... The dimensions are defined as follows:

[0162] The first dimension represents the tightening torque of bolt number 5. The unit is Nm, collected in real time via OPC UA node ns=3;s="Torque_05", with an original range of [45, 55], and normalized using a formula. Map to the interval [0,1].

[0163] The second dimension represents the robot's second-axis positioning repeatability. The unit is mm, collected by a laser displacement sensor and calculated using a 100-point moving average. The original range is [0.005, 0.02], and the normalization formula is... , mapped to [0,1].

[0164] The third dimension is the effective value of vibration acceleration. The unit is m / s², collected by a vibration sensor and calculated using a 50-point moving average. The original range is [0, 5]. The normalization formula is... , mapped to [0,1].

[0165] The fourth dimension is the urgency of the production cycle. Obtained from production scheduling data in the MES system, the original range is [0,1], and can be used directly.

[0166] The fifth dimension is the material hardness deviation. The unit is HRC, obtained from incoming material inspection data, with an original range of [-2, 2], and a normalization formula. , mapped to [0,1].

[0167] The 6th dimension is ambient temperature. The unit is °C, collected by a temperature and humidity sensor, with an initial range of [20, 30], and a normalization formula. , mapped to [0,1].

[0168] The 7th dimension is ambient humidity. The unit is %, collected by a temperature and humidity sensor, with an initial range of [30, 80], and a normalization formula. , mapped to [0,1].

[0169] The 8th dimension is the pass rate of the preceding process. The unit is %, obtained from the MES quality statistics module, with an original range of [90, 100] and a normalization formula. , mapped to [0,1].

[0170] The 9th dimension is the remaining lifespan of the equipment. The unit is %, obtained from the predictive maintenance module, with an original range of [0, 100], and a normalization formula. , mapped to [0,1].

[0171] Dimensions 10 to 12 are the material batch number code. The data is obtained from the batch management database and uses One-Hot encoding. The three batch types correspond to a three-dimensional vector with values ​​of {0,1}³.

[0172] The 13th dimension is the operator's proficiency. Skills were obtained from HR system skill profiles, ranging from level 1 to 5, using the normalization formula. , mapped to [0,1].

[0173] The 14th dimension is the tool wear amount. The unit is mm, collected by the tool in-situ detection sensor, with an initial range of [0, 0.3], and a normalization formula. , mapped to [0,1].

[0174] The 15th dimension is the historical fluctuation trend of quality. It is calculated using the formula (current CPK - target CPK) / target CPK, with an original range of [-1, 1], which can be used directly.

[0175] For discrete variables such as root cause types, a one-hot encoding method is used to embed the state vector, with the encoding dimension equal to the total number of root cause categories. The normalized state vector serves as the input to the policy network and value network of the reinforcement learning agent.

[0176] The action space of the reinforcement learning agent is defined as follows: A continuous vector, where each dimension corresponds to the relative adjustment ratio of an adjustable control parameter, with values ​​ranging from [value range missing]. The policy network output layer uses the Tanh activation function to restrict actions to this range. Taking the precision reducer assembly scenario as an example, the action dimension... The dimensions are defined as follows:

[0177] The first dimension corresponds to the torque adjustment ratio. , representing the relative rate of change of the target torque of bolt #5, is mapped using the following formula: It is executed by writing to register 40001 via Modbus.

[0178] The second dimension corresponds to the calibration frequency adjustment ratio. , representing the relative rate of change of robot calibration frequency, is mapped using the following formula: This is accomplished by updating the MES process parameter table.

[0179] The third dimension corresponds to the clamping force adjustment ratio. , representing the relative rate of change of the clamping force of the fixture, is mapped by the formula: It is executed through the PLC analog output module.

[0180] The fourth dimension corresponds to the heating time adjustment ratio. , representing the relative rate of change of heating time in the hot-loading process, is mapped using the following formula: It is executed by preset values ​​of the PLC timer.

[0181] The 5th dimension corresponds to the feed rate adjustment ratio. , representing the relative rate of change of the assembly robot's feed speed, is mapped using the following formula: It is executed via servo drive speed commands.

[0182] The 6th dimension corresponds to the air pressure adjustment ratio. , representing the relative rate of change of the working pressure of the pneumatic system, is mapped by the formula: It is executed by controlling the signal through a proportional valve.

[0183] The 7th dimension corresponds to the nozzle distance adjustment ratio. , representing the relative rate of change of the distance between the adhesive / spray nozzle and the workpiece, is mapped by the formula: It is executed through stepper motor position commands.

[0184] The 8th dimension corresponds to the motor current limiting adjustment ratio. , representing the relative rate of change of the maximum current limit of the drive motor, is mapped by the following formula: The execution is performed by writing parameters through the servo driver.

[0185] The scaling factor (e.g., 0.05, 0.2, etc.) in the mapping formula represents the upper limit of the safe adjustment range, which is preset according to the process specifications and can be modified online. Before execution, the action must be verified by the shadow execution module to ensure that the adjusted physical value does not exceed the equipment's allowable range.

[0186] The reinforcement learning algorithm employs a policy gradient-based optimization framework, whose reward function and quality objective function... The increment is positively correlated. The agent takes the current state (including root cause type, system parameters, resource state, etc.) as input, calculates the value of each candidate solution through the policy network, and selects the solution with the highest value to execute.

[0187] In this embodiment, the reinforcement learning agent adopts a deep deterministic policy gradient framework, which includes four deep neural networks: an online policy network. Target-Policy Network Online value network and target value network .

[0188] The specific structures of each network are as follows: (1) Online policy network The input layer dimension is state vector (In this embodiment) Hidden layer 1 is a fully connected layer with 400 neurons, using ReLU activation function, Xavier uniform weight distribution, and bias initialization to a constant of 0.1. Hidden layer 2 is a fully connected layer with 300 neurons, using ReLU activation function, Xavier uniform weight distribution, and bias initialization to a constant of 0.1. The output layer is a fully connected layer with [number missing]. (In this embodiment) The activation function is Tanh, which restricts the output to... The interval and weights are initialized using a uniform distribution. The bias is initialized to a constant of 0.1.

[0189] (2) Online value network : State vector in the state input path A 400-dimensional feature vector is obtained through a fully connected layer (400 neurons, ReLU activation). In the action input path, the action vector... A 400-dimensional feature vector is obtained through a fully connected layer (400 neurons, ReLU activation). The merging layer adds the 400-dimensional vectors output from the two paths element-wise, followed by ReLU activation. The output layer is a fully connected layer (1 neuron, no activation function), outputting a scalar value with uniformly distributed weights. The bias is initialized to a constant of 0.1.

[0190] (3) Target Policy Network and target value network The structure is exactly the same as that of the online policy network and the online value network, respectively. Its parameters are copied from the online network parameters through a soft update method, and the update formula is:

[0191]

[0192] in: This is the soft update coefficient, with a range of values. In this embodiment, .

[0193] (4) Training hyperparameter configuration: The optimizer used is the Adam optimizer, and the policy network learning rate is [missing information]. Value network learning rate The weight decay coefficients are all The experience replay pool capacity is... Transfer samples, sampling small batch size Discount factor This is used to calculate the discounted value of the cumulative reward. The exploration noise uses Ornstein-Uhlenbeck random process noise, with noise parameters... , To enhance exploration capabilities.

[0194] (5) Decision-making and update process:

[0195] ① Initialization: Randomly initialize the parameters of the policy network and value network, copy the target network parameters from the online network, and clear the experience replay pool.

[0196] ② Decision: In the number of time-series iteration steps Collect state vectors Input the action of the online strategy network During the training phase, Ornstein-Uhlenbeck noise was superimposed on the movements and then the movements were executed with amplitude limiting.

[0197] ③ Storage: Observe the new state after executing the action. and rewards Transfer samples Store in the experience replay pool.

[0198] ④ Update: Each execution Step (in this embodiment, take) Randomly sample a batch of samples from the playback pool and calculate the temporal difference objective:

[0199] By minimizing the mean square error loss function Update value network parameters Then update the policy network parameters according to the policy gradient formula. Finally, based on the soft update coefficient. Update the target network parameters.

[0200] ⑤ Iteration: Repeat steps ②-④ until the cumulative reward converges or the preset number of iterations is reached.

[0201] If logical redundancy is detected at a process node, the system will automatically trigger the process refactoring module. This module dynamically adjusts the edge weights in the quality knowledge graph, cutting off inefficient paths and establishing efficient connections. Simultaneously, the agent calculates the reward after execution (i.e.,...). The policy network is continuously updated (by increments), and new and excellent solutions are stored in the policy library, thereby enabling the policy library to be self-enriched and optimized.

[0202] Furthermore, the system writes the optimized logical parameters back to the underlying metadata database, enabling self-evolution at the management logic level.

[0203] To ensure the smoothness of the system optimization process, this invention specifically sets up a shadow execution mode. The logic evolution module first runs the optimized management strategy in a simulation environment built based on digital twin technology. This digital twin environment, through fitting a physics engine with historical big data, can simulate the key dynamic responses of a real production line with engineering-acceptable high precision. Verification shows that its prediction error is less than 5% of actual production fluctuations, sufficient to reliably assess the potential impact of the new strategy. Only when the optimization scheme shows convergence after 1000 consecutive cycles in shadow mode, and all risk indicators (such as equipment load peaks) are within a safe range, will the system dynamically adjust the edge weights in the quality knowledge graph through the logic reconstruction module and automatically update the metadata databases in the cloud and at the edge.

[0204] Furthermore, to enhance the interpretability and execution efficiency of the optimization strategy, a policy distillation unit is configured within the logic evolution module. This unit employs a teacher-student model architecture, where the teacher model is a deeply trained high-dimensional reinforcement learning neural network, and the student model is a lightweight decision tree or logistic regression model. By utilizing the predicted probabilities (soft labels) output by the teacher model to guide the training of the student model, the system successfully "distills" complex policies into concise logical rules that are easily understood by frontline operators and executed by the underlying low-power controller. This design, while maintaining deep self-optimization capabilities, significantly reduces the computational overhead and cognitive costs of the system in actual operation.

[0205] In a preferred embodiment of the present invention, the system also integrates a quality traceability module based on blockchain technology. This module employs a consortium blockchain architecture composed of nodes from various functional departments within the enterprise. Each time the PDCA cycle completes a closed-loop evolution, the system encapsulates the key parameters of this cycle, including the initial topological hash, statistical characteristic values ​​during execution, the causal weight matrix for deviation analysis, and the final evolution correction instructions. A globally unique evidence snapshot is generated using the SHA-256 hash algorithm, and standard UTC timestamp information is attached. This data block is synchronized to all nodes on the chain using a consensus algorithm (such as PBFT or Raft). This design not only ensures the immutability of the quality self-optimization trajectory but also provides an undeniable data foundation for subsequent compliance audits and long-term quality traceability.

[0206] In terms of hardware support, the cloud computing cluster adopts a distributed storage architecture and uses a high-speed solid-state drive array to form a cold and hot data layered storage system, ensuring that the quality knowledge graph containing tens of millions of relation edges can achieve millisecond-level structured queries.

[0207] The edge computing nodes utilize an industrial-grade protective housing made of aluminum-magnesium alloy, with a wide operating temperature range of -40℃ to +70℃ and IP67 waterproof and dustproof capabilities, allowing direct deployment in welding workshops or high-humidity painting workshops with strong electromagnetic interference. The end-point sensing sensor network connects to an isolated industrial bus (such as Profibus-DP) via shielded twisted-pair cables. Its signal conditioning circuitry features a 1.5kV opto-isolation layer, ensuring the integrity of weak sensor signals transmitted in complex industrial environments.

[0208] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to specific implementation examples.

[0209] Example 1: Deploying the system described in this invention in a precision reducer digital factory. The production line comprises 12 main processes and involves 25 automated devices.

[0210] Planning Phase: The logic planning module retrieves historical data from the quality knowledge graph and identifies a correlation weight of 0.85 between the coaxiality of the planetary carrier and the bearing press-fitting force. Weighting coefficients are then set. , , Based on the quality objective function, an assembly topology containing a "real-time torque feedback closed loop" was generated after 500 iterations using a genetic algorithm. This structure requires the pressing station to collect pressing force data every 10ms and compare it with a standard curve in real time.

[0211] Execution Phase: The intelligent execution terminal collects the pressing curve in real time at the pressing station. The real-time calculation engine uses a sampling period of 10ms and a sliding window size of 100 sampling points (i.e., 1 second) to calculate the moving average and sliding variance of the pressing peak torque.

[0212] Inspection phase: The deviation analysis module detected the residual vectors of 5 consecutive products. The peak torque component exceeded a preset threshold (the threshold was set to ±5% of the baseline value). A causal graph was constructed using a causal inference algorithm, incorporating variables such as pressing force, material hardness, and equipment vibration. The average causal effect (ACE) of each variable on the pressing quality fluctuation was calculated, where the pressing quality index Y was defined as the root mean square error (RMSE) between the peak pressing force and the standard curve, in kN. The results showed that the average causal effect (ACE) of the microcracks in the pressing head (indirectly manifested through vibration characteristics) on the RMS error was 0.35 kN, and the material hardness fluctuation... The value is 0.15 kN, the vibration amplitude of the equipment base. The value is 0.05 kN. This is due to the micro-cracks in the press-fit head. The largest value indicates that it has the strongest causal effect on the fluctuation of the pressing force. Therefore, the system accurately locates the micro-crack at a certain point in the pressing head, which is the main cause of the force transmission loss.

[0213] In the handling phase: Based on the root cause analysis results, the logic evolution module retrieves the "preventive maintenance" plan from the strategy library and triggers an early warning. The system first executes the immediate action of "adjusting the frequency of subsequent inspection processes" (from sampling 1 out of 10 pieces to full inspection) as an emergency response. Simultaneously, based on the causal diagnosis result of "micro-cracks in the pressing head," the logic evolution module automatically optimizes the maintenance strategy: dynamically adjusting the preventive maintenance cycle of the pressing head from "every 100,000 pressing cycles" based on fixed operating time to a health status prediction model based on vibration spectrum characteristics. When the energy value of a specific frequency band (e.g., 5kHz-8kHz) in the vibration spectrum exceeds a dynamic threshold, a maintenance work order is automatically triggered. The reinforcement learning model employs the Q-learning algorithm. To handle continuous states (e.g., equipment vibration characteristic values, production plan urgency), the state space is discretized (e.g., equal-frequency discretization) or a function approximation method such as a deep Q-network is used. Its state space includes the current root cause type (microcracks, material hardness, etc.), equipment vibration characteristic values, and production plan urgency; the action space includes optional immediate handling solutions, such as "adjusting the detection frequency" and "modifying control parameters"; the reward function is defined as the increment of the quality objective function after executing the immediate action, i.e.

[0214]

[0215] in, and These are the quality objective function values ​​calculated before and after the action was performed. Because... Comprehensive consideration of process compliance Resource efficiency and quality stability This reward function comprehensively reflects the improvement effect of the handling actions on the overall quality system efficiency. The model updates the Q-table based on the reward obtained from this event, thereby increasing the priority of the "adjust detection frequency" strategy in similar situations, and storing the early vibration characteristics of the microcrack in the press-fitting head as new state characteristics in the strategy library, so as to identify similar risks earlier in the future. For long-term actions such as "preventive maintenance of equipment," their effects will be indirectly reflected through changes in the state space in subsequent PDCA cycles, and will affect future decisions.

[0216] Comparative Example 1: Under the same production environment, a traditional manual inspection combined with a fixed threshold control statistical model was adopted. Quality management personnel discovered an anomaly in the pressing process through daily reports and organized an engineering meeting 24 hours later to investigate the root cause. Due to a lack of causal inference tools, the investigation process took 4 hours, and the initial conclusion was mistakenly attributed to a material tolerance issue.

[0217] To quantify the technical superiority of the present invention, the core indicators of the above embodiments and comparative examples were recorded and compared over a one-month operating period. Specific data are detailed in Table 1 below.

[0218] Table 1: Comparison of the operational efficiency of a quality system self-optimization system based on the PDCA cycle;

[0219]

[0220] Based on the quantitative experimental data in Table 1, it can be concluded that this invention achieves a leap from "days" to "seconds" in anomaly perception and decision-making time; the real-time computing engine and deviation analysis module constructed by this invention can instantly capture quality fluctuations. Regarding root cause identification accuracy, due to the introduction of causal inference and residual neural networks, this invention can effectively remove environmental noise and accurately pinpoint core deviation factors, improving accuracy to 96.8%. In the most critical aspect of quality loss cost control, the self-optimizing system, through "instantaneous correction" and "autonomous updating of management logic," significantly reduces the batch production of defective products caused by system lag, resulting in a monthly average loss cost reduction of 86.7%. Furthermore, the system's resource scheduling efficiency score improved from 0.58 to 0.92, demonstrating the effectiveness of the quality objective function. The logic planning demonstrates superior performance in balancing compliance, efficiency, and stability.

[0221] Furthermore, after six months of continuous operation, the number of nodes in the quality knowledge graph generated by the system increased from the initial 12,000 to 85,000, and the edge weights were refined through reinforcement learning reward feedback. This means that the system's modeling accuracy for complex business scenarios continuously improves with increasing runtime. In shadow execution mode, the system can predict and avoid more than 90% of potential process conflicts, ensuring the high robustness of the production system in dynamic environments.

[0222] Example 2: This example is a further optimized solution based on Example 1. It is especially suitable for high-end manufacturing scenarios with complex processes, strong variable coupling, and extremely high requirements for root cause localization accuracy, such as aero-engine blade assembly and precision reducer core component assembly.

[0223] This embodiment introduces a structural causal model for quantitative root cause diagnosis and employs strategy distillation technology to achieve lightweight and secure logical evolution, thereby significantly improving optimization depth and reliability while maintaining system real-time performance.

[0224] Step A: High-fidelity full-element perception and causal structure learning;

[0225] High-frequency data is collected through a network of end-point sensing sensors deployed on the production site. Taking the assembly of high-pressure compressor rotor blades for aero-engines as an example, this production line includes 12 key workstations such as heat fitting, tightening, and clearance measurement, involving more than 200 process variables.

[0226] Sensor configuration: A dual-axis laser displacement sensor (measurement accuracy ±0.001mm, sampling frequency 2kHz) is used to monitor the blade installation angle; a dynamic torque sensor (sampling frequency 5kHz) is used to capture instantaneous torque fluctuations during the tightening process; and an infrared thermal imager (thermal sensitivity 0.05℃) is used to monitor the uniformity of the heat-fitting temperature. All sensor signals are converged to the edge computing node via an isolated industrial bus (Profinet IRT) to ensure time synchronization accuracy better than 1μs.

[0227] Data preprocessing: Edge computing nodes filter and resample the raw time-series data to generate data containing process variables. and final quality indicators Standardized dataset (e.g., blade tip gap uniformity, in mm) It is then uploaded to the cloud computing cluster in real time via 10 Gigabit fiber optic cable.

[0228] Cause-and-effect graph construction: In the cloud, the deviation analysis module is based on the dataset. A fast causal inference algorithm is used for causal structure learning. This algorithm can handle complex scenarios with latent variables. The algorithm performs the following operations:

[0229] Conditional independence test: Calculate the partial correlation coefficient and perform a Fisher's Z test (significance level). ), to determine the conditional independence between any two variables.

[0230] Latent variable detection: when three variables , , satisfy and Independent, but given hour and When they are not independent, it indicates the presence of unobserved latent variables. The algorithm identifies latent variables by detecting this pattern of conditional independence and constructs a causal graph containing the latent variables.

[0231] Skeleton and Direction Recognition: An undirected graph skeleton is constructed based on the independence test results. The direction of the edges is determined using the V-structure and Meek rules, and the final output is a directed acyclic graph, i.e., a causal graph. The nodes in the graph represent all variables, and the directed edges... express yes The direct cause.

[0232] For example, in actual production lines, the fast causal inference algorithm, through conditional independence tests, found that the partial correlation coefficient between "tightening torque of bolt #5" and "blade tip clearance uniformity" is less than that of other observed variables. Fisher's Z-test The null hypothesis of conditional independence cannot be rejected; however, given the "flange deformation", the partial correlation coefficient between the two is... , The conditional independence relationship was broken. This successfully identified "flange deformation" as the key latent variable connecting "tightening torque" and "gap uniformity," thus constructing a precise causal graph containing the latent variable.

[0233] Step B: Quantitative root cause diagnosis based on Do-Calculus;

[0234] When the multi-model fusion anomaly detection strategy (same as in Example 1) detects a quality indicator Exceeding specification limits (e.g.) When this occurs, the system initiates root cause diagnosis based on do-calculus. This embodiment calculates each candidate causal variable. right The average causal effect is used to quantify its true contribution.

[0235] The average causal effect is defined as:

[0236]

[0237] in:

[0238] Describing the cause variable For outcome variables The average causal effect;

[0239] This indicates that intervention will Forced to set as value and cut off all pointers Causal edge;

[0240] This indicates an abnormal state (such as the current fluctuating torque value);

[0241] Represents the standard state (such as the target value of the process specification);

[0242] This represents the expected value.

[0243] To estimate the average causal effect unbiasedly from observational data, this invention employs the backdoor criterion. The system utilizes the causal graph constructed in step A. Automatically identify the minimum set of adjustment variables that satisfy the backdoor criterion. Then from the historical database Extract all The conditional probability of the sample is calculated using a nonparametric method of kernel density estimation. Therefore, we get:

[0244] Among them, summation traversal of the adjustment variable set All possible combinations of values ​​for all variables. Substitute them. and The average causal effect is then obtained.

[0245] Example data: In a single event of excessive tip clearance, the system calculated the following:

[0246] variable The average causal effect value of (tightening torque of bolt #5) is 0.032 mm;

[0247] variable The average causal effect value of (blade heating temperature) is 0.001 mm;

[0248] variable The average causal effect value of (positioning repeatability of the second axis of the assembly robot) is 0.028 mm.

[0249] Although, in terms of correlation, temperature and gap The Pearson coefficient may be higher (e.g., 0.6), but the average causal effect value clearly indicates that the tightening torque... and positioning accuracy This is the root cause of quality fluctuations, and temperature... This is merely an accompanying phenomenon. The system thus achieves a leap from "correlation analysis" to "causal diagnosis," providing a physical basis for precise policy implementation.

[0250] Step C: Lightweight logic evolution and shadow execution based on strategy distillation;

[0251] This step transforms the causal diagnostic results into safe, actionable, and low-overhead field control strategies.

[0252] Reinforcement Learning Agent: The logical evolution module incorporates a reinforcement learning agent based on a deep deterministic policy gradient algorithm. Its state space... Including the real-time value of the root dependent variable ( , Equipment remaining life, material inventory and production plan; motion space The set of compensation strategies includes, for example, "adjusting the target torque of bolt #5" and "increasing the frequency of online calibration of the robot's second axis"; the reward function. Defined as the degree of improvement in quality indicators after taking action. The agent is trained offline on historical data to generate a high-dimensional teacher policy network. .

[0253] Policy distillation: To deploy complex policies to edge computing nodes with limited computing power, this invention employs a teacher-student model for knowledge distillation.

[0254] Teacher Model: A trained deep deterministic policy gradient agent It runs in the cloud.

[0255] Student model: Lightweight gradient boosting decision tree model, or a set of interpretable IF-THEN rules.

[0256] Distillation process: The teacher runs the process in a simulation environment, generating a massive amount of (states) Action probability distribution Data pairs (including soft labels). Student model with Using this probability distribution as input, and taking this probability distribution as the learning objective, the model is trained by minimizing the cross-entropy loss. The final student model is obtained. It is only a few hundred KB in size and can be deployed on edge nodes.

[0257] Shadow Execution: To ensure evolution safety, any new student model or instructions generated by it are not directly applied to the real production line. The system launches a digital twin running synchronously with the physical production line in a backup computing unit on an edge computing node. The new strategy is first "shadow executed" in the digital twin for at least 10 complete production cycles. The simulation accuracy of the digital twin meets engineering verification requirements, reflecting the key dynamic characteristics of the real production line with a prediction error of less than 5%, ensuring the reliability of the new strategy evaluation, and accurately predicting the impact of the new strategy on quality and equipment load. Only when the shadow execution results show that the risks are controllable (e.g., equipment load is not exceeded and cycle times are not delayed) will the system seamlessly switch the new strategy to the real environment through blue-green deployment technology.

[0258] Through the above steps, this embodiment achieves real-time, lightweight, and secure decision-making at the edge while maintaining the powerful learning capabilities of the cloud, completing a deep closed loop from "causal diagnosis" to "secure handling".

[0259] In actual testing on an aero-engine blade assembly line, this embodiment achieved the following significant results compared to an anomaly detection scheme that only uses a general PDCA cycle and data-driven approach (such as the comparative example in Embodiment 1):

[0260] For complex coupled faults with 50 random injections, traditional correlation methods achieved an accuracy of 62%, while this embodiment achieved a causal diagnosis accuracy of 97%, and could quantitatively provide the contribution of each factor (average causal effect value). Through policy distillation, the size of the edge decision model was compressed from >500MB to <5MB, and the single decision inference time was shortened from >200ms (requiring communication with the cloud) to <10ms (local execution). The shadow execution mode successfully intercepted two potential risks of equipment overload due to overly aggressive policies, ensuring zero production interruption. Repetitive quality anomalies caused by the same root cause were reduced by 90%, and downtime caused by quality anomalies was reduced by 85%.

[0261] Example 3: This example is a further development based on Example 1. It addresses the pain points of complex causal coupling and high trial-and-error costs in reinforcement learning during the assembly of precision reducers by introducing a causal-guided reinforcement learning framework to achieve efficient and safe strategy optimization.

[0262] The system architecture and module interactions are as follows: Based on the hardware of Example 1, this example adds a causal effect prediction network to the cloud computing cluster and deeply integrates it with the original reinforcement learning agent (logic evolution module). The system operation flow is as follows:

[0263] Cause-effect graph construction and average causal effect calculation (same as Example 1): The bias analysis module uses a fast causal inference algorithm to learn the cause-effect graph from historical data. And calculate each causal variable. Average causal effect on key quality indicator (tip clearance uniformity) The data is stored in a causal knowledge base. In this embodiment, the quality indicators... This represents the root mean square error of the blade tip gap uniformity, in mm.

[0264] Training a causal effect prediction network: Extracting all historical intervention events from a causal knowledge base , where: state : Includes real-time values ​​of root variables (such as the tightening torque of bolt #5) Robot's second axis positioning repeatability The data includes 15 dimensions, such as equipment vibration characteristic values ​​and production plan urgency. Specifically: (Nm) (mm), Effective value of vibration acceleration (m / s²), production cycle urgency (Dimensionless, 0~1) Material hardness deviation (HRC), ambient temperature (°C), ambient humidity (%), Pass rate of preceding processes (%), Remaining equipment lifespan (%), current batch material batch number code (one-hot) Operator proficiency (Grade 1~5) Tool Wear (mm), clamping force (kN), Lubrication status (0 / 1) Historical fluctuation trend of quality (-1~1).

[0265] action 8-dimensional continuous motion, with each dimension normalized to This corresponds to the percentage change in the actual controlled quantity. Specific actions include: (Torque adjustment ratio) (Calibration frequency adjustment ratio) (Clamping force adjustment ratio) (Heating time adjustment ratio) (Feed rate adjustment ratio) (Air pressure adjustment ratio) (Nozzle distance adjustment ratio) (Motor current limit adjustment ratio).

[0266] : The change in the root mean square error of the blade tip clearance uniformity of 10 consecutive products after the action is performed, in mm. If the quality improves after the action (root mean square error decreases), then It is a negative value; if it worsens, it becomes a positive value.

[0267] Constructing a three-layer fully connected neural network The input layer dimension is the sum of the state dimension and the action dimension (15+8=23 dimensions), the number of hidden layer neurons are 256 and 128 respectively, the activation function is ReLU, and the output layer dimension is the number of quality indicators (in this embodiment). (i.e., the output layer is 1-dimensional). Using the Adam optimizer with a learning rate of 0.001, training for 500 epochs resulted in a loss function that converged to 0.002 mm². During training, early stopping (patience=20) was employed to prevent overfitting, and the loss was monitored on the validation set. The loss function was defined as:

[0268]

[0269] in:

[0270] The loss function;

[0271] Parameters of the network for predicting causal effects;

[0272] Historical intervention dataset;

[0273] For the first The state of each sample;

[0274] For the first The actions of each sample;

[0275] For the first The change in quality indicators corresponding to each sample;

[0276] For predicting causal effects, the network pairs The predicted value;

[0277] The Euclidean norm of a vector;

[0278] This indicates the number of samples in the historical intervention dataset.

[0279] The historical intervention data was obtained through the following methods:

[0280] In the initial stage of system deployment: adopt an initialization strategy based on domain expert rules or -greedy explores strategies to safely generate large amounts of simulated intervention data in a digital twin simulation environment, covering the broadest possible state-action space, for pre-training causal effect prediction networks.

[0281] During system operation: continuously extract naturally occurring intervention events and their corresponding quality indicator changes from the historical logs of the production system to form a real intervention dataset.

[0282] Continuous learning: Using newly generated real intervention data, the causal effect prediction network is continuously fine-tuned (online learning) to improve its prediction accuracy.

[0283] Online optimization of causal-guided reinforcement learning:

[0284] For clarity, in this embodiment, the system performance function is defined. (Definitions same as in Example 1) Used for macroscopic evaluation of the overall quality system's operational effectiveness; key quality indicators defined. This is a specific, physically measurable vector of quality parameters (in this embodiment, it is the root mean square error of the blade tip gap uniformity).

[0285] state space 15 dimensions, same as above. .

[0286] Action space 8-dimensional continuous motion, normalized to .

[0287] Actual reward According to the system efficiency function The increment, i.e. ,in The calculation methods for each indicator are the same as in Example 1. Here... It evaluates system effectiveness from a macro-process level, while causal effect prediction networks Output It is aimed at key quality indicators The predicted change in (tip clearance uniformity). Together, these two constitute a multi-dimensional assessment of the system state.

[0288] Cause-and-effect guided reward Predicting network output based on causal effects The opposite of the weight (This embodiment only considers a single quality indicator) to obtain, that is .because A negative value indicates an improvement in prediction quality (i.e., a reduction in root mean square error). Taking the opposite value... A positive value indicates a positive reward for actions that improve quality, which aligns with the design principles of reinforcement learning reward functions.

[0289] Total Rewards: . The search range (0.1~1.0, step size 0.1) was pre-optimized in a simulation environment using a grid search method, with the average value after convergence of the final strategy being used as the benchmark. Values ​​are indicators. This is the causal guidance coefficient, used to balance the weights of actual rewards and causal guidance rewards.

[0290] Both the policy network and the value network employ a deep deterministic policy gradient architecture with two hidden layers (400, 300). The policy network output layer is 8-dimensional, and tanh activation is used to restrict actions to... Value network outputs 1D state-action value. Network parameters are initialized using Xavier.

[0291] Causal gradient guidance during policy network updates: An additional causal gradient term is added when calculating the policy gradient. ,in This was also determined through simulation pre-optimization. This incentive strategy selection can produce a negative causal effect (i.e., a significant improvement in prediction quality, because...). When it is negative, The gradient term (positive) guides the policy towards accelerating its exploration towards the optimization objective. The gradient term is used here instead of the reward directly to more directly inject causal knowledge into the policy network's update process, in conjunction with the total reward. They complement each other.

[0292] The formula for calculating the policy gradient is:

[0293]

[0294] in:

[0295] For policy gradient;

[0296] For policy network parameters;

[0297] This refers to the number of samples in a small batch.

[0298] The output of the value network is used to evaluate the state. Take action below Expected cumulative reward;

[0299] For the policy network in the 1st Individual sample status The action to be output;

[0300] This indicates that the value network output corresponds to the action. The gradient;

[0301] The policy network output is related to the parameters. The gradient;

[0302] To guide the exploration of causal factors;

[0303] Predict network output for causal effects;

[0304] The square of the output modulus of the causal effect prediction network is represented by the action. The gradient;

[0305] This represents the Euclidean norm of a vector.

[0306] The training and execution process is as follows:

[0307] In the action phase of each PDCA cycle, the system starts from the current state. Initially, the policy network outputs actions. .

[0308] The causal effect prediction network immediately provides... and calculate .

[0309] action The commands are converted into actual control instructions and issued to the production line (after verification by shadow execution). After execution, actual control results are observed. and new status And calculate the actual reward. .

[0310] Experience replay pool storage and for training causal effect prediction networks .

[0311] Each step updates the network by randomly sampling a mini-batch (batch size=64) from the replay pool: the value network is updated by minimizing the temporal difference error, the policy network is updated using the gradient formula mentioned above, and the causal effect prediction network is updated using the newly collected... The data is continuously fine-tuned (online learning, learning rate lowered to 0.0001).

[0312] Shadow Execution and Deployment: Similar to Example 1, the new strategy is first validated in a digital twin environment for 10 cycles to ensure that the risks are controllable before switching to the real production line via blue-green deployment. The simulation accuracy of the digital twin environment is higher than 99%, which can accurately predict the impact of the new strategy on quality and equipment load.

[0313] The key calculation process is illustrated below:

[0314] In an abnormal event, the system detected that the tip clearance was out of tolerance. Causal diagnosis gives the main root cause. (Average causal effect = 0.032 mm) and (Average causal effect = 0.028 mm). Current state. The specific values ​​are shown in Table 2 below:

[0315] Table 2:

[0316]

[0317] Policy network output actions As shown in Table 3:

[0318] Table 3:

[0319]

[0320] Causal effect prediction network input (15-dimensional) and (8-dimensional), obtained through forward computation This means that the root mean square error of the gap will decrease by 0.018 mm after the action is performed.

[0321] Actual reward The total reward for the current step needs to be calculated after the action is executed. When used to train a value network, predictions are used. As an immediate guide. Value network based on The updated policy network simultaneously maximizes the value network output and the causal gradient, making subsequent actions more likely to produce a more significant causal effect.

[0322] In practice, a three-month comparative test was conducted at a precision reducer factory, running both the standard deep deterministic policy gradient (without causal guidance) and the causal-guided deep deterministic policy gradient of this embodiment. The results are shown in Table 4 below:

[0323]

[0324] Causal-guided rewards provide prior knowledge, enabling agents to avoid ineffective exploration and halving the number of iterations. The final strategy more accurately eliminates root causes, reducing monthly losses by 32.1%. Causal prediction allows agents to quickly select effective actions, shortening anomaly recovery time by nearly half. Causal guidance aligns the strategy with physical mechanisms, making it easier for experts to understand and trust. Causal gradient terms guide the strategy to avoid aggressive actions that could lead to overreactions by the device, resulting in zero accidents.

[0325] In practical applications, reinforcement learning requires extensive trial and error in complex manufacturing environments, which is costly and may lead to safety accidents; agents struggle to understand the causal relationship between actions and quality indicators, are easily misled by spurious correlations, and have low learning efficiency; optimization strategies lack interpretability and are difficult to gain the trust and adoption of field engineers.

[0326] This embodiment leverages the prior knowledge provided by the causal effect prediction network, enabling the agent to avoid ineffective actions from the outset, significantly reducing trial-and-error attempts and mitigating the risk of quality fluctuations during system self-optimization. Causal-guided rewards encourage the agent to prioritize actions with direct causal effects on the root cause, avoiding superficial local optimizations and ensuring the final strategy fundamentally eliminates quality defects. Causal gradient terms constrain the policy exploration direction, preventing equipment overload; simultaneously, the interpretability of the causal effect prediction network output allows each decision to be traced back to its physical basis, facilitating engineer review and intervention. This embodiment organically integrates causal inference (diagnosis) and reinforcement learning (decision-making) through mathematical formulas. Diagnostic results directly participate in reward shaping and policy gradient calculation, forming a tightly coupled closed loop of "diagnosis-prediction-decision-verification," significantly enhancing the system's intelligence level.

[0327] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0328] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-optimization method for a quality system based on the PDCA cycle, characterized in that, Includes the following steps: Step 1: Collect multi-source heterogeneous quality data and construct a quality knowledge graph containing quality element nodes and their logical association edges. The quality element nodes include personnel nodes, equipment nodes, material nodes, method nodes, environment nodes, and measurement data nodes. Step 2: Based on the quality knowledge graph, a causal inference algorithm is used to construct a causal graph and calculate the average causal effect of each causal variable on the quality index to locate the root cause variable of quality fluctuation; Step 3: Using the real-time monitoring values ​​of the root dependent variable, system parameters, and resource status as the state space, and the disposal schemes in the policy library as the action space, a reinforcement learning agent makes decisions. The reinforcement learning agent is a neural network model built on a deep deterministic policy gradient framework, including a policy network and a value network. The policy network takes the state space vector as input and outputs continuous control parameters in the action space. The value network takes the state-action pair as input and outputs an estimate of the expected cumulative reward. Step 4: Set up a causal effect prediction network, take the current state and candidate actions as input, output the predicted causal effect vector of the quality index, and calculate the causal guidance reward based on the inverse of the predicted causal effect vector. Step 5: Combine the actual reward from the environmental feedback with the causal guidance reward in a weighted manner to obtain the total reward, which is used to update the value network and policy network of the reinforcement learning agent; Actual rewards for environmental feedback According to the quality objective function The increment is calculated; The quality objective function Defined as: ， in: , , The weighting coefficients are preset according to the business scenario, representing the relative importance of compliance, efficiency, and stability, respectively. The process compliance score represents the number of time-series iteration steps; Resource utilization efficiency representing the number of iteration steps in the time series; The stability of quality output represents the number of iteration steps in the time series. Indicates the number of timing iterations; The policy network's update gradient includes a causal guidance term, which is determined based on the gradient of the predicted causal effect vector with respect to the action. The value network updates parameters by minimizing the temporal difference error, and the policy network updates parameters by policy gradient ascent. During the update, historical transfer samples are sampled from the experience replay pool for offline training. Step 6: After verifying the security of the optimal solution output by the reinforcement learning agent in shadow execution mode, deploy it to the production environment for execution.

2. The quality system self-optimization method based on the PDCA cycle according to claim 1, characterized in that, The collection of multi-source heterogeneous quality data in step 1 includes: Access to the operating status parameters of production equipment, process log files of business management system, and structured reports of quality inspection process is achieved through a distributed data acquisition gateway. For unstructured document quality standards and regulations, a natural language processing engine is used for feature extraction. The natural language processing engine uses a bidirectional long short-term memory network combined with an attention mechanism to transform the input text into structured knowledge containing entities and their semantic attributes. The extracted entities are used as nodes in the quality knowledge graph, and the relationships between entities are used as directed edges. Each node is associated with a node feature vector, which is formed by concatenating a semantic feature vector and a structured attribute vector.

3. The quality system self-optimization method based on the PDCA cycle according to claim 2, characterized in that, The natural language processing engine extracts quality-related entities including personnel, equipment, materials, methods, environment, and measurement data entity types, and extracts relationship types including logical associations of "specification", "requirement", "responsibility", and "cause". The semantic feature vectors are obtained by encoding the contextual semantics of entities through a bidirectional long short-term memory network, with the dimension uniformly set to 512. If a node has no corresponding text description, the semantic feature vector is filled with all zeros or a default vector is generated by a pre-trained model. The structured attribute vector is obtained by normalizing or embedding the structured attributes of the nodes, and its dimension is preset to a fixed value; the structured attributes are stored in the nodes in the form of key-value pairs for easy independent querying.

4. The quality system self-optimization method based on the PDCA cycle according to claim 1, characterized in that, The causal inference algorithm in step 2 includes: A fast causal inference algorithm or PC algorithm is used to learn and construct a causal graph containing all relevant variables from historical data. The causal graph is a directed acyclic graph, where nodes are quality elements and directed edges represent causal relationships. For suspected causal variables, the do-calculus method is used to calculate the intervention distribution, and the minimum set of adjustment variables is identified using the backdoor or frontdoor criterion. The average causal effect of each causal variable on the quality index is then calculated. in, Indicates the average causal effect; As a causal variable; For quality indicators; Indicates an abnormal state; Indicates the standard state; This indicates an intervention operation, which is about to begin. Forced to be set as And cut off other variables The impact; Indicates the expected value; When calculating the average causal effect, the backdoor criterion or the frontdoor criterion is used to estimate the causal effect from the observed data, quantify the contribution rate of each causal variable to the quality fluctuation, and determine the causal variable with the highest contribution rate as the root cause variable.

5. The quality system self-optimization method based on the PDCA cycle according to claim 1, characterized in that, In step 4, the causal effect prediction network is a three-layer fully connected neural network structure. The input layer dimension is the sum of the state dimension and the action dimension. The number of hidden layer neurons is 256 and 128, respectively. The activation function is ReLU. The output layer dimension is the number of quality indicators. The training process of the causal effect prediction network includes: using historical intervention datasets Network parameters are trained by minimizing the mean squared error loss function. : in, The loss function; Parameters of the network for predicting causal effects; Historical intervention dataset; For the first The state of each sample; For the first The actions of each sample; For the first A vector of changes in quality indicators corresponding to each sample; For predicting causal effects, the network pairs The predicted value; The Euclidean norm of a vector; Indicates the number of samples in the historical intervention dataset; The historical intervention data is obtained through the following methods: extracting naturally occurring intervention events from the historical logs of the production system, generating simulated intervention data in a digital twin simulation environment, or collecting real intervention data through a security exploration strategy in the early stages of system deployment.

6. The self-optimization method for a quality system based on the PDCA cycle according to claim 1 or 5, characterized in that, The formula for calculating the total reward in step 5 is as follows: in, Total reward; A tangible reward for environmental feedback; This is the causal guidance coefficient, used to balance the weights of actual rewards and causal guidance rewards; The reward is causally guided; the actual reward. The increment of the quality objective function is defined as: in, To enhance the number of sequential iterations in learning interaction The corresponding comprehensive quality objective function value is based on the number of time-series iteration steps. The calculation is based on real-time monitoring data within a preset time window that is initially slid forward. The number of iterations in the previous interaction sequence. The corresponding comprehensive quality objective function value; The causal guided reward The output calculation based on the causal effect prediction network is defined as follows: in, The quantity of quality indicators; For the first The preset weights of each quality indicator satisfy ; The first prediction of the network output for causal effects Each quality indicator in the number of time-series iteration steps The predicted change.

7. The quality system self-optimization method based on the PDCA cycle according to claim 1, characterized in that, In step 5, the reinforcement learning agent adopts a deep deterministic policy gradient framework, which includes a policy network and a value network. The value network Used to assess the state Take action below The expected cumulative reward is output as a scalar value; the policy gradient calculation formula for the policy network is: ， in, For policy gradient; For policy network parameters; This refers to the number of samples in a small batch. Output to the value network; For the policy network in the 1st Individual sample status The action to be output; This indicates that the value network output corresponds to the action. The gradient; The policy network output is related to the parameters. The gradient; To guide the exploration of causal factors; Predict network output for causal effects; The square of the output modulus of the causal effect prediction network is represented by the action. The gradient; The value network is updated by minimizing the temporal difference error.

8. The quality system self-optimization method based on the PDCA cycle according to claim 1, characterized in that, The shadow execution mode in step 6 is specifically as follows: The optimized management strategy is run in a simulation environment built on digital twin technology. The digital twin environment simulates the key dynamic responses of the real production line with a prediction error of less than a preset percentage of the actual production fluctuations by fitting a physical engine with historical big data. When the optimization strategy runs continuously for a preset number of cycles in the simulation environment and passes the convergence verification, and all risk indicators are within the safe range, the optimization strategy is switched to the real production environment through blue-green deployment technology; the risk indicators include equipment load peak and cycle time delay.

9. The self-optimization method for a quality system based on the PDCA cycle according to claim 1, characterized in that, It also includes a strategy distillation step: A teacher-student model architecture is used to distill the high-dimensional policy network generated by reinforcement learning into a lightweight decision model. The teacher model is a trained high-dimensional reinforcement learning neural network, and the student model is a gradient boosting decision tree model or a decision tree model. During the distillation process, the action probability distribution output by the teacher model is used as a soft label, and the student model takes the state as input and the soft label as the learning target, and is trained by minimizing the cross-entropy loss. The trained student models are deployed on edge computing nodes to execute quality control decisions in real time.

10. A quality system self-optimization system based on the PDCA cycle, used to execute the quality system self-optimization method based on the PDCA cycle according to any one of claims 1-9, characterized in that, include: The all-element perception module is configured to collect multi-source heterogeneous quality data and construct a quality knowledge graph containing quality element nodes and their logical association edges. The quality element nodes include personnel nodes, equipment nodes, material nodes, method nodes, environment nodes, and measurement data nodes. The causal diagnosis module is configured to construct a causal graph based on the quality knowledge graph and use a causal inference algorithm to calculate the average causal effect of each causal variable on the quality index, thereby locating the root cause variable of quality fluctuation; the causal effect prediction network adopts a three-layer fully connected neural network structure, takes the current state and candidate actions as input, and outputs a predicted causal effect vector for the quality index. The reinforcement learning agent is configured to use the root cause variable, system parameters and resource status as the state space and the disposal schemes in the policy library as the action space for decision-making. The total reward function of the reinforcement learning agent is a weighted combination of the actual reward from environmental feedback and the causal guided reward, wherein the causal guided reward is calculated based on the output of the causal effect prediction network. The shadow execution module is configured to verify the security of optimization strategies in a simulation environment built on digital twin technology, and deploy the optimization strategies to the production environment after successful verification.

Images