Wire harness production quality control system based on artificial intelligence

Abstract

The invention provides a wire harness production quality control system based on artificial intelligence. The wire harness production quality control system comprises a wire harness data acquisition module, a wire harness knowledge graph module, a wire harness quality correlation analysis module, a wire harness quality causal reasoning module, a wire harness quality prediction module, a wire harness parameter optimization module and a wire harness decision support module. The system identifies complex association through a graph neural network, discovers problem root causes through time sequence analysis and causal reasoning, optimizes production parameters in combination with a virtual environment, and provides scientific decision support. According to the method, the technical problems of complex correlation identification, time-delay effect capture, causal relationship determination, parameter optimization and the like in wire harness production are solved, and intelligent quality control of the whole process is realized.

Description

Technical Field

[0001] The present invention belongs to the field of industrial intelligent manufacturing, and particularly relates to a wire harness production quality control system based on artificial intelligence. Background Art

[0002] As a key interconnection component in modern industrial products, the quality of wire harnesses directly affects the performance and reliability of the whole machine. The current technical problems in wire harness production quality control mainly include the following:

[0003] 1. The wire harness production process involves multiple processes such as wire stripping, crimping, and packaging. The relationship between the parameters of each process and the final quality is complex and difficult to accurately describe.

[0004] 2. There is a time-lag effect in the adjustment of production parameters, and traditional analysis methods are difficult to effectively capture the delayed quality impact caused by parameter changes.

[0005] 3. Existing technologies are mostly based on simple statistical correlation analysis and fail to effectively identify multi-level causal relationships, resulting in root cause analysis relying on empirical judgment.

[0006] 4. The optimization of process parameters mainly relies on trial-and-error methods, lacking a systematic multi-objective balance optimization mechanism, with low efficiency and high cost.

[0007] Existing technologies have not systematically solved the problems of quality prediction, root cause analysis, and parameter optimization in wire harness production. There is an urgent need for a wire harness production quality control system that can comprehensively apply advanced artificial intelligence algorithm technologies. Summary of the Invention

[0008] To solve the above technical problems, the present invention provides a wire harness production quality control system based on artificial intelligence, and the system composition is as follows:

[0009] Wire harness data acquisition module: Set on the wire harness production line and enterprise server, used to collect and manage wire harness production data; this module connects wire harness production equipment through PROFINET, EtherCAT, and OPC UA industrial communication protocols, collects raw material parameters, process parameters, image data, and electrical test data, performs data cleaning and anomaly detection based on the Savitzky-Golay algorithm and the modified Z-score method, extracts time-domain, frequency-domain, and statistical features, and then classifies and stores the data; this module is connected to the wire harness knowledge graph module and the wire harness quality correlation analysis module through a standard data interface to transmit the processed production data.

[0010] Harness Knowledge Graph Module: It is set on the enterprise application server and is used to build a structured knowledge model for harness production. This module extracts harness production knowledge from the rule base, historical data, and production documents, and constructs a three-layer graph structure according to the production factor layer (including five types of nodes: raw materials, equipment, tools, environment, and personnel), the process layer (including three types of nodes: processes, operations, and inspections), and the quality feature layer (including four types of nodes: physical features, electrical features, reliability features, and appearance features). It uses the TransE algorithm and graph attention network to realize knowledge vectorization and stores it in the Neo4j graph database. This module is connected to the harness quality correlation analysis module and the harness quality causal reasoning module through knowledge interfaces to transmit structured knowledge.

[0011] Harness Quality Correlation Analysis Module: It is set on the enterprise GPU server and is used to mine the complex correlation relationships between harness production factors and quality features. This module uses a multi-relational heterogeneous graph convolutional network (MR-HGCN) to process the knowledge graph, realizes heterogeneous information fusion through the node type embedding layer, relationship perception layer, and message passing layer, identifies key influencing factors based on the attention mechanism, and uses an improved bidirectional Dijkstra algorithm to extract quality influence paths. This module is connected to the harness quality causal reasoning module and the harness quality prediction module through analysis interfaces to transmit correlation analysis results.

[0012] Harness Quality Causal Reasoning Module: It is set on the enterprise application server and is used to identify real causal relationships from the correlation relationships. This module processes the time series data of production parameters, establishes a multi-order time delay effect model, constructs an initial causal skeleton through conditional independence tests based on partial correlation coefficients, and verifies the causal relationships through virtual intervention and average treatment effect calculations. This module is connected to the harness parameter optimization module through causal interfaces to transmit causal knowledge.

[0013] Harness Quality Prediction Module: It includes a training part set on the enterprise server and an inference part set on the edge device. This module uses a teacher-student knowledge distillation architecture to build a quality prediction model. The teacher model integrates three sub-networks: a multi-layer perceptron, a long short-term memory network, and a graph convolutional network to process multi-source data. The student model uses a lightweight structure to adapt to edge computing resources, realizes knowledge transfer through a knowledge distillation loss function, and is deployed to the edge device for real-time prediction after quantization optimization. This module is connected to the harness decision support module through prediction interfaces to transmit quality prediction results.

[0014] Harness Parameter Optimization Module: It is set on the enterprise high-performance computing server and is used to realize the intelligent optimization of production parameters. This module constructs a virtual production environment, integrates a physical model based on the finite element method and a data-driven model based on a feedforward neural network, and uses the Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning algorithm to find the optimal parameter combination that balances quality, efficiency, and resource consumption in the virtual environment. This module is connected to the harness decision support module through an optimization interface to transmit the parameter optimization plan.

[0015] Harness Decision Support Module: It includes a computing part set on the enterprise server and a display part set on the operation terminal. This module uses the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to find the parameter plan that balances quality, efficiency, cost, and energy consumption, evaluates the risk of the plan based on the parameter sensitivity index, generates multi-dimensional radar charts and parallel coordinate charts to visualize the characteristics of the plan, and generates a decision report. This module interacts with production management personnel through the user interface to provide decision support information.

[0016] The above-mentioned functional modules are connected through standardized data interfaces to form a complete closed-loop system: The data flow of the harness data collection module flows to the harness knowledge graph module and the harness quality correlation analysis module. The knowledge of the harness knowledge graph module flows to the harness quality correlation analysis module and the harness quality causal reasoning module. The analysis results of the harness quality correlation analysis module flow to the harness quality causal reasoning module and the harness quality prediction module. The reasoning results of the harness quality causal reasoning module flow to the harness parameter optimization module. The prediction results of the harness quality prediction module and the optimization plan of the harness parameter optimization module flow to the harness decision support module. The decision suggestions of the harness decision support module guide production adjustment and trigger new data collection, forming a closed-loop optimization and constituting four working modes:

[0017] 1. Quality Monitoring Mode: The harness data collection module regularly collects production data, and the harness quality prediction module uses the optimized edge deployment model to predict real-time quality indicators. When the predicted value deviates from the preset target range, the harness decision support module generates a warning message and transmits it to the operation terminal. This mode enables the system to identify potential risks before actual quality problems occur, continuously monitor the production status, and ensure the stability of harness quality.

[0018] 2. Problem Diagnosis Mode: Automatically starts when the system detects quality anomalies or receives a manual diagnosis request; the harness quality correlation analysis module identifies the impact paths related to the problem based on a multi-relational heterogeneous graph convolutional network; the harness quality causal reasoning module applies conditional independence testing and intervention verification techniques to analyze the root causes and rank them according to the probability of occurrence; the harness decision support module generates a structured root cause analysis report and specific solution suggestions based on the analysis results; this mode provides technicians with accurate problem positioning and solutions, significantly shortening the problem-solving cycle.

[0019] 3. Parameter Optimization Mode: Starts in response to new product introduction or process improvement requirements; the harness parameter optimization module systematically explores the parameter space in a virtual production environment, combines physical models and data-driven models to evaluate the effects of parameter combinations; the system generates a Pareto optimal solution set through non-dominated sorting genetic algorithm and evaluates the performance of each solution in dimensions such as quality, efficiency, and cost; the harness decision support module generates final optimization suggestions based on the Pareto optimal solution set and risk assessment results; this mode realizes the scientific adjustment of harness production parameters, effectively improving product quality and production efficiency.

[0020] 4. Knowledge Accumulation Mode: Automatically starts regularly by the system or when new knowledge patterns are identified; the harness knowledge graph module extracts new knowledge from incremental production data and updates the three-layer graph structure; the harness quality prediction module and the harness parameter optimization module adjust the model parameters according to the new data; the system continuously improves the knowledge base and algorithm models through continuous learning, realizing continuous improvement of quality control capabilities; this mode enables the system to automatically accumulate and inherit knowledge in the field of harness production, reducing the dependence on expert experience.

[0021] Advantages of the present invention:

[0022] 1. Improve the ability of correlation recognition: By adopting a three-layer knowledge graph and multi-relational heterogeneous graph convolutional network technology, it can identify multi-level impact paths in harness production (such as the complex correlation of "crimping temperature → wire deformation → contact resistance → harness conductivity"), significantly enhancing the depth of correlation recognition and the ability to discover hidden correlations, providing a more accurate technical basis for quality optimization.

[0023] 2. Achieve accurate causal reasoning: Through a multi-order time-delay effect model and a causal verification method based on virtual intervention, it effectively distinguishes correlation from causality, significantly reducing the false positive rate of false associations, greatly shortening the time for locating the root cause of quality problems, and improving problem-solving efficiency.

[0024] 3. Accelerate the adaptation of the model to new products: Based on the teacher-student knowledge distillation architecture, it realizes the rapid adaptation and efficient deployment of the quality prediction model, significantly shortening the training cycle of the new product model, maintaining low-latency inference ability on edge devices, and reducing the dependence on labeled samples at the same time.

[0025] 4. Optimize the production parameter combination: Combine physical models, data-driven models, and reinforcement learning algorithms to systematically explore the parameter space, significantly improve the parameter optimization efficiency without affecting actual production, and improve the balance between quality and resource consumption.

[0026] 5. Enhance the scientific nature of decision-making: Through multi-objective optimization algorithms and risk assessment methods, provide parameter solutions that take into account quality, efficiency, and cost, significantly improve the decision-making accuracy and parameter adjustment consistency, and reduce the dependence on manual experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Shows the schematic architecture diagram of the system of the present invention;

[0028] Figure 2 Shows the flowchart of the working steps of the system of the present invention; DETAILED DESCRIPTION OF THE INVENTION

[0029] The exemplary embodiments of the present invention will be described in detail below with reference to the drawings.

[0030] Combined with Figure 1 , the present invention provides a wire harness production quality control system based on artificial intelligence, and the system composition is as follows:

[0031] A. Wire harness data acquisition module

[0032] The wire harness data acquisition module is set on the wire harness production line and the enterprise server, and is used to collect and manage wire harness production data. This module connects wire harness production equipment through industrial communication protocols, collects raw material parameters, process parameters, image data, and electrical test data, and classifies and stores the data after performing data cleaning and feature extraction; this module is connected to the wire harness knowledge graph module and the wire harness quality correlation analysis module through data interfaces to transmit the processed production data.

[0033] In this embodiment, the wire harness data acquisition module uses multi-protocol data acquisition technology to collect data throughout the wire harness production process. This module connects high-precision equipment such as wire stripping machines and crimping machines through the PROFINET protocol (cyclic update time 1ms), connects high-speed detection equipment through the EtherCAT protocol (synchronous jitter <100ns), and connects PLCs and testing equipment through the OPC UA protocol to form a data acquisition network covering all processes. The collected data is divided into four categories: raw material parameters (wire material, cross-sectional area, initial resistance), with a sampling frequency of once per batch to form a batch-level quality baseline; process parameters (wire stripping force, wire stripping length, crimping pressure, crimping temperature, packaging tension), with a sampling frequency of once per product to capture process fluctuations; image data (wire harness appearance image, crimping cross-section image) for identifying visual defects; and electrical test data (wire harness resistance value, insulation resistance value, conduction test result), with a sampling frequency of once per product to record the final quality characteristics. All types of data are associated with the product ID and timestamp to establish a data chain throughout the production process.

[0034] To address the technical problem of the complex relationship between wire harness production parameters and quality, this module implements a three-level signal processing method. The first level realizes signal cleaning. For the problem of high-frequency noise interference in process parameters, a filtering process based on the Savitzky-Golay algorithm is designed, and its mathematical expression is:

[0035]

[0036] where represents the value of the jth data point after filtering, y j+i represents the value of the original data point, c i represents the convolution coefficient (obtained by solving polynomial fitting through the least squares method), m represents the half-width of the window, i represents the offset within the window, and the window length is 2m + 1. Compared with the traditional moving average filter, this algorithm can effectively suppress high-frequency noise while maintaining the peak characteristics of the signal, and is suitable for retaining key feature points during the crimping process. In wire harness production, the crimping force curve contains key feature points such as the initial contact point, maximum pressure point, and rebound point. The accurate extraction of these feature points is crucial for judging the crimping quality. In this embodiment, the window length is set to 7 (i.e., m = 3, equivalent to a 140ms time window in the process parameter sampling), and the polynomial order is set to 3. While retaining the fast-changing characteristics of the crimping process, high-frequency interference caused by equipment vibration is filtered out.

[0037] The second level realizes outlier detection. For the characteristic of asymmetric parameter distribution in wire harness production, the modified Z-score method is used for outlier identification, and its mathematical expression is:

[0038]

[0039] Among them, Z′ i represents the corrected Z-score value of the i-th data point, x i represents the data point value to be detected, median(X) represents the median of the data set X, and MAD represents the absolute median difference. The calculation formula is MAD = median(|x i - median(X)|), and 0.6745 is a normalization factor (to make MAD equivalent to the estimate of the standard deviation of the normal distribution). When Z′ i > 3.5, the data point is determined to be an outlier. This method shows higher detection accuracy for non-normal distribution data such as wire harness pressing force, avoiding misjudgment of the standard Z-score method for skewed distribution data.

[0040] The third level realizes feature extraction. In view of the diverse characteristics of factors affecting wire harness quality, a multi-domain feature set is extracted from the cleaned data. Basic statistical features are extracted from the time domain, including: arithmetic mean (characterizing the central tendency of parameters), standard deviation (characterizing the degree of dispersion of parameters), maximum value (characterizing extreme value characteristics), peak-to-peak value (characterizing the fluctuation range), root mean square value (characterizing the effective value), and waveform factor, peak factor, pulse factor, margin factor, and kurtosis factor (characterizing waveform characteristics). From the frequency domain perspective, the fast Fourier transform is applied to calculate the signal spectrum, and the main frequency, power spectral density, frequency center, and frequency band energy distribution are extracted to capture the periodic change characteristics of process parameters. From the statistical perspective, skewness, kurtosis, interquartile range, and sample entropy are calculated. All features are standardized by the Z-score method:

[0041]

[0042] Among them, x′ i represents the standardized feature value, x i represents the original feature value, μ represents the feature mean, and σ represents the feature standard deviation.

[0043] Through the above implementation methods, this module solves key technical problems such as multi-process parameter acquisition, noise interference filtering, outlier identification, and feature extraction in wire harness production, providing a high-quality and multi-dimensional data basis for subsequent knowledge graph construction and quality correlation analysis.

[0044] B. Wire Harness Knowledge Graph Module

[0045] The wiring harness knowledge graph module is set on the enterprise application server and is used to build a structured knowledge model for wiring harness production. This module extracts wiring harness production knowledge from the rule base, historical data, and production documents, constructs a three-layer graph structure according to the production element layer, process layer, and quality feature layer, uses the TransE algorithm and graph attention network to realize knowledge vectorization, and stores it in the Neo4j graph database; this module is connected to the wiring harness quality correlation analysis module and the wiring harness quality causal reasoning module through a knowledge interface to transmit structured knowledge.

[0046] In this embodiment, the wiring harness knowledge graph module realizes the structured representation and vectorized encoding of knowledge in the wiring harness production field. This module first extracts wiring harness production knowledge from three types of knowledge sources: the rule base (process specification documents, quality standard documents, operation procedure documents), historical data (production record database, quality inspection result table, fault case dataset), and production documents (technical manual files, design document libraries, expert experience summary documents). For structured data, SQL query statements are used to extract factual knowledge from relational databases, regular expression matching is used to extract parameter knowledge from semi-structured texts, and domain-specific rules are used to extract standard knowledge from process specifications; for unstructured texts, a sequence annotation method based on BERT is used for named entity recognition and relationship extraction, with the hidden layer dimension set to 768 and the number of attention heads set to 12. After the extracted knowledge passes the consistency check, it is converted into the RDF triple format as the basic data for graph construction.

[0047] This module uses a three-layer structure to design the wiring harness knowledge graph, systematically expressing the knowledge in the wiring harness production field. The first layer is the production element layer, which contains 5 types of nodes: (1) Raw material nodes: specifically including wire nodes, terminal nodes, sheath nodes, insulation layer nodes, connector nodes, fixing part nodes, identification part nodes, protection part nodes, sealing part nodes, shielding material nodes, strengthening material nodes, lubricating material nodes, glue nodes, soldering flux nodes, solder nodes, cleaning agent nodes, anti-corrosion material nodes, packaging material nodes; (2) Equipment nodes: specifically including wire stripping machine nodes, crimping machine nodes, tester nodes, winding machine nodes, assembly table nodes, marking machine nodes, injection molding machine nodes, ultrasonic welding machine nodes, heat shrinkage machine nodes, wire cutting machine nodes, tensile tester nodes, microscope nodes; (3) Tool nodes: specifically including crimping die nodes, cutting tools nodes, measuring tool nodes, fixing tool nodes, auxiliary tool nodes, detection tool nodes, cleaning tool nodes, adjustment tool nodes; (4) Environment nodes: specifically including temperature nodes, humidity nodes, cleanliness nodes, static electricity nodes, light nodes; (5) Personnel nodes: specifically including operator nodes, quality inspector nodes, process engineer nodes, maintenance worker nodes.

[0048] These nodes are interconnected through 10 types of relationship, namely "usage" relationship, "attached to" relationship, "contains" relationship, "compatible" relationship, "substitute" relationship, "operation" relationship, "monitoring" relationship, "influence" relationship, "maintenance" relationship and "adjustment" relationship, forming a knowledge network of the basic elements of production. For example, the "operator" node is connected to the "wire stripping machine" node through the "operation" relationship and to the "crimping die" node through the "usage" relationship, and this structure reflects the operation relationship in actual production.

[0049] The second layer is the process layer, which includes the processes, operations and inspection nodes of wire harness production. The process nodes specifically include wire stripping node, crimping node, winding node, assembly node, packaging node, testing node, marking node, fixing node, injection molding node; the operation nodes specifically include: (1) operation nodes under the wire stripping process: insulation layer removal operation node, wire stripping length adjustment operation node, wire stripping force control operation node, wire stripping speed adjustment operation node, wire stripping position positioning operation node; (2) operation nodes under the crimping process: terminal positioning operation node, crimping force control operation node, crimping temperature adjustment operation node, crimping time control operation node, crimping die replacement operation node, crimping speed adjustment operation node; (3) operation nodes under the winding process: winding turns control operation node, winding tension adjustment operation node, winding pitch control operation node; (4) operation nodes under the assembly process: wire harness positioning operation node, terminal assembly operation node, connector insertion operation node, fixing part installation operation node; (5) operation nodes under the packaging process: wire harness binding operation node, wire harness winding operation node, protective sleeve installation operation node; (6) operation nodes under the testing process: test probe connection operation node, test voltage setting operation node, test program selection operation node; (7) operation nodes under the marking process: label positioning operation node, marking information setting operation node, marking temperature control operation node; the quality inspection nodes specifically include: (1) inspection nodes under the wire stripping process: wire stripping length inspection node, insulation layer residue inspection node; (2) inspection nodes under the crimping process: crimping height inspection node, tensile test node, crimping cross-section inspection node, crimping end face inspection node; (3) inspection nodes under the winding process: winding turns inspection node, winding firmness inspection node; (4) inspection nodes under the assembly process: assembly integrity inspection node, insertion and locking inspection node; (5) inspection nodes under the testing process: continuity inspection node, insulation inspection node, resistance value inspection node; (6) inspection nodes under the packaging process: packaging integrity inspection node, appearance quality inspection node; The nodes in the process layer are connected through 6 types of relationship, namely "preceding" relationship, "succeeding" relationship, "contains" relationship, "inspection" relationship, "dependency" relationship and "substitute" relationship, which describe the timing and logical dependencies between processes.

[0050] The third layer is the quality feature layer, which contains multi-dimensional features of the wire harness quality: (1) Physical feature nodes: specifically including dimension nodes, shape nodes, strength nodes, curvature nodes, tensile force nodes, pressure nodes, weight nodes, density nodes, surface roughness nodes, hardness nodes, flexibility nodes, wear resistance nodes, and thermal expansion nodes; (2) Electrical feature nodes: specifically including resistance nodes, insulation nodes, conduction nodes, impedance nodes, capacitance nodes, signal attenuation nodes, shielding effect nodes, and crosstalk nodes; (3) Reliability feature nodes: specifically including temperature cycle nodes, vibration nodes, lifespan nodes, waterproofness nodes, corrosion resistance nodes, and anti-interference nodes; (4) Appearance feature nodes: specifically including surface nodes, identification nodes, color nodes, integrity nodes, and packaging nodes. The quality feature nodes are connected through 5 types of relationship types, namely, "composition" relationship, "influence" relationship, "correlation" relationship, "opposition" relationship, and "dependency" relationship, which describe the composition and interaction relationships between features.

[0051] The three-layer knowledge graph forms a complete knowledge network through cross-layer relationships. Nine relationship types are defined from the production factor layer to the process layer, namely, "participation" relationship, "execution" relationship, "assistance" relationship, "influence" relationship, "operation" relationship, "control" relationship, "provision" relationship, "supervision" relationship, and "constraint" relationship, which describe the effects of production factors on processes; seven relationship types are defined from the process layer to the quality feature layer, namely, "determination" relationship, "contribution" relationship, "influence" relationship, "detection" relationship, "control" relationship, "improvement" relationship, and "degradation" relationship, which describe the formation effects of processes on quality features; five relationship types are defined from the production factor layer to the quality feature layer, namely, "direct influence" relationship, "indirect influence" relationship, "determination" relationship, "constraint" relationship, and "detection" relationship, which describe the influence mechanisms of production factors on quality features.

[0052] To achieve the vector representation of knowledge, this module uses the TransE algorithm for graph embedding, mapping the entities and relationships of the knowledge graph to a low-dimensional continuous vector space. The mathematical expression is:

[0053] ‖h+r-t‖ L1 / L2

[0054] where h, h ′ and t, t ′ are the vector representations of the head entity and tail entity for positive sampling and negative sampling respectively, r is the vector representation of the relationship, and ‖·‖ L1 / L2 represents the L1 norm or L2 norm. The vector dimension is set to 128 dimensions, and the following loss function is used for model training:

[0055] L=∑ (h,r,t)∈S ∑ (h′,r,t′)∈S′max(0, γ + d(h + r, t) - d(h′ + r, t′))

[0056] where S is the set of correct triples, S ′ is the set of incorrect triples obtained by negative sampling, γ is the margin parameter (set to 1.0), and d is the L2 distance function. The training uses the mini-batch stochastic gradient descent algorithm, with a batch size of 256, an initial learning rate of 0.01, 500,000 valid triples in the training data, and 100 training epochs. This algorithm learns the vector representations of entities and relationships by minimizing the distances of correct triples and maximizing the distances of incorrect triples, achieving a computational representation of knowledge.

[0057] For the heterogeneous characteristics of nodes, this module implements a node feature extraction method based on the Graph Attention Network (GAT). The calculation formula of the attention mechanism is as follows:

[0058] e ij = a T · [W · h i ‖ W · h j

[0059]

[0060] where e ij is the attention coefficient, representing the importance of node j to node i; h i and h j are the feature vectors of nodes i and j respectively; W is the weight matrix; a is the attention vector; h′ i is the updated feature vector of node i after calculation by the graph attention layer, σ is the ReLU activation function; N i is the neighbor set of node i; ‖ represents the vector concatenation operation. The GAT network structure contains 3 graph attention layers, with a hidden layer dimension of 64, 8 attention heads, and a negative slope parameter of 0.2 for LeakyReLU. This attention mechanism can adaptively learn the importance weights of different neighbor nodes, which is particularly applicable in the wire harness knowledge graph. For example, for the "crimping" process node, the "crimping pressure" node is usually more important than the "ambient temperature" node, and GAT can automatically learn this weight difference.

[0061] This module uses the Cypher query language to implement knowledge graph queries. Query optimization includes index optimization (establishing B+ tree indexes and hash indexes for frequently queried attributes), cache mechanism (managing the query result cache using the LRU strategy, with a capacity of 2GB), and query rewriting (optimizing the paths and execution plans for complex queries). The knowledge graph is stored in the Neo4j graph database.

[0062] ​Through the above implementation manners, the wiring harness knowledge graph module realizes the systematic expression and computational representation of knowledge in the field of wiring harness production, providing a standardized knowledge basis for subsequent quality correlation analysis and causal reasoning.

[0063] C. Wiring Harness Quality Correlation Analysis Module

[0064] The wiring harness quality correlation analysis module is set on the enterprise GPU server and is used to mine the complex correlation relationships between the wiring harness production factors and quality characteristics. This module uses a multi-relational heterogeneous graph convolutional network to process the knowledge graph, identifies key influencing factors through a message passing mechanism and a graph attention mechanism, and extracts quality influence paths; this module is connected to the wiring harness quality causal reasoning module and the wiring harness quality prediction module through an analysis interface to transmit the correlation analysis results.

[0065] In this embodiment, the wiring harness quality correlation analysis module uses a multi-relational heterogeneous graph convolutional network (MR-HGCN) for correlation analysis to solve the problem of identifying complex correlations with multi-level transmission in wiring harness production. This network consists of three main structures: a node type embedding layer, a relation-aware layer, and a message passing layer, forming an end-to-end correlation analysis architecture.

[0066] The node type embedding layer designs a specific embedding strategy for heterogeneous nodes in the wiring harness knowledge graph. For each type of node t, a dedicated embedding matrix is defined where |V t | is the number of nodes of this type, d is the embedding dimension (set to 64), represents the set of real numbers. The initial node features are generated through a type-related embedding layer:

[0067]

[0068] where, is the initial feature vector of node i, x i is the one-hot encoding of node i, and E type(i) is the embedding matrix corresponding to the type of node i. According to the wiring harness production characteristics, the embedding matrices of each type of node are initialized respectively: the raw material nodes are initialized based on material characteristics, the process nodes are initialized based on timing correlation, and the quality nodes are initialized based on the hierarchical structure, realizing the unified representation of heterogeneous nodes. For example, although "wire material" and "crimping temperature" belong to different types of nodes, their representations in the embedding space can reflect their relevance in wiring harness production, providing a basis for subsequent correlation analysis.

[0069] The relation-aware layer designs dedicated transformation matrices for different relation types in wiring harness production to capture the semantic information of the relations. For each relation r ∈ R defined in the knowledge graph, a relation-specific transformation matrix is constructed When information is passed along the relation r, the features of the source node are transformed by this matrix:

[0070]

[0071] in, represents a message passed from node j to node i through relationship r, represents the feature representation of node j at layer l, W r is the transformation matrix corresponding to relation r. This design enables the network to distinguish the different roles of different relation types, such as "participation," "influence," and "determination," in the wiring harness quality process, accurately modeling relational semantics. For example, in the wiring harness crimping process, the influence of "crimping pressure" on "crimping strength" through the "determination" relation is typically higher than the influence on "conductivity" through the "influence" relation. The relational perception layer accurately captures this difference.

[0072] The message passing layer implements multi-relation message aggregation based on the attention mechanism and adaptively learns the importance weights of different relations. For node i, its l+1 layer representation is calculated as follows:

[0073]

[0074] in, represents the set of neighbor nodes connected to node i through relationship r; represents the attention weight of node j to node i through relationship r; W r Represents the transformation matrix of relation r at layer l; represents the representation vector of node j in layer l; σ is the activation function, using LeakyReLU with a negative slope of 0.2. The calculation formula of attention weight is:

[0075]

[0076] Among them, a r is the attention vector of relation r, which is learned through end-to-end training; ‖ represents the vector concatenation operation, is the feature vector of node k, and T is the transpose of the vector. The attention mechanism enables the model to adaptively assign weights based on the actual importance of different parameter relationships in wire harness production, achieving accurate association recognition.

[0077] To fuse information at different levels, this module uses residual connections and inter-layer attention mechanisms:

[0078]

[0079] Among them, β l is the attention weight of the lth layer, normalized by the softmax function:

[0080]

[0081] Among them, q1 and q2 are learnable parameters. The model is trained in an end-to-end manner, and the loss function is designed as the sum of three parts:

[0082] L = L node + λ1·L path + λ2·‖Θ‖2

[0083] Among them, L node is the node classification loss, which is used to optimize the node representation; L path is the path prediction loss, which is used to optimize the influence path recognition; ‖Θ‖2 is the L2 regularization term, which is used to control the model complexity; λ1 and λ2 are trade-off coefficients, which are set to 0.5 and 0.01 respectively. The design of the multi-objective loss function ensures that the model can optimize the node representation and path recognition simultaneously, improving the comprehensive performance of the model.

[0084] Based on the trained multi-relational heterogeneous graph convolutional network, this module implements the function of extracting the quality influence path. The influence path is defined as a directed path from the production factor node (source node) to the quality feature node (target node), which is expressed as:

[0085] P = (v0, r1, v1, r2,..., r k , v k )

[0086] Among them, v0 is the source node, v k is the target node, r i is the relationship type, and v i is the intermediate node. The path importance score is calculated based on the node importance and relationship strength:

[0087]

[0088] Among them, w(v i ) is the importance weight of the node, which is calculated by node centrality; w(r i ) is the strength weight of the relationship, which is calculated by relationship frequency; sim(v i , v i+1 ) is the similarity between node v i and v i+1 , which is calculated by vector cosine similarity; λ is the path attenuation factor, which is set to 0.8, indicating that the longer the path, the lower its importance. This scoring mechanism comprehensively considers the node importance, relationship strength and path length, and can accurately evaluate the actual influence of the path.

[0089] The path search adopts an improved bidirectional Dijkstra algorithm, and its core steps are as follows:

[0090] Initialize the forward search boundary set F s = S (source node set), and the backward search boundary set B t = T (target node set); Alternately perform forward and backward searches, and expand F s and B t ; When , find the cross nodes, construct the complete path; Score and sort the found paths according to the Score function; Return the top K paths with the highest scores (K = 5 in this embodiment). To improve the search efficiency, a pruning strategy is added to the algorithm: (1) Depth limit, the upper limit of the path length is set to 6; (2) Similarity pruning, discard the new paths whose similarity to the existing high-score paths exceeds 0.8; (3) Redundancy detection, remove the paths containing loops. These optimization strategies significantly improve the search efficiency and can complete the path search from thousands of source nodes to target nodes within 5 seconds in practical applications.

[0091] Based on the extracted impact paths, this module constructs a quality impact graph to achieve a systematic expression of the association relationship. The quality impact graph is defined as G ′ = (V ′ , E ′ , W ′ ), where V ′ is the node set, including key production factors and quality characteristics; E ′ is the edge set, representing the influence relationship; W ′ is the weight set, representing the influence intensity. Node selection is based on the importance index:

[0092] Importance(v) = α·Cen(v) + β·Cov(v) + γ·Con(v)

[0093] where Cen is the node centrality, Cov is the node coverage rate, Con is the node connectivity, and α, β, γ are weight coefficients, which are 0.5, 0.3, and 0.2 respectively. Select the top 50% of the nodes ranked by importance and include them in the impact graph. Edge construction is based on path compression to generate direct influence relationships, and the compression formula is:

[0094]

[0095] where, Let \(P\) denote the set of all paths from node \(s\) to node \(t\), \(|P|\) denote the length of path \(P\), and \(Score(P)\) denote the importance score of path \(P\). The generated initial influence graph is further simplified by threshold pruning (removing edges with weights lower than \(\theta\), where \(\theta\) is set to the 30th percentile of all edge weights) and edge merging (merging multiple relationship edges with the same source and target nodes into one comprehensive edge) to form the final quality influence graph. For example, in a certain wire harness quality analysis, the system successfully identified multiple influence paths from "crimping temperature" to "conductivity" and compressed them into one weighted edge with a weight value of 0.78, indicating that the crimping temperature has a significant impact on conductivity.

[0096] To improve the interpretability of the correlation analysis results, this module implements three types of explanation generation functions:

[0097] (1) Node explanation generation: Generate natural language explanations describing node characteristics and relationships based on the graph embedding representation and attribute information of the node; (2) Path explanation generation: Generate causal chain explanations describing the influence mechanism based on the node sequence and relationship chain of the path; (3) Influence graph visualization: Implement the visual representation of the influence graph using a force-directed layout algorithm, where the node size is proportional to the importance, the edge thickness is proportional to the influence strength, and different types of nodes and relationships are encoded using different graphical symbols and colors.

[0098] These explanation functions make the correlation analysis results easier to understand and apply. For example, for the influence path "crimping temperature → wire deformation → contact resistance → wire harness conductivity", the explanation generated by the system is: "The crimping temperature (average 280°C) affects the degree of plastic deformation of the copper wire (correlation coefficient 0.82), changing the contact resistance between the wire and the terminal (average reduction of 0.3 mΩ / 10°C), and ultimately having a significant impact on the wire harness conductivity (contribution 38%)."

[0099] Through the above implementation methods, the wire harness quality correlation analysis module solves the technical problem of difficult identification and characterization of multi-level influence relationships in the wire harness production process, providing a correlation knowledge basis for subsequent quality prediction and causal reasoning. The analysis results of this module are directly input into the wire harness quality causal reasoning module to further distinguish between correlation and causality, and at the same time input into the wire harness quality prediction module to provide feature selection and relationship weight information for quality prediction.

[0100] D. Wire Harness Quality Causal Reasoning Module

[0101] The wire harness quality causal reasoning module is set on the enterprise application server and is used to identify true causal relationships from the correlation relationships. This module processes the time-series data of production parameters, establishes a time-delay effect model, and constructs a causal relationship network through conditional independence testing and intervention verification; this module is connected to the wire harness parameter optimization module through a causal interface to transmit causal knowledge.

[0102] In this embodiment, the harness quality causal inference module designs a complete causal inference method for the time-delay effect and causal relationship identification problems in the harness production process. This method first determines the delayed impact of parameter changes on quality through time-delay effect modeling, then constructs an initial causal skeleton through conditional independence testing, and finally verifies the true causal relationship through intervention verification.

[0103] The time-delay effect modeling part solves the problem of delayed quality impact caused by parameter changes in harness production. For typical time-delay phenomena such as the fact that it takes a certain amount of time for the crimping temperature change to be reflected in the contact resistance, this module defines the time-delay effect function of parameter p on quality characteristic q at time t:

[0104]

[0105] where K is the time-delay order, set to 5; τ i is the i-th order time-delay, with the range [0, τ max , τ max is the maximum time-delay, set to 3 times the production cycle; ω i is the i-th order time-delay weight, satisfying:

[0106]

[0107] This multi-order time-delay modeling can capture the complex impact pattern of parameter changes on quality and describe the actual production situation more accurately than a single time-delay model. In harness production, the parameter impact usually shows a multi-stage delay effect. For example, the impact of crimping temperature change on contact resistance includes an immediate effect (material thermal expansion), a medium-term effect (thermal stress adjustment), and a long-term effect (material microstructure change). It is difficult for traditional single time-delay models to accurately capture this complex time-dependent relationship.

[0108] The time-delay parameter estimation adopts a time-delay parameter optimization method based on cross entropy, and the optimization objective is:

[0109]

[0110] where denotes taking the minimum value of the expression, p(q(t)|f p→q(t)) is the conditional probability distribution, λ is the regularization coefficient, set to 0.01; Ω(τ,ω) is the regularization term, and L1 regularization is used to control the model complexity. The optimization adopts an alternating iteration strategy: fix τ to optimize ω, and then fix ω to optimize τ until convergence. This method can effectively handle typical time-delay relationships such as temperature-resistance and pressure-deformation in the wire harness crimping process. In practical applications, this method successfully identifies the multi-order time-delay effect of the crimping temperature on the contact resistance. The time-delay order is 3, and the time-delay values are 10s, 60s, and 300s respectively, with corresponding weights of 0.4, 0.35, and 0.25, which highly coincides with the characteristics of the thermodynamic process in materials science.

[0111] To quantify the intensity of the time-delay effect, this module defines an index of the time-delay effect intensity:

[0112]

[0113] where, |ω i | represents the absolute value of the weight of the i-th order time delay, λ is the time decay factor, set to 0.2, indicating that the influence weight of a longer time delay is lower. This index comprehensively considers the weights of each order of time delay and the time decay factor, and can objectively reflect the overall influence intensity of the parameters on the quality. For example, the time-delay effect intensity of the crimping pressure on the contact resistance is 0.85, while the time-delay effect intensity of the ambient temperature on the contact resistance is only 0.25, indicating that the crimping pressure is a more important factor affecting the contact resistance.

[0114] Based on the time-delay effect modeling, this module uses an improved PC (Peter-Clark) algorithm to construct an initial causal skeleton. The PC algorithm starts from a fully connected undirected graph and gradually removes edges of non-direct causal relationships through conditional independence tests. For data considering the time-delay effect, the conditional independence test adopts the following steps:

[0115] Initialize a fully undirected graph, where V is all variable nodes, including all possible undirected edges; for each pair of adjacent nodes (v i ,v j ), perform a conditional independence test: the conditional set S is initially an empty set; sequentially increase the size k of the conditional set S = 0, 1, 2,...; for each conditional set of size k Calculate the conditional independence statistic; if v i and v j are conditionally independent given the conditional set S, then remove the edge from the graph and record the separation set S(v i ,v j ) = S; terminate when the size k of the conditional set exceeds the maximum threshold (in this embodiment, k max = 5) or all possible conditional sets have been tested.

[0116] For each V-structure v i -v k -v j (i.e., both v i and v j are adjacent to v k but v i and v j are not adjacent), check whether the separating set S(v i , v j ) does not contain v k : If v k is not in S(v i , v j ), then convert the undirected edges v i -v k and v k -v j to directed edges v i →v k and v j →v k ; According to the determined edge directions, propagate the direction information through the orientation rules; Repeat applying the orientation rules until no more edge directions can be determined.

[0117] The conditional independence test adopts a method based on partial correlation coefficients, calculates the partial correlation coefficient ρ(v i , v j |S) considering time lags, and converts it to a Z statistic through Fisher-Z transformation:

[0118]

[0119] where n is the sample size and |S| is the size of the conditional set. If the Z statistic is less than the threshold 1.96 (corresponding to a 95% confidence level), then v i and v j are considered conditionally independent under the given conditional set S. This method can effectively handle the conditional independence test of time-lagged data and accurately identify the direct causal relationships in wire harness production.

[0120] To integrate time-lag information into the causal skeleton, this module adopts the following strategy:

[0121] For each directed edge v i →v j constructed by the PC algorithm, query the time-lag effect modeling results to obtain the optimal time-lag order n * (v i →v j ) and the influence coefficient Calculate the time-lag influence strength S(v i →v j ); Incorporate the time-lag information (n *,{ω},S) is associated with edge v i →v j For edges whose directions cannot be determined by the PC algorithm, if there is a significant Granger causal relationship in only one direction, then that direction is determined as the direction of the causal edge; if there is a significant Granger causal relationship in both directions, then the direction with greater influence is retained; if the direction cannot be determined by time lag information, then it is retained as an undirected edge.

[0122] To verify the validity of causal relationships, this module designs a causal verification method based on virtual intervention. First, a simulation model M of the wiring harness production process is constructed, including: a parameter-quality mapping model trained based on historical data, a constraint relationship model of mutual influence between parameters, and an evolution model of quality characteristics. Then, for each causal edge v to be verified, i →v j , design intervention experiments:

[0123] Control group: keep the parameter v i The natural distribution remains unchanged, and the quality characteristics v j The response distribution P(v j );

[0124] Experimental group: parameter v i To intervene, set it to a specific value v i =v ′ , record quality characteristics v j The response distribution P(v j |do(v i =v ′ )), where do(·) represents the intervention operation; the distribution difference before and after the intervention is compared D(P(v j ),P(v j |do(v i =v ′ The intervention value design uses a stratified sampling strategy: the parameter range is divided into five equal intervals, three intervention values are randomly selected in each interval, and the experiment is repeated 10 times for each intervention value to eliminate the influence of random error. The intervention effect is evaluated using the average treatment effect (ATE):

[0125]

[0126] Where K is the total number of intervention values (in this embodiment, K=15). The formula for calculating the standardized treatment effect is:

[0127]

[0128] Among them, σ(v j) is the standard deviation of the quality characteristics. The causal relationship confirmation rule is: if nATE > 0.2, then the causal relationship is confirmed, and the causal confidence level is min(1, nATE / 0.8); if nATE ≤ 0.2, then the causal relationship is rejected, and this edge is removed from the causal graph. This method verifies the causal relationship through the actual intervention effect, avoiding the limitation of judging causality only relying on statistical correlation.

[0129] Based on the verification results, this module constructs a complete causal network, where each edge contains the following attributes: the time-delay order n(v i →v j ), the set of influence coefficients {ω0, ω1, …, ω n}, the standardized average treatment effect nATE(v i →v j ), and the causal confidence level c(v i →v j ). The edge weight calculation formula is: w(v i →v j ) = nATE(v i →v j ) · c(v i →v j ). The causal network is simplified by removing edges with weights lower than 0.1 and merging multiple consecutive edges on the path to construct a direct influence relationship. This processing makes the causal network more concise and highlights the key causal paths.

[0130] This module implements three core causal reasoning functions based on the constructed causal network: transfer chain identification, root cause analysis, and impact prediction. Transfer chain identification is used to find all causal paths from parameters to quality characteristics. The depth-first search algorithm is used to traverse the graph G to calculate the path weight:

[0131]

[0132] Among them, w(P): the weight of the path, w(e) the weight of the edge e in the path, len(P) is the path length, and λ is the path attenuation factor, set to 0.3. Sort in descending order of weight and output the top 5 paths. This function can accurately identify the causal link from process parameters to quality characteristics. For example, the system successfully identifies the causal transfer chain of "crimping temperature → wire deformation → contact resistance → conductivity", with a weight of 0.42, explaining the mechanism of how crimping temperature affects conductivity.

[0133] Root cause analysis is used to trace the source of the reasons for quality anomalies. The input is the set of abnormal quality characteristics Q ′ = {q1, q2, …, q m}, and a reverse-tracing subgraph G ′ is constructed, which contains all pairs of Q ′Nodes that have a direct or indirect impact on the middle node. Calculate the impact score for each parameter node:

[0134] Impact(p) = ∑ q∈Q′ ∑ P∈Paths(p,q) w(P)·Severity(q)

[0135] where Paths(p,q) is the set of all paths from p to q, and Severity(q) is the anomaly severity of quality characteristic q. Sort in descending order of the impact score and output the top 10 root cause parameters. This function can quickly locate the root cause of quality problems and greatly improve the problem-solving efficiency.

[0136] Impact prediction is used to predict quality changes based on parameter changes. The input is the parameter change set ΔP = {Δp1, Δp2, …, Δp n}, and calculate the expected change for each quality characteristic:

[0137]

[0138] where Δq j is the expected change amount of quality characteristic q j , k is the time step, Δp i (t - k) represents the change amount of parameter p at time t - k (t - k) represents the change amount of parameter p at time t - k i , and is the weight coefficient. The confidence interval of the prediction is [Δq j - 1.96·σ j , Δq j + 1.96·σ j , where σ j is the standard deviation of the prediction error. This function can accurately predict the impact of process parameter adjustment on quality and provide prediction support for parameter optimization.

[0139] Through the above implementation methods, the wire harness quality causal reasoning module has achieved accurate modeling and effective reasoning of complex causal relationships in wire harness production, solved the technical problems of time-delay effect and difficult identification of causal relationships in wire harness production, and provided accurate causal knowledge and reasoning ability for wire harness quality control. The causal knowledge of this module is directly input into the wire harness parameter optimization module to provide a causal basis for parameter optimization and ensure the correctness of the optimization direction.

[0140] E. Wire Harness Quality Prediction Module

[0141] The wire harness quality prediction module includes a training part set in the enterprise server and an inference part set in the edge device. This module constructs a quality prediction model using the teacher-student knowledge distillation architecture, optimizes the model and deploys it to the edge device for real-time prediction; this module is connected to the wire harness decision support module through a prediction interface to transmit the quality prediction results.

[0142] In this embodiment, the wiring harness quality prediction module adopts a distributed structure to solve the problem of wiring harness production quality prediction. The training part is deployed on an enterprise server with high-performance computing resources, and the inference part is deployed on the edge devices of the production line. The two parts are connected through the knowledge distillation technology.

[0143] The training part constructs a teacher model with a fusion network structure, which includes three sub-networks: a multi-layer perceptron (MLP) sub-network processes the structured parameter data obtained from the data acquisition module; a long short-term memory network (LSTM) sub-network processes the time-series process data; a graph convolutional network (GCN) sub-network processes the structured knowledge obtained from the wiring harness knowledge graph module. The output features of the three sub-networks are merged through an attention fusion method:

[0144]

[0145] where h i is the output feature of the i-th sub-network, and α i is the corresponding attention weight:

[0146]

[0147] e i = v T tanh(W[h i ‖h j )

[0148] where i and j represent the i-th and j-th sub-networks respectively, v and W are learnable weight vectors and matrices respectively, and T represents transpose. This attention mechanism can adaptively adjust the weights of each sub-network according to the importance of different input data, improving the adaptability and accuracy of the model. The fused features are mapped to the quality index prediction value through a fully connected layer. The teacher model is trained using the following loss function:

[0149]

[0150] where MSE is the mean squared error loss, is the L2 regularization term, λ is the regularization coefficient, set to 0.001. The training uses the Adam optimizer, the initial learning rate is 0.001, and the learning rate decay strategy is adopted. It decays to 0.9 times the original value every 10 epochs. The teacher model is trained on the enterprise server, using GPU acceleration, the batch size is 64, and the number of training rounds is 100 rounds.

[0151] The inference part constructs a lightweight student model, which only contains a simplified multi-layer perceptron network. The input is a subset of the key features identified by the teacher model, and the output is the predicted value of the quality index. The student model adopts a two-layer structure, with the hidden layer dimension of 32, using the ReLU activation function, and the output layer using a linear activation function (for regression tasks) or a Sigmoid activation function (for classification tasks).

[0152] The student model interacts and learns with the teacher model through the following loss function:

[0153] L student = α·MSE(y s , y true ) + (1 - α)·KL(σ(y s / T), σ(y t / T))

[0154] where y s and y t are the predicted outputs of the student model and the teacher model respectively, α is the balance coefficient, set to 0.5, KL is the KL divergence, σ is the softmax function, T is the temperature parameter, set to 3.0. The KL divergence term enables the student model to learn not only the correct predicted values but also the prediction distribution of the teacher model, acquiring richer knowledge. The temperature parameter T controls the "softening" degree of knowledge distillation. A higher temperature value makes the output distribution of the teacher model smoother, helping the student model learn the similarity structure between categories.

[0155] This module implements a fast adaptation method for new products, adopting a model fine-tuning strategy:

[0156]

[0157] where θ base is the initial parameter of the student model, θ new is the parameter after fine-tuning, η is the fine-tuning learning rate, set to 0.0005, and L new is the loss function based on a small number of new product samples. The fine-tuning process uses a small learning rate and a small number of iterations (usually no more than 50 rounds) to avoid overfitting the small sample data. This method enables the model to quickly adapt to new products, and only 10 - 20 samples are required to complete the adaptive learning, significantly reducing the cost and time of new product introduction.

[0158] Through the above implementation methods, the wiring harness quality prediction module realizes the collaborative work of the complex model and the lightweight model, solves the technical problem of real-time quality prediction in wiring harness production, and provides a prediction basis for subsequent decision-making support.

[0159] F. Wiring Harness Parameter Optimization Module

[0160] The wire harness parameter optimization module is set on the enterprise high-performance computing server and is used to realize the intelligent optimization of production parameters. This module constructs a virtual production environment, integrates a physical model based on the finite element method and a data-driven model based on a feedforward neural network, and uses the Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning algorithm to find the optimal parameter combination that balances quality, efficiency, and resource consumption in the virtual environment; this module is connected to the wire harness decision support module through an optimization interface to transmit the parameter optimization plan.

[0161] In this embodiment, the wire harness parameter optimization module designs a parameter optimization method based on a virtual environment for the problem of optimizing process parameters in wire harness production. This method first constructs a high-fidelity virtual environment for wire harness production, and then uses a model-based reinforcement learning algorithm to explore the optimal parameter combination in the virtual environment to achieve parameter optimization without affecting actual production.

[0162] The virtual production environment is constructed by integrating a physical model and a data-driven model to form a digital twin system of the wire harness production process. The physical model is established based on the basic principles of wire harness manufacturing, including:

[0163] Mechanical model of the wire stripping process: Describes the quantitative relationship between the wire stripping force and the wire material and geometric characteristics. The formula includes shear force, friction force, and material elastic coefficient;

[0164] Thermodynamic model of the crimping process: Describes the distribution function of the crimping temperature over time and space, and the influence function of temperature on the metal deformation characteristics;

[0165] Electrical performance model: Establishes the quantitative mapping relationship between the crimping geometric parameters (compression ratio, crimping height) and the contact resistance value.

[0166] The physical model is solved using the finite element method, and a fine grid with a grid density of 0.01 mm is set at key positions to improve the simulation accuracy. The mathematical expression of the physical model is:

[0167]

[0168] Among them, M is the mass matrix, C is the damping matrix, K is the stiffness matrix, u is the displacement vector, and F is the external force vector. The heat conduction equation is:

[0169]

[0170] Among them, ρ is the density, c pis the specific heat capacity, T is the temperature field, k is the thermal conductivity, and q is the heat source term. These equations are discretized and solved by the finite element method to simulate the mechanical and thermodynamic behaviors during the wire harness production process. During the wire harness crimping process, these physical models can accurately simulate the effects of different crimping parameters (pressure, temperature, time) on the crimping quality, providing a theoretical basis for parameter optimization.

[0171] The data-driven model adopts a feedforward neural network structure, specifically including 1 input layer, 3 fully connected hidden layers, and 1 output layer. The input layer receives process parameters, and the dimension is the same as the number of process parameters; the first hidden layer contains 128 neurons, the second hidden layer contains 64 neurons, and the third hidden layer contains 32 neurons. All hidden layers use the ReLU activation function; the number of neurons in the output layer is the same as the number of quality indicators. The linear activation function is used for continuous value indicators, and the Sigmoid activation function is used for classification indicators. The model is trained using the Adam optimizer, with a learning rate of 0.001, a batch size of 64, and the training loss function is:

[0172]

[0173] where N is the number of quality indicators, w i is the weight of each indicator, L i is the indicator-specific loss function (mean squared error for continuous indicators, cross-entropy for classification indicators), λ is the regularization coefficient value of 0.001, is the L2 regularization term. By learning the parameter-quality mapping relationship in historical production data, the data-driven model can quickly predict the impact of parameter changes on quality, making up for the deficiency of physical models in capturing complex nonlinear relationships.

[0174] The physical model and the data-driven model are integrated through a model fusion method to form a high-fidelity virtual environment:

[0175] y = α·y physics +(1 - α)·y data

[0176] where y physics is the output of the physical model, y data is the output of the data-driven model, α is the weight coefficient, with an initial value set to 0.5, and it is dynamically adjusted inversely according to the prediction errors of the two models:

[0177]

[0178] where e physics and e dataThey are the prediction errors of the physical model and the data-driven model respectively. This adaptive fusion method utilizes the theoretical basis of the physical model and the learning ability of the data-driven model, improving the accuracy and generalization ability of the virtual environment. Based on the constructed virtual environment, this module uses the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to find the optimal parameter combination. The reinforcement learning problem is defined as:

[0179] State space S: It includes 8 process parameters such as wire stripping force, wire stripping speed, wire stripping length, crimping pressure, crimping temperature, crimping time, crimping speed, and packaging tension, and 12 state variables such as wire deformation, terminal deformation, contact area, contact pressure, temperature distribution, resistance value, gravitational test value, vibration test value, temperature cycle test value, conductivity test value, insulation test value, and appearance inspection value. The dimension of the state vector is 20;

[0180] Action space A: It includes 6 adjustable process parameters such as wire stripping force, wire stripping speed, crimping pressure, crimping temperature, crimping time, and crimping speed. The dimension of the action vector is 6;

[0181] Reward function R:

[0182] R(s,a) = w1·Q(s ′ ) + w2·E(s,a) + w3·C(s,a)

[0183] where, Q(s ′ ) is the quality score in the new state s ′ The calculation formula is the weighted sum of quality indicators; E(s,a) is the efficiency score, and the calculation formula is the output per unit time; C(s,a) is the resource consumption score, and the calculation formula is the negative weighted sum of energy consumption and material loss; w1, w2, and w3 are weight coefficients, set to 0.6, 0.3, and 0.1 respectively, reflecting the optimization strategy of quality priority.

[0184] The network structure design of the TD3 algorithm is as follows:

[0185] Policy network πφ: Adopt a fully connected neural network, including 1 input layer, 2 hidden layers, and 1 output layer. The dimension of the input layer is the state space dimension 20, the first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the dimension of the output layer is the action space dimension 6. The activation function of the hidden layer is ReLU, and the activation function of the output layer is tanh (scaled by the action range).

[0186] [[ID=3X]]Double Q-value network and Adopt a similar structure. The input is the concatenation of the state and action (dimension 26), and the output is a single Q value. Use two Q networks to take the minimum value as the estimate to reduce the overestimation problem:

[0187]

[0188] Among them, γ is the discount factor, set to 0.99; θ i ′ is the parameter of the target Q network; φ ′ is the parameter of the target policy network; ∈ is the exploration noise, following a truncated normal distribution and σ is 0.2.

[0189] Soft update of the target network: Achieve slow update through Polyak averaging;

[0190] Delayed policy update: Update the policy network once every d times of updating the Q network. d is set to 2 to reduce policy oscillation.

[0191] The algorithm is trained and run for 100 iterations, and each iteration contains 10,000 environment steps. The convergence criterion is set that the reward change rate of 20 consecutive iterations is less than 1%. After training is completed, extract the optimal parameter combination from the obtained policy to form a parameter optimization plan.

[0192] This module implements the parameter sensitivity analysis function, and calculates the parameter sensitivity index through the Sobol method:

[0193]

[0194] Among them, S i is the first-order sensitivity index of parameter i, V(Y) is the total variance of the model output Y, and V i = V[E(Y|X i )] is the variance of the conditional expectation when X i is fixed. The sensitivity analysis uses 10,000 Monte Carlo samples to calculate the sensitivity index of each parameter and its 95% confidence interval. The sensitivity analysis results are used to evaluate the stability and risk of the parameter plan and assist in decision-making support. For example, the analysis shows that the sensitivity index of the crimping temperature is 0.42, significantly higher than 0.15 of the crimping time, which means that the fluctuation of the crimping temperature has a greater impact on the quality, and the temperature stability should be key controlled in production.

[0195] Through the above implementation methods, the wire harness parameter optimization module realizes the intelligent optimization of wire harness production parameters, solves the technical problems of low efficiency and high cost of the traditional trial-and-error method, and provides a scientific parameter configuration plan for wire harness production.

[0196] G. Wire Harness Decision Support Module

[0197] The wire harness decision support module includes a calculation part set in the enterprise server and a display part set in the operation terminal. This module uses the Non-dominated Sorting Genetic Algorithm (NSGA-II) to find a parameter solution that balances quality, efficiency, cost, and energy consumption, evaluates the risk of the solution based on the parameter sensitivity index, generates multi-dimensional radar charts and parallel coordinate charts to visualize the solution characteristics, and generates a decision report. This module interacts with production management personnel through the user interface to provide decision support information.

[0198] In this embodiment, the wire harness decision support module adopts a multi-objective optimization method to generate a decision solution for the multi-objective balance problem in wire harness production. This module receives the outputs of the wire harness quality prediction module and the wire harness parameter optimization module, and provides optimized parameter configuration suggestions for production management personnel by weighing four key indicators: quality, efficiency, cost, and energy consumption.

[0199] This module formulates the wire harness production decision problem as an optimization problem in a four-dimensional objective space:

[0200] max{Q(x),E(x),-C(x),-P(x)}

[0201] Where max{·} represents taking the maximum value of the expression, x is the process parameter vector, which includes 6 parameters: stripping force, stripping speed, crimping pressure, crimping temperature, crimping time, and crimping speed; Q(x) is the quality score function, defined as the weighted sum of four key quality indicators: conduction rate, insulation performance, tensile test, and vibration test:

[0202] Q(x)=w1·q1(x)+w2·q2(x)+w3·q3(x)+w4·q4(x)

[0203] E(x) is the efficiency function, defined as the number of products produced per unit time; C(x) is the cost function, defined as the sum of the material cost, labor cost, equipment cost, and energy cost per product; P(x) is the energy consumption function, defined as the total energy consumption for producing each product.

[0204] This module uses the Non-dominated Sorting Genetic Algorithm (NSGA-II) to solve the above multi-objective optimization problem. NSGA-II is an efficient multi-objective optimization algorithm that can generate a set of Pareto optimal solutions in a single run. Each solution corresponds to a parameter configuration plan, representing different trade-offs among the four objectives. The main steps of the algorithm are as follows:

[0205] Initialization: Randomly generate the population P0 with a population size set to 100; Non-dominated sorting: Stratify and sort the population according to the Pareto dominance relationship; Crowding degree calculation: Calculate the crowding degree of each individual within the same non-dominated layer, and preferentially select individuals with a larger crowding degree to maintain diversity; Parent selection: Adopt the binary tournament selection strategy to select parent individuals based on non-dominated sorting and crowding degree; Mutation and crossover: Apply simulated binary crossover (SBX) and polynomial mutation operations to generate offspring, with a crossover probability of 0.9 and a mutation probability of 0.1; Elite retention: Combine the parents and offspring, perform non-dominated sorting and crowding degree calculation, and retain the optimal individuals equal to the population size for the next generation; Termination condition: Reach the maximum number of iterations (set to 100) or the change rate of the hypervolume index of the optimal solution set is less than 0.1% for 10 consecutive generations.

[0206] The NSGA-II algorithm outputs a set of Pareto optimal solutions to form a decision-making solution set, and each solution represents different trade-offs in four objective dimensions. For example, solution A may focus more on quality, while solution B focuses more on efficiency. Both are non-dominated optimal solutions, and the decision-maker selects according to actual needs.

[0207] To assist production management personnel in making decisions, this module evaluates and demonstrates the optimization solutions in the following ways:

[0208] Comprehensive scoring: Calculate the comprehensive score of the solution through fixed weights, and the formula is:

[0209] Score(x) = 0.5·Q norm (x) + 0.3·E norm (x) - 0.1·C norm (x) - 0.1·P norm (x)

[0210] where the subscript norm represents normalization processing, mapping each index value to the interval [0, 1]. The weight configuration reflects the quality-first strategy and can be adjusted according to enterprise needs. The system arranges the solutions in descending order of the comprehensive score, by default recommends the solution ranked first, and at the same time provides other high-score solutions for selection and reference.

[0211] Risk assessment: Based on the results of parameter sensitivity analysis, calculate the stability index S of the solution:

[0212]

[0213] where S i is the sensitivity index of parameter i, and R i is the fluctuation range ratio of parameter i, and the calculation formula is

[0214] R i = Δx i / x i,max-x i,min ,Δx i is the actual fluctuation range of parameter i, and x i,max and x i,min are the allowable maximum and minimum values of parameter i. According to the stability index value, the risk of the solution in this module is divided into three levels: when S < 0.2, it is a low risk, indicating that the parameter fluctuation has little impact on the quality; when 0.2 ≤ S < 0.5, it is a medium risk, indicating that the parameter fluctuation will have a certain impact on the quality; when S ≥ 0.5, it is a high risk, indicating that the parameter fluctuation will significantly affect the quality stability. The risk assessment results are displayed in color coding (green, yellow, red) on the decision interface, intuitively reflecting the risk level of the solution.

[0215] Visualization display: Through multi-dimensional radar charts and parallel coordinate charts, present the performance of different solutions in each target dimension, supporting the intuitive comparison of solutions. The radar chart adopts a four-axis design, representing the four target dimensions of quality, efficiency, cost, and energy consumption respectively. The scales on each axis are normalized, enabling the comparison of indicators with different dimensions in the same chart. The parallel coordinate chart arranges multiple target dimensions as parallel coordinate axes, and each solution is represented as a broken line connecting each axis, effectively showing the relationships and trends between high-dimensional data.

[0216] Decision report generation: The system automatically generates a structured decision report, which includes the following main contents:

[0217] Solution overview: Briefly describe the parameter configuration and expected effects of the recommended solution; Multi-objective comparison: Display the performance of the solution in each target dimension through tables and charts; Risk analysis: Explain the stability index and potential risk factors of the solution; Implementation suggestions: Provide specific steps for parameter adjustment and monitoring key points; Sensitivity analysis: Show the degree of influence of parameter fluctuation on quality, guiding the control of the production process.

[0218] Through the above implementation methods, the wiring harness decision support module realizes the generation of a parameter solution with multi-objective balance, provides scientific and intuitive decision support for wiring harness production management, and effectively solves the technical problem of difficult balance among quality, efficiency, cost, and energy consumption in wiring harness production.

[0219] Based on the output of the decision support module, four working modes are realized:

[0220] 1. Quality monitoring mode: The wiring harness data acquisition module regularly collects production data, and the wiring harness quality prediction module uses the optimized edge deployment model to predict real-time quality indicators; when the predicted value deviates from the preset target range, the wiring harness decision support module generates a warning message and transmits it to the operation terminal; this mode enables the system to identify potential risks before actual quality problems occur, continuously monitor the production status, and ensure the quality stability of the wiring harness.

[0221] 2. Problem Diagnosis Mode: Automatically starts when the system detects quality anomalies or receives a manual diagnosis request; the wire harness quality correlation analysis module identifies the impact paths related to the problem based on the multi-relationship heterogeneous graph convolutional network; the wire harness quality causal reasoning module applies conditional independence testing and intervention verification techniques to analyze the root causes and rank them according to the probability of occurrence; the wire harness decision support module generates a structured root cause analysis report and specific solution suggestions based on the analysis results; this mode provides technicians with accurate problem positioning and solutions, significantly shortening the problem-solving cycle.

[0222] 3. Parameter Optimization Mode: Starts in response to new product introduction or process improvement requirements; the wire harness parameter optimization module systematically explores the parameter space in a virtual production environment, and evaluates the effects of parameter combinations by combining physical models and data-driven models; the system generates a Pareto optimal solution set through the non-dominated sorting genetic algorithm and evaluates the performance of each solution in dimensions such as quality, efficiency, and cost; the wire harness decision support module generates final optimization suggestions based on the Pareto optimal solution set and risk assessment results; this mode realizes the scientific adjustment of wire harness production parameters, effectively improving product quality and production efficiency.

[0223] 4. Knowledge Accumulation Mode: Automatically starts regularly by the system or when new knowledge patterns are identified; the wire harness knowledge graph module extracts new knowledge from incremental production data and updates the three-layer graph structure; the wire harness quality prediction module and the wire harness parameter optimization module adjust the model parameters according to the new data; the system continuously improves the knowledge base and algorithm models through continuous learning, realizing the continuous improvement of quality control capabilities; this mode enables the system to automatically accumulate and inherit knowledge in the field of wire harness production, reducing the dependence on expert experience.

[0224] Combined Figure 2 , the working process of the system of the present invention includes the following seven steps, and each step is connected through a standardized data interface to form a complete quality control closed loop:

[0225] Step 1: Collect and preprocess wire harness production data

[0226] The wire harness data collection module collects the whole process data of wire harness production through industrial communication protocols, including raw material parameters, process parameters, image data, and electrical test data. Perform noise filtering, anomaly detection, and feature extraction on the collected data to form a standardized feature data set. Classify and store the processed data in the database for subsequent knowledge graph construction and correlation analysis.

[0227] Step 2: Construct a wire harness production knowledge graph

[0228] The wire harness knowledge graph module extracts wire harness production knowledge from the rule base, historical data, and production documents, and constructs a three-layer graph structure according to the production element layer, process layer, and quality feature layer. Intra-layer relationships and inter-layer relationships are established to form a complete knowledge network. Through graph embedding technology, knowledge is vectorized and stored in the graph database, providing a structured knowledge foundation for subsequent correlation analysis and causal reasoning.

[0229] Step 3: Analyze the wire harness quality correlation

[0230] The wire harness quality correlation analysis module uses graph neural network technology to analyze the complex correlations between wire harness production elements and quality features, and converts the node relationships in the knowledge graph into computational influence paths. Extract the key influence paths from production elements to quality features, evaluate the importance of the paths, and construct a quality influence graph, providing a correlation knowledge foundation for subsequent causal reasoning and quality prediction.

[0231] Step 4: Perform wire harness quality causal reasoning

[0232] The wire harness quality causal reasoning module processes the time-series data of production parameters, establishes a time-delay effect model, and represents the delayed impact of parameter changes on quality. An initial causal skeleton is constructed through conditional independence testing, and the true causal relationship is verified through intervention. A complete causal network is constructed to achieve functions such as transfer chain identification, root cause analysis, and impact prediction, providing a causal basis for parameter optimization.

[0233] Step 5: Predict the wire harness production quality

[0234] The wire harness quality prediction module constructs a quality prediction model using a knowledge distillation architecture, trains a complex teacher model on the enterprise server, migrates the knowledge to a lightweight student model, and deploys it to edge devices. A rapid adaptation mechanism for new products is achieved, and the deployment efficiency is improved through model optimization, providing real-time prediction support for quality monitoring.

[0235] Step 6: Optimize the wire harness production parameters

[0236] The wire harness parameter optimization module constructs a virtual production environment, integrates physical models and data-driven models to simulate the wire harness production process. The state space and action space are defined, and a reward function that comprehensively considers quality, efficiency, and resource consumption is designed. A reinforcement learning algorithm is used to find the optimal parameter combination, and the parameter sensitivity is analyzed, providing a parameter optimization solution for decision support.

[0237] Step 7: Provide wire harness production decision support

[0238] The wire harness decision support module formulates the parameter optimization problem as a multi-objective optimization problem, taking into account four objective dimensions: quality, efficiency, cost, and energy consumption. It generates a Pareto optimal solution set, evaluates the risk of the solution, visualizes the solution characteristics, and generates a decision report. It provides optimization suggestions to production managers, and the execution results are fed back to the data acquisition module to form a complete decision-execution-feedback closed loop.

[0239] The above seven steps are selectively activated according to different system operation modes: the quality monitoring mode activates steps 1, 5, and 7; the problem diagnosis mode activates steps 1, 3, 4, and 7; the parameter optimization mode activates steps 1, 6, and 7; the knowledge accumulation mode activates steps 1, 2, 5, and 6.

[0240] In summary, the present invention uses multiple artificial intelligence technologies such as knowledge graphs, graph neural networks, causal reasoning, and reinforcement learning to construct an artificial intelligence-based wire harness production quality control system. By combining data-driven and knowledge-guided methods, it identifies quality associations, infers causal relationships, predicts quality indicators, and optimizes production parameters, providing a comprehensive quality control solution for the wire harness manufacturing industry. This system can be widely applied to the production processes of various wire harness products such as automotive wire harnesses, household appliance wire harnesses, and industrial equipment wire harnesses, and has significant technical value and broad application prospects.

[0241] Although this document demonstrates the specific applications of the present invention through specific embodiments, these embodiments are for illustrative purposes only and do not imply a limitation on the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims, and any appropriate modifications, equivalent substitutions, or improvements based on the principles of the present invention shall be considered as falling within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be given the broadest interpretation to cover all such modifications, equivalent structures, and functions.

Claims

1. A wire harness production quality control system based on artificial intelligence, characterized in that, Including: A wire harness data acquisition module, which is used to collect raw material parameters, process parameters, image data, and electrical test data through industrial communication protocols, and perform data cleaning and feature extraction; A wire harness knowledge graph module, which is used to construct a three-layer graph structure including a production factor layer, a process layer, and a quality feature layer, and use the TransE algorithm and graph attention network to realize knowledge vectorization; A wire harness quality correlation analysis module, which is used to process the knowledge graph through a multi-relational heterogeneous graph convolutional network, identify key influencing factors through a node type embedding layer, a relationship perception layer, and a message passing layer, extract quality influence paths, and mine the complex correlation relationships between wire harness production factors and quality features; A wire harness quality causal reasoning module, which is used to establish a multi-order time-delay effect model, construct an initial causal skeleton through conditional independence testing, verify causal relationships through virtual intervention, and identify true causal relationships from correlation relationships; A wire harness quality prediction module, which is used to construct a quality prediction model through a teacher-student knowledge distillation architecture, where the teacher model integrates three sub-networks: a multi-layer perceptron, a long short-term memory network, and a graph convolutional network, and the student model is deployed on edge devices to predict the production quality of wire harnesses in real time; A wire harness parameter optimization module, which is used to construct a virtual production environment integrating physical models and data-driven models, find the optimal parameter combination that balances quality, efficiency, and resource consumption through reinforcement learning algorithms, and perform intelligent optimization of production parameters; A wire harness decision support module, which is used to find a parameter solution that balances quality, efficiency, cost, and energy consumption through a multi-objective optimization algorithm, evaluate the risk of the solution, and provide decision support information; Among them, the data of the wire harness data acquisition module flows to the wire harness knowledge graph module and the wire harness quality correlation analysis module, the knowledge of the wire harness knowledge graph module flows to the wire harness quality correlation analysis module and the wire harness quality causal reasoning module, the analysis results of the wire harness quality correlation analysis module flow to the wire harness quality causal reasoning module and the wire harness quality prediction module, the reasoning results of the wire harness quality causal reasoning module flow to the wire harness parameter optimization module, the prediction results of the wire harness quality prediction module and the optimization solutions of the wire harness parameter optimization module flow to the wire harness decision support module, and the decision-making suggestions of the wire harness decision support module guide production adjustments and trigger new data acquisitions, forming a closed-loop optimization.

2. The system according to claim 1, characterized in that, The wire harness data acquisition module is connected to high-precision devices such as wire stripping machines and crimping machines through the PROFINET protocol, connected to high-speed detection devices through the EtherCAT protocol, and connected to PLCs and test devices through the OPC UA protocol, forming a data acquisition network covering the entire process; data cleaning and anomaly detection are performed based on the Savitzky-Golay algorithm and the modified Z-score method, and time-domain, frequency-domain, and statistical features are extracted and standardized through the Z-score method.

3. The system according to claim 1, wherein The production factor layer in the wire harness knowledge graph module includes five types of nodes: raw materials, equipment, tools, environment, and personnel; the process layer includes three types of nodes: processes, operations, and inspections; The quality feature layer contains four types of nodes: physical features, electrical features, reliability features, and appearance features. The complete knowledge network is formed by connecting these layers through cross-layer relationships and stored in the Neo4j graph database.

4. The system according to claim 1, wherein The multi-relational heterogeneous graph convolutional network of the wire harness quality correlation analysis module includes: A node type embedding layer that designs a dedicated embedding matrix for different types of nodes; A relation perception layer that designs a dedicated transformation matrix for different relation types; A message passing layer that realizes multi-relational message aggregation based on the attention mechanism; The module uses an improved bidirectional Dijkstra algorithm to extract the quality influence path, which includes path length limitation, similarity pruning, and redundancy detection strategies.

5. The system according to claim 1, wherein The multi-order time-delay effect model established by the wire harness quality causal inference module contains multiple time-delay orders and corresponding time-delay weights, which are used to quantify the different time-delay effects of parameter changes on quality. The time-delay order is set to K = 5, and the maximum time-delay is set to 3 times the production cycle. An initial causal skeleton is constructed through conditional independence tests based on partial correlation coefficients, the causal relationship is verified by calculating the average treatment effect, and a complete causal network is constructed based on the verification results.

6. The system according to claim 1, wherein In the wire harness quality prediction module, the teacher model adopts the following structure: A multi-layer perceptron sub-network processes structured parameter data; A long short-term memory network sub-network processes time-series process data; A graph convolutional network sub-network processes structured knowledge; The output features of the three sub-networks are combined through an attention fusion method; The knowledge distillation architecture realizes the transfer of knowledge from the teacher model to the student model through a composite loss function adjusted by the temperature parameter. After quantization optimization, the student model is deployed to edge devices.

7. The system according to claim 1, wherein The virtual production environment of the wire harness parameter optimization module includes: A mechanical model of the wire stripping process that describes the relationship between the wire stripping force and the wire material and geometric characteristics; A thermodynamic model of the crimping process that describes the distribution of the crimping temperature over time and space; An electrical performance model that establishes the mapping relationship between the crimping geometric parameters and the contact resistance value; The output of the physical model and the data-driven model are integrated through a model fusion method. Combining with the Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning algorithm, the state space, action space, and reward function are defined to find the optimal parameter combination.

8. The system according to claim 1, characterized in that, The wire harness decision support module formulates the wire harness production decision problem as a multi-objective optimization problem including four dimensions: quality score, efficiency, cost, and energy consumption. The non-dominated sorting genetic algorithm is used to generate the Pareto optimal solution set. The comprehensive score of the solution is calculated through fixed weights. The risk of the solution is evaluated based on the parameter sensitivity index and divided into three levels: low risk, medium risk, and high risk. The characteristics of the solution are visualized through multi-dimensional radar charts and parallel coordinate charts.

9. The system according to claim 1, wherein The system supports four working modes: The quality monitoring mode. The wire harness data acquisition module regularly collects production data. The wire harness quality prediction module uses the optimized edge deployment model to predict real-time quality indicators. When the predicted value deviates from the preset target range, the wire harness decision support module generates a warning message; Problem diagnosis mode, which is automatically started when the system detects quality anomalies or receives a manual diagnosis request. The wire harness quality correlation analysis module identifies the impact paths related to the problem, and the wire harness quality causal reasoning module analyzes the root cause by applying conditional independence testing and intervention verification techniques; Parameter optimization mode, which is started for new product introduction or process improvement requirements. The wire harness parameter optimization module systematically explores the parameter space in a virtual production environment, and the wire harness decision support module generates the final optimization suggestions based on the Pareto optimal solution set; Knowledge accumulation mode, which is automatically started by the system regularly or when new knowledge patterns are identified. The wire harness knowledge graph module extracts new knowledge from incremental production data and updates the three-layer graph structure. The quality prediction module and the parameter optimization module adjust the model parameters according to the new data.

Images