Door cover assembly failure mode root cause analysis method based on graph similarity matching

Abstract

The application discloses a door cover assembly failure mode root cause analysis method based on graph similarity matching, and belongs to the technical field of automobile manufacturing and quality management. The method comprises the following steps: collecting multi-source data in the whole process of stamping, welding, coating and final assembly, and performing layered anomaly detection; constructing a layered knowledge graph comprising material, process parameters, intermediate and final quality nodes based on process mechanism, and learning the causal weight between nodes by using a graph attention network; when failure occurs, an abnormal propagation subgraph is separated from the graph, which is encoded into an embedded vector by using a graph convolution network, and similarity matching and weighted voting are performed with a historical case library to quickly infer the most possible root cause. The application realizes automatic and accurate root cause tracing of door cover assembly failure of cross-plant and multi-coupling factors, significantly shortens the analysis cycle, and reduces the dependence on personal experience.

Description

Technical Field

[0001] This invention belongs to the field of automotive manufacturing and quality management technology, specifically a root cause analysis method for door and hood assembly failure modes based on graph similarity matching. Background Technology

[0002] As competition in the new energy vehicle market intensifies, the market has transitioned from a "blue ocean" to a "red ocean," making product quality a key factor in determining brand reputation and customer repurchase rates. Door panels (such as car doors, hoods, and trunk lids) are among the most frequently touched body assembly components, and their assembly quality directly impacts the overall vehicle's appearance, sealing performance, opening and closing feel, and noise control. Common door panel assembly defects include excessive gap differences, abnormal door closing noises, poor sealing, and inconsistent opening and closing force. These problems are easily perceived by users and can lead to questions about the brand's manufacturing processes.

[0003] Currently, when faced with quality failures during door assembly, engineers typically follow the traditional analytical path of "problem identification - line stoppage and isolation - manual investigation - expert consultation - trial and error verification - root cause identification" to trace the source. However, attributing failures in door assembly faces unique and severe challenges: First, its quality formation chain spans four major process workshops: stamping, welding, painting, and final assembly, involving dozens of key control points such as the precision of the body-in-white door frame, the dimensions of the door panel assembly, the installation of hinges / locks / limiters, and the application of sealing strips. The data sources are highly dispersed and heterogeneous. Second, the adjustment processes between each process link have a high degree of coupling and nonlinear effects. For example, a small positioning deviation in the welding workshop may be amplified by thermal deformation after painting and baking, ultimately manifesting as a serious gap surface difference problem in the final assembly stage. This complex causal transmission relationship makes single, localized data analysis ineffective.

[0004] The existing traditional root cause analysis model has the following limitations: (1) Long analysis cycle: It relies heavily on manual item-by-item investigation and expert consultation, resulting in high trial and error costs and failing to meet the rapid response requirements of high-speed production lines; (2) Static attribution logic: The analysis process is often based on fixed empirical rules or isolated statistical results, making it difficult to dynamically capture and quantify the complex interactions between multiple links and parameters; (3) Fragmented data value: Data from each workshop and system is stored in isolation outside the process, lacking the ability to trace and integrate the entire process chain from stamping, welding, painting to final assembly. Therefore, traditional methods are no longer suitable for the current stringent requirements of high quality, high efficiency and high consistency in automobile manufacturing.

[0005] In summary, there is an urgent need in this field for an analytical method that can deeply integrate manufacturing process mechanisms, effectively correlate multi-source data throughout the entire process, and intelligently and quickly locate the root cause, in order to solve the problem of tracing the source of quality failures in the specific scenario of door and cover assembly. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a root cause analysis method for door and cover assembly failure modes based on graph similarity matching. By constructing a knowledge graph that integrates multi-source data and process mechanisms from stamping, welding, painting, and final assembly, and using graph neural networks for rapid similarity matching and reasoning, the method can achieve automated and accurate root cause localization of door and cover assembly failures, thereby significantly improving analysis efficiency and quality consistency.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A root cause analysis method for failure modes of door and cover assembly based on graph similarity matching includes the following steps: Step 1: Data Processing: Collect control point data of the vehicle throughout the entire process in the stamping workshop, welding workshop, painting workshop and final assembly workshop, and store the data in association with the vehicle identification number (VIN); Step 2: Anomaly Identification: A hierarchical detection method is used to perform real-time anomaly identification on the control point data in order to identify and mark abnormal data; Step 3: Knowledge Graph Construction: A root cause analysis knowledge graph is constructed based on historical failure data. The knowledge graph includes a node system consisting of material node layers, process parameter node layers, intermediate quality node layers, and final quality node layers, as well as edges representing causal relationships between nodes based on process mechanisms. The weights of the edges are learned through a graph attention network. The knowledge graph is stored in a graph database and associated with historical failure labels, historical root cause labels, and historical remedial measures. Step 4: Root Cause Analysis: For the currently manufactured vehicle, extract the anomaly propagation graph from the knowledge graph based on its abnormal data; use a graph convolutional network encoder to encode the anomaly propagation graph into a fixed-dimensional graph embedding vector; calculate the cosine similarity between the graph embedding vector and the graph embedding vectors of all historical cases in the knowledge base; retrieve the top K historical cases with the highest similarity, and determine the most likely root cause of the current failure through weighted voting based on their root cause labels.

[0008] Furthermore, in step one, a hierarchical storage architecture is adopted for data management, and the method is as follows: The system accesses the data streams of equipment in each workshop in real time via message middleware and stores high-frequency sensor data in a memory cache layer. Use a time-series database to store recent time-series data; Use a relational database to store structured data on material batches, personnel operations, and inspection results; Archive historical data that has exceeded the specified period to the data lake.

[0009] Furthermore, in step two, the layered detection method includes: First layer: Rule engine, which sets hard threshold rules for known failure modes and provides millisecond-level responses; The second layer: statistical process control, using X-bar / R-EWMA control charts to detect mean drift and process capacity decay; The third layer: unsupervised machine learning, which uses isolated forests to detect anomalies in high-dimensional parameter combinations and an autoencoder to detect reconstruction errors in time-series waveforms; The fourth layer is a deep temporal model, which uses a Transformer-based anomaly detection model to identify contextual anomalies and cumulative effects in long-term temporal dependent scenarios.

[0010] Furthermore, in step three, the method of learning edge weights using a graph attention network is as follows: Failed samples in historical data are marked as positive samples, and qualified samples are marked as negative samples; Construct a GAT network with binary cross-entropy as the loss function, and train it to accurately predict the out-of-range labels of the final quality nodes. After training, the attention weights of each layer of GAT are extracted as the causal influence strength of the edges, where the nodes... For nodes Attention weights The calculation formula is: in: For nodes For nodes The original attention score; For nodes For nodes The weight of causal influence; and These are nodes and nodes eigenvectors; The weight matrix is ​​a learnable matrix; This is the attention parameter vector; Indicates a splicing operation; Let v represent the set of incoming edge neighbors of node v.

[0011] Furthermore, in step three, the storage structure of historical cases in the knowledge base is as follows: in: For the first Anomalous propagation diagram of a historical case; For the first Records of handling measures for each historical case; For the first Expiration tags for historical cases; For the first Root cause labeling of historical case failures; For the first The vehicle identification number associated with each historical case; For the first The timestamp of each historical case.

[0012] Furthermore, in step four, the method of encoding the anomaly propagation graph into a graph embedding vector using a graph convolutional network encoder is as follows: The node features are updated through L-layer graph convolution. The formula for updating the node features of a layer is: in: Adjacency matrix for adding self-loops; for The degree matrix; It is the first The learnable weight matrix of the layer; It is the ReLU activation function; After L layers of graph convolution, attention-weighted pooling is used for graph-level readout to obtain the graph embedding vector. : Where: attention weight The calculation formula is: in: and These are learnable parameters; and To ensure that after propagation through L layers in a Graph Convolutional Network (GCN), the nodes and nodes The final layer feature vector; It is a set of nodes.

[0013] Furthermore, a multi-scale feature fusion approach is used to generate graph embedding vectors: in: For the first Embedding after layer graph convolution.

[0014] Furthermore, the method for encoding the anomaly propagation graph into a fixed-dimensional graph embedding vector is as follows: in: To extract anomaly propagation graphs from knowledge graphs; The cosine similarity between the graph embedding vector and the graph embedding vectors of all historical cases in the knowledge base is calculated: in: For the first Pre-computed graph embedding vectors for each historical case.

[0015] Furthermore, in step four, the top K most similar historical cases retrieved are used... : The most likely root cause of the current failure is determined by weighted voting. The formula is: in: For the first The root cause label of a historical case failure.

[0016] Furthermore, the confidence level of the root cause inference results is calculated: When the confidence level is lower than a preset threshold, the current case is pushed to the manual confirmation queue, and the knowledge base is updated after expert confirmation.

[0017] The beneficial effects of this invention are as follows: This invention, based on graph similarity matching, provides a root cause analysis method for failure modes in door and cover assembly. Through the synergistic effect of four key steps, it produces significant technical benefits: First, by collecting and linking data from the entire process of stamping, welding, painting, and final assembly, a unified quality traceability chain was constructed, breaking down the traditional data silos between workshops and laying a data foundation for achieving root cause tracing across processes. Secondly, a hierarchical anomaly detection method was adopted, which integrates rule-based, statistical, unsupervised learning and deep learning, to accurately identify progressive, complex and contextual anomalies in multi-source heterogeneous data, providing high-quality anomaly event input for subsequent analysis; Furthermore, a hierarchical knowledge graph (including materials, parameters, intermediate and final quality nodes) reflecting the unique process mechanism of door and cover assembly was constructed, and causal edge weights were learned using a graph attention network (GAT). This enables the graph to not only statically describe process relationships, but also dynamically quantify the influence intensity of different factors on the final quality, forming a computable and reasonable domain knowledge model. Finally, a graph convolutional network (GCN) is used to encode the real-time anomaly propagation graph into a vector, and similarity matching and weighted voting are performed with historical cases to automate and intelligently infer root cause analysis. This method transforms the traditional lengthy investigation process that relies on expert experience into a second-level computation based on data and models, greatly improving the efficiency and accuracy of analysis, and can handle novel unknown failures through a confidence assessment mechanism. Attached Figure Description

[0018] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart of an embodiment of the root cause analysis method for door and cover assembly failure modes based on graph similarity matching of the present invention; Figure 2 A schematic diagram of a tiered data storage architecture; Figure 3 Define a causal relationship edge graph for the scene with the gap in the right front door; Figure 4 This is a schematic diagram illustrating the process of extracting the anomaly propagation graph. Figure 5 This is a schematic diagram illustrating the principle of graph similarity calculation. Figure 6 This is a schematic diagram of the GCN neural network structure. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0020] like Figure 1 As shown in the figure, the root cause analysis method for failure modes of door and cover assembly based on graph similarity matching in this embodiment includes the following steps.

[0021] Step 1: Data Processing: Collect control point data of the vehicle throughout the entire process of stamping workshop, welding workshop, painting workshop and final assembly workshop, and store the data in association with the vehicle identification number (VIN).

[0022] (1) Data collection Specifically, the assembly quality of the door cover is the cumulative result of stamping dimensional accuracy, welding positioning accuracy, coating film thickness uniformity, and final assembly adjustment processes. Therefore, it is necessary to collect data from all control points in stamping, welding, painting, and final assembly to trace the true root cause of failure. In this embodiment, data from all control points in the four major workshops of stamping, welding, painting, and final assembly are collected, specifically including the following:

[0023] Stamping Workshop: Collect equipment operation logs, sensor data, process parameters, material information, and personnel operation records, covering data such as press tonnage, sheet thickness, die temperature, forming force curve, and blank holder force. Specifically, the control points in the stamping workshop mainly include the sheet metal incoming inspection station and the stamping forming station. The data to be collected includes: equipment operation logs (press tonnage, slide stroke, SPM cycle time, and downtime records), sensor data (sheet thickness measurements, die temperature, forming force curve, and drawbead resistance), process parameters (blank holder force, air cushion pressure, lubricant application amount, and stamping speed), material information (steel coil batch number, material grade, supplier code, and mechanical property test report), and personnel operation records (die change time and operator's employee number, and first and last piece inspection results).

[0024] Welding workshop: Data collection includes welding robot trajectory data, welding current and voltage waveforms, weld nugget diameter, door frame opening dimensions, and hinge mounting hole position accuracy. Specifically, control points in the welding workshop are concentrated at the door assembly station, hinge mounting point welding station, and the body-in-white online measurement station. Data collection requires the following: equipment operation logs (welding robot trajectory data, welding torch electrode grinding times, fixture cylinder operation cycles), sensor data (real-time welding current and voltage waveforms, ultrasonic weld nugget diameter, door frame opening dimensions and hinge mounting hole position accuracy measured online by a coordinate measuring machine or laser), process parameters (welding time, electrode pressure, preheating pulse parameters, adhesive application trajectory and amount), material information (welding wire batch number, hinge assembly part number and supplier traceability code), and personnel operation records (records of manual repair welding of non-standard weld points, and dimensional deviation adjustment handling forms).

[0025] In the painting workshop: data such as electrophoretic film thickness, baking temperature curve, sealant coating parameters, and bath liquid circulation pump status are collected. In the final assembly workshop: data such as tightening torque angle curve, gap surface difference measurement, door closing force, and locking sound pressure level are collected. The control points in the painting workshop are located in the pretreatment electrophoresis line, sealant coating station, and topcoat baking section. It is necessary to collect equipment operation logs (battery liquid circulation pump operation status, oven chain speed, robot coating trajectory deviation alarm), sensor data (electrophoretic film thickness measurement, temperature curves of each section of the baking oven, wet film thickness and sagging detection), process parameters (electrophoretic voltage-time curve, PVC sealant temperature and viscosity, spray atomization pressure and fan width), material information (electrophoretic paint batch number, sealant expiration date and opening time, paint formula number), and personnel operation records (paint film defect sanding and repair records, manual touch-up records).

[0026] The control points in the final assembly workshop are distributed at the door and cover sub-assembly line, assembly and adjustment station, and final inspection gap and surface difference measurement station. It is necessary to collect equipment operation logs (torque and angle curves of electric tightening guns, positioning accuracy of assisted robotic arms, and delivery time of AGVs), sensor data (tightening torque and angle values ​​of hinge bolts, optical gap and surface difference measurement values, door closing force and locking sound pressure level), process parameters (step-by-step torque setting values ​​of bolt tightening strategy, operating current threshold of window regulator, and assembly pressing force of sealing strips), material information (batch and heat treatment furnace number of hinge bolts, vulcanization date of sealing strips, and serial number of window regulator motors), and personnel operation records (thickness and number of adjustments of gap and surface difference adjustment shims, troubleshooting work orders for abnormal door closing noise, and rework station work content and responsible persons).

[0027] The data from each workshop is collected in real time through the MES system and linked to the vehicle's VN code to form a complete traceability chain for door and hood assembly quality.

[0028] (2) Tiered storage architecture This addresses the aforementioned industrial data scenarios characterized by multi-source heterogeneity, time-intensive processing, and the need to support full-process traceability of stamping, welding, and coating. For example... Figure 2 The implementation example shown adopts a hierarchical data storage architecture to balance the low latency requirements of real-time monitoring with the long-link traceability needs of root cause analysis across workshops. The hierarchical storage architecture includes a real-time layer, a time-series layer, a relational layer, an analysis layer, and a unified service layer. The data management method using this hierarchical storage architecture is as follows: real-time access to data streams from equipment in each workshop is achieved through a message middleware, and high-frequency sensor data is stored in a memory cache layer; a time-series database is used to store recent time-series data; a relational database is used to store structured data on material batches, personnel operations, and inspection results; and historical data exceeding a specified period is archived to a data lake.

[0029] Specifically, in this embodiment, Apache Kafka is used as a message middleware to access the data streams of equipment in each workshop in real time. Downstream, Redis is used as a memory cache layer to store high-frequency sensor data (such as welding current waveforms and tightening torque curves) from the past 24 hours, which is used by the monitoring model for real-time anomaly detection and online inference. The read latency is controlled at the millisecond level.

[0030] Real-time layer (hot data): Uses a time series database to store time series data for the past 30 days, indexed by three levels of tags: workshop-workstation-equipment. It supports the model to quickly aggregate and query the fluctuation trend of process parameters within a specific time window, meeting the time correlation backtracking requirements of root cause analysis.

[0031] Time series layer (warm data): Fast aggregation query of parameter fluctuation trends to meet the time correlation backtracking needs in root cause analysis.

[0032] Relationship layer (structured data): PostgreSQL is used to store structured data such as material batch information, personnel operation records, and inspection results. Foreign key associations are established between VIN codes and time series data, supporting full-process data join queries by vehicle dimension.

[0033] Analysis layer (cold data): Historical data exceeding 30 days is archived to a data lake (such as MinIO based on Parquet format), partitioned by workshop and bucketed by date for offline model training and long-term quality trend analysis, and read in batches through the Spark engine.

[0034] Unified Service Layer: A data service API gateway is deployed on the upper layer to encapsulate the joint query logic across storage engines. The model side pulls the associated data of the vehicle in the entire process of stamping, welding and painting by VIN code through a unified interface, shielding the heterogeneity of the underlying storage.

[0035] Step 2: Anomaly Identification: A hierarchical detection method is used to identify anomalies in the control point data in real time, so as to identify and mark abnormal data.

[0036] Anomaly data is identified and labeled to construct node vectors for a knowledge graph. This embodiment employs a hierarchical detection method for real-time anomaly identification of multi-source heterogeneous data from the stamping, welding, and coating industries, including the following:

[0037] First layer: Rule engine (millisecond-level response), sets hard threshold rules for known failure modes to provide millisecond-level responses.

[0038] Hard threshold rules are set for known failure modes to trigger immediate alarms. This is suitable for safety-related parameters and parameters with clearly defined process boundaries. It is commonly used for single-point threshold data such as press tonnage, blank holder force, welding current and voltage, electrode pressure, tightening torque and angle, drying oven chain speed, electrophoresis voltage, and AGV arrival time. The detection targets are safety boundaries and known failure modes.

[0039] The second layer is statistical process control (SPC) with second-level response, which uses X-bar / R-EWMA control charts to detect mean drift and process capability decay.

[0040] X-bar / R-EWMA control charts are established for key quality characteristics to detect mean drift and process capability Cpk decay, and to identify progressive anomaly trends. The X-bar chart monitors changes in group means to detect whether the process center of gravity has drifted, while the R-bar chart monitors changes in subgroup ranges to detect whether process variability has changed. The EWMA (Exponentially Weighted Moving Average) control chart assigns exponentially decaying weights to historical data, making it more sensitive to small mean drifts than the X-bar chart, and particularly suitable for detecting progressive anomalies. It is commonly used for quality characteristic data such as sheet thickness, solder joint weld nugget diameter, frame opening size, electrophoretic film thickness, gap surface difference, closing force, and sealing strip indentation force, with the detection targets being mean drift and process capability decay.

[0041] The third layer is unsupervised machine learning (second-level response), which uses isolated forests to detect anomalies in high-dimensional parameter combinations and autoencoders to detect reconstruction errors in time-series waveforms.

[0042] For high-dimensional process parameter combinations, an Isolation Forest algorithm is used. For time-series waveform data, an autoencoder-based reconstruction error detection method is employed; if the reconstruction loss exceeds a threshold, the waveform is identified as abnormal. This method is commonly used for waveform / high-dimensional data such as forming force curves, welding current waveforms, tightening torque-angle curves, electrophoresis voltage curves, and multi-parameter combinations. The detection target is waveform anomalies.

[0043] The fourth layer is a deep temporal model (sub-second response), which uses a Transformer-based anomaly detection model to identify contextual anomalies and cumulative effects in long-term temporal dependent scenarios.

[0044] For long-term time-dependent scenarios, a Transformer-based anomaly detection model is employed. This model learns the temporal distribution of normal production patterns and detects contextual and aggregate anomalies. It is commonly used for long-term data such as oven temperature waste curves, electrode grinding accumulation, mold temperature trends, and adjustment / rework frequency trends. The detection targets are contextual anomalies and cumulative effects.

[0045] Step 3: Knowledge Graph Construction: Construct a root cause analysis knowledge graph based on historical failure data. The knowledge graph includes a node system consisting of a material node layer, a process parameter node layer, an intermediate quality node layer, and a final quality node layer, as well as edges representing causal relationships between nodes based on process mechanisms. The weights of the edges are learned through a graph attention network. The knowledge graph is stored in a graph database and associated with historical failure labels, historical root cause labels, and historical handling measures.

[0046] Specifically, this embodiment constructs a root cause analysis knowledge graph of door cover failures and their process anomalies based on historical data. This graph is used for subsequent similarity comparison in graph similarity-based root cause analysis to find the root causes of unknown failures and anomalies. The specific implementation process of constructing the fault knowledge graph with the right front door clearance exceeding tolerance as the target fault mode is as follows.

[0047] (1): Knowledge graph node system construction: Define material node layer, process parameter node layer, intermediate quality node layer, and final quality node layer to form a hierarchical and classified node system.

[0048] Based on the process flow and quality transfer mechanism of door assembly, a node system for a fault knowledge graph is defined, which adopts a hierarchical classification structure. The specific nodes for the right front door gap scenario in this implementation example are as follows.

[0049] Material node layer example: N_steel steel coil batch node, including mean thickness, standard deviation of thickness, yield strength, elongation, supplier code, etc.

[0050] Examples of process parameter node layers: N_stamp_param stamping process node, including blank holder force, air cushion pressure, stamping speed, mold temperature, etc.; N_weld_param welding process node, including welding current, voltage, time, electrode pressure, number of electrode grinding cycles, etc.; N_bake_param baking process node, including temperature settings for each section, chain speed, etc.

[0051] Intermediate quality node layer examples: N_stamp_dim stamping dimension node, containing key dimension measurements of the inner door panel; N_fame_dim door frame opening node, containing the door frame opening width, height, and diagonal dimensions; N_hinge_pos hinge hole position node, containing the three-dimensional coordinate deviation of the upper and lower hinge mounting holes; N_bake_deform baking deformation node, feature vector dimension 4, containing the dimensional changes before and after baking.

[0052] Final quality node layer example: N_gap door gap node, which includes the top, bottom, front, and rear gap measurements and out-of-tolerance labels.

[0053] All of the above feature values ​​correspond to an anomaly label value stored in the node, serving as one of the elements of the node's feature vector.

[0054] (2) Definition of causal edge: Define the causal relationship edges between nodes based on the process mechanism, such as Figure 2 As shown.

[0055] (3) Edge weight learning: The edge weights are learned by using graph attention network (GAT) and the causal influence strength between nodes is calculated through the attention mechanism.

[0056] The Graph Attention Network (GAT) is used to learn edge weights. The training process is as follows: Samples with out-of-range gaps in historical data are labeled as positive samples, and samples with acceptable gaps are labeled as negative samples. A three-layer GAT network is constructed, with hidden layer dimensions of 64, 32, and 16, respectively. Binary cross-entropy is used as the loss function, and the training objective is to accurately predict the out-of-range labels of the N_gap nodes. After training, the attention weights of each layer of the GAT are extracted as the causal influence strength of the edges. The formula for calculating edge weights based on the attention mechanism is: in: For nodes For nodes The original attention score; For nodes For nodes The weight of causal influence; and These are nodes and nodes eigenvectors; The weight matrix is ​​a learnable matrix; This is the attention parameter vector; Indicates a splicing operation; Let v represent the set of incoming edge neighbors of node v.

[0057] (4) Graph storage: Store the knowledge graph in the graph database, establish a VIN code index, extract the anomaly propagation graph from the causal subgraph along the abnormal data, and store it in association with failure labels, root cause labels, and disposal measures.

[0058] The constructed knowledge graph is stored in the Neo4j graph database, with node and edge attributes stored in JSON format. A VN code index is built to support millisecond-level retrieval of causal subgraphs. For historical data on door failures, possible failure modes are categorized, and the categorization set is as follows: ,in This is the f-th invalid label. The classification set is... , Classify the root causes for r fault labels. Extract the anomaly propagation graph from the cause-effect subgraph along the anomaly data, as follows: Figure 4 As shown. The storage structure of the Neo4j graph database is as follows: in: For the first Anomalous propagation diagram of a historical case; For the first Records of handling measures for each historical case; For the first Expiration tags for historical cases; For the first Root cause labeling of historical case failures; For the first The vehicle identification number associated with each historical case; For the first The timestamp of each historical case.

[0059] Step 4: Root Cause Analysis: For the currently manufactured vehicle, extract the anomaly propagation graph from the knowledge graph based on its abnormal data; use a graph convolutional network encoder to encode the anomaly propagation graph into a fixed-dimensional graph embedding vector; calculate the cosine similarity between the graph embedding vector and the graph embedding vectors of all historical cases in the knowledge base; retrieve the top K historical cases with the highest similarity, and determine the most likely root cause of the current failure through weighted voting based on their root cause labels. Specifically, in this embodiment, the graph similarity matching method for fast root cause inference includes the following: Graph encoding: The graph of the current vehicle's abnormal propagation is encoded into a fixed-dimensional embedding vector using a graph convolutional network (GCN), and attention-weighted pooling and multi-scale feature fusion are used to extract graph structure information; Similarity calculation: The cosine similarity between the current graph embedding and all historical cases in the knowledge base is calculated; Root cause inference: The top K historical cases with the highest similarity are retrieved, and weighted voting is used to determine the most likely root cause of the current failure; Confidence assessment: The confidence of the root cause inference is calculated. When the confidence is lower than the threshold, the case is pushed to the manual confirmation queue, and the knowledge base is updated after expert confirmation.

[0060] like Figure 5 As shown, when a door cover failure occurs in a vehicle under construction, similarity matching is used to find the anomaly propagation map with the highest similarity to the anomaly propagation map of the vehicle under construction from the Neo4j graph database. To support efficient similarity matching, a graph convolutional neural network is used to encode the anomaly propagation map into a fixed-dimensional embedding vector. A graph convolutional network (GCN) is then used to extract features from the anomaly propagation map. The GCN neural network results are shown below. Figure 6 As shown. Among them, the first The formula for updating the node features of a layer is: in: Adjacency matrix for adding self-loops; for The degree matrix; It is the first The learnable weight matrix of the layer; It is the ReLU activation function; This is the initial node feature matrix.

[0061] After L layers of graph convolution, graph-level readout operations are used to aggregate node features into graph embedding vectors: in: It is a set of nodes.

[0062] Attention-weighted pooling is used for graph-level readout to obtain graph embedding vectors. : Where: attention weight The calculation formula is: in: and These are learnable parameters; and To ensure that after propagation through L layers in a Graph Convolutional Network (GCN), the nodes and nodes The final layer feature vector.

[0063] To capture graph structure information at different levels, a multi-scale feature fusion approach is used to generate graph embedding vectors: in: For the first Embedding after layer graph convolution.

[0064] Furthermore, the anomaly propagation graph of the current vehicle is matched with the historical failure root cause knowledge graph to find the most similar historical cases, thereby quickly inferring the root cause. A graph embedding vector is calculated for the current anomaly propagation graph of the in-service vehicle, that is, the anomaly propagation graph is encoded as a fixed-dimensional graph embedding vector. in: To extract anomaly propagation graphs from knowledge graphs.

[0065] The cosine similarity between the graph embedding vector and the graph embedding vectors of all historical cases in the knowledge base is calculated: in: For the first Pre-computed graph embedding vectors for each historical case.

[0066] The top K most similar historical cases retrieved : Based on the root cause labels of Top-K similar cases, the most likely root cause of the current failure is determined through weighted voting. : in: For the first The root cause label of a historical case failure.

[0067] Furthermore, the confidence level of the root cause inference results is calculated: When the confidence level is lower than a preset threshold, the current case is pushed to the manual confirmation queue. After expert confirmation, the knowledge base is updated. In this embodiment, the preset threshold is 90%. That is, when Confidence < 90%, it indicates that the current failure case may belong to a novel failure mode not fully covered in the knowledge base, and the current case needs to be pushed to the manual confirmation queue. If the root cause analysis results of the manual analysis are inconsistent with the algorithm analysis results, domain experts will confirm and label the root causes and add them to the historical failure root cause knowledge graph database.

[0068] Compared with the prior art, the beneficial effects of this embodiment are as follows.

[0069] (1) By using a graph similarity algorithm, the abnormal data of vehicles currently under construction are intelligently matched with the abnormal propagation graph in the historical database. This method enables rapid root cause analysis of typical problems such as doors and hoods, significantly reducing the time cost of failure tracing. This not only helps to shorten the R&D cycle of new models, but also directly improves the debugging and production efficiency of mass production of models.

[0070] (2) A data collection and storage platform built on the industrial brain systematically gathers multi-dimensional process data related to doors and hoods. This platform provides a data foundation for in-depth analysis of the correlation between process parameters and assembly quality, thereby effectively supporting process optimization and continuously improving the appearance quality and manufacturing consistency of vehicles.

[0071] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A root cause analysis method for failure modes of door and cover assembly based on graph similarity matching, characterized in that: Includes the following steps: Step 1: Data Processing: Collect control point data of the vehicle throughout the entire process in the stamping workshop, welding workshop, painting workshop and final assembly workshop, and store the data in association with the vehicle identification number (VIN); Step 2: Anomaly Identification: A hierarchical detection method is used to perform real-time anomaly identification on the control point data in order to identify and mark abnormal data; Step 3: Knowledge Graph Construction: A root cause analysis knowledge graph is constructed based on historical failure data. The knowledge graph includes a node system consisting of material node layers, process parameter node layers, intermediate quality node layers, and final quality node layers, as well as edges representing causal relationships between nodes based on process mechanisms. The weights of the edges are learned through a graph attention network. The knowledge graph is stored in a graph database and associated with historical failure labels, historical root cause labels, and historical remedial measures. Step 4: Root Cause Analysis: For the currently manufactured vehicle, extract the anomaly propagation graph from the knowledge graph based on its abnormal data; use a graph convolutional network encoder to encode the anomaly propagation graph into a fixed-dimensional graph embedding vector; calculate the cosine similarity between the graph embedding vector and the graph embedding vectors of all historical cases in the knowledge base; retrieve the top K historical cases with the highest similarity, and determine the most likely root cause of the current failure through weighted voting based on their root cause labels.

2. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 1, characterized in that: In step one, a hierarchical storage architecture is used for data management, and the method is as follows: The system accesses the data streams of equipment in each workshop in real time via message middleware and stores high-frequency sensor data in a memory cache layer. Use a time-series database to store recent time-series data; Use a relational database to store structured data on material batches, personnel operations, and inspection results; Archive historical data that has exceeded the specified period to the data lake.

3. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 1, characterized in that: In step two, the layered detection method includes: First layer: Rule engine, which sets hard threshold rules for known failure modes and provides millisecond-level responses; The second layer: statistical process control, using X-bar / R-EWMA control charts to detect mean drift and process capacity decay; The third layer: unsupervised machine learning, which uses isolated forests to detect anomalies in high-dimensional parameter combinations and an autoencoder to detect reconstruction errors in time-series waveforms; The fourth layer is a deep temporal model, which uses a Transformer-based anomaly detection model to identify contextual anomalies and cumulative effects in long-term temporal dependent scenarios.

4. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 1, characterized in that: In step three, the method of learning edge weights using a graph attention network is as follows: Failed samples in historical data are marked as positive samples, and qualified samples are marked as negative samples; Construct a GAT network with binary cross-entropy as the loss function, and train it to accurately predict the out-of-range labels of the final quality nodes. After training, the attention weights of each layer of GAT are extracted as the causal influence strength of the edges, where the nodes... For nodes Attention weights The calculation formula is: in: For nodes For nodes The original attention score; For nodes For nodes The weight of causal influence; and These are nodes and nodes eigenvectors; The weight matrix is ​​a learnable matrix; This is the attention parameter vector; Indicates a splicing operation; Let v represent the set of incoming edge neighbors of node v.

5. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 1, characterized in that: In step three, the storage structure of historical cases in the knowledge base is as follows: in: For the first Anomalous propagation diagram of a historical case; For the first Records of handling measures for each historical case; For the first Expiration tags for historical cases; For the first Root cause labeling of historical case failures; For the first The vehicle identification number associated with each historical case; For the first The timestamp of each historical case.

6. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to any one of claims 1-5, characterized in that: In step four, the method for encoding the anomaly propagation graph into a graph embedding vector using a graph convolutional network encoder is as follows: The node features are updated through L-layer graph convolution. The formula for updating the node features of a layer is: in: Adjacency matrix for adding self-loops; for The degree matrix; It is the first The learnable weight matrix of the layer; It is the ReLU activation function; After L layers of graph convolution, attention-weighted pooling is used for graph-level readout to obtain the graph embedding vector. : Where: attention weight The calculation formula is: in: and These are learnable parameters; and To ensure that after propagation through L layers in a Graph Convolutional Network (GCN), the nodes and nodes The final layer feature vector; It is a set of nodes.

7. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 6, characterized in that: Graph embedding vectors are generated using a multi-scale feature fusion approach. in: For the first Embedding after layer graph convolution.

8. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 7, characterized in that: The method for encoding anomaly propagation graphs into fixed-dimensional graph embedding vectors is as follows: in: To extract anomaly propagation graphs from knowledge graphs; The cosine similarity between the graph embedding vector and the graph embedding vectors of all historical cases in the knowledge base is calculated: in: For the first Pre-computed graph embedding vectors for each historical case.

9. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 8, characterized in that: In step four, the top K most similar historical cases retrieved are used... : The most likely root cause of the current failure is determined by weighted voting. The formula is: in: For the first The root cause label of a historical case failure.

10. The root cause analysis method for door and cover assembly failure modes based on graph similarity matching according to claim 9, characterized in that: Calculate the confidence level of the root cause inference results: When the confidence level is lower than a preset threshold, the current case is pushed to the manual confirmation queue, and the knowledge base is updated after expert confirmation.

Images