Soft package battery production process multi-process abnormal intelligent detection method and system

Abstract

The application discloses a kind of soft package battery production process multi-process abnormality intelligent detection method and system, the method includes, by tracing code to the same cell in multiple production processes Process parameters are aligned, and multi-process parameter vector is constructed;The partial correlation coefficient matrix between adjacent processes is calculated in sliding window, and the current window correlation structure diagram is constructed in combination with each process parameter statistic;Current window correlation structure diagram and reference correlation structure diagram established based on historical normal production data are input into graph change detection network, respectively encoded to obtain node embedding and calculate the difference vector of corresponding node, to obtain batch level abnormality score and process level abnormality contribution score;According to batch level abnormality score, multi-process linkage abnormality is judged, and according to process level abnormality contribution score, the abnormal source process is positioned.The application can detect the abnormal change of the correlation between multiple processes, and realize the abnormal source process positioning.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing and production process quality inspection technology for lithium-ion batteries, specifically to an intelligent detection method and system for multiple abnormalities in the production process of soft-pack batteries. Background Technology

[0002] The production process of soft-pack lithium-ion batteries involves multiple continuous steps, including coating, rolling, slitting, stacking, tab welding, aluminum-plastic film encapsulation, and electrolyte injection. The processing results of each step are not only affected by the equipment status and process parameters of that step, but also continue to be passed on as input to subsequent steps. Therefore, there are complex parameter correlations and quality transfer relationships between different steps. With the increasing automation and digitalization of battery production lines, production equipment can typically collect process parameters such as temperature, pressure, speed, dimensions, and vacuum level, and combine this with product traceability information for quality monitoring. Currently, anomalies are commonly identified on the production floor using methods such as process upper and lower limits, statistical process control, and final inspection. Related data analysis methods are also increasingly being used for process quality prediction and anomaly detection. These methods are effective in identifying situations where a single parameter deviates significantly from its set range or where a single step experiences significant fluctuations.

[0003] However, due to changes in raw material batches, gradual equipment wear, or slow shifts in the production environment, the parameters of each process may still be within the allowable range, while the original relationships between processes have changed. Such anomalies are characterized by cross-process transmission and gradual accumulation, making them difficult to detect in a timely manner when judged based on a single parameter or single process. Furthermore, it is difficult to accurately identify the source process of the anomaly after it occurs, affecting the efficiency of quality traceability and anomaly investigation in the production process. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this manual provides a method and system for intelligent detection of anomalies in multiple processes during the production of pouch batteries.

[0005] According to a first aspect of the embodiments of this application, a method for intelligent detection of multi-process anomalies in the production process of pouch batteries is provided, including: By using traceability codes, process parameters collected from multiple production processes of the same battery cell are aligned according to the process sequence to construct a multi-process parameter vector at the battery cell level. Within the sliding window, the partial correlation coefficient matrix between adjacent processes is calculated for the multi-process parameter vectors of multiple cells, and the direct correlation strength between each pair of adjacent processes after controlling the parameters of the remaining processes is obtained. Using the partial correlation coefficient matrix as edge features and the parameter statistics of each process within the window as node features, a correlation structure graph of the current window is constructed. The correlation structure graph uses each process as a node and the material flow direction between adjacent processes as directed edges. The association structure graph of the current window and the benchmark association structure graph established based on historical normal production data are input into the graph change detection network. The graph change detection network encodes each node embedding in the two graphs respectively, calculates the difference vector of the corresponding node embedding, and outputs the batch-level anomaly score and the process-level anomaly contribution score based on the difference vector. The batch-level anomaly score determines whether there is a multi-process linkage anomaly in the current window, and the process source of the anomaly is located based on the process-level anomaly contribution score.

[0006] As an optional approach, the calculation of the partial correlation coefficient matrix between adjacent processes includes: calculating the covariance matrix of the multi-process parameter vectors of all cells within the sliding window; adding a regularization term to the covariance matrix and then inverting it to obtain the precision matrix; extracting sub-block elements corresponding to each pair of adjacent process parameters from the precision matrix; calculating the partial correlation coefficient between each pair of adjacent process parameters based on the sub-block elements and forming the partial correlation coefficient matrix.

[0007] As an optional approach, the node features are vectors composed of the mean and standard deviation of each parameter of each process within the sliding window, and the edge features are vectors flattened from the partial correlation coefficient matrix between corresponding adjacent processes. Each node feature and each edge feature is mapped to a unified embedding dimension through a process-specific linear projection layer, and then used as the input features of the association structure graph.

[0008] As an optional approach, the encoder of the graph change detection network is a graph attention network containing multiple graph attention layers; each graph attention layer adopts an asymmetric attention mechanism when calculating the attention weights between nodes: for information transmitted from downstream process nodes to upstream process nodes, the attention weights are multiplied by a decay factor to suppress it; for information transmitted from upstream process nodes to downstream process nodes, the attention weights remain unchanged.

[0009] As an optional approach, the attention coefficients of each attention head in the graph attention layer are calculated by concatenating the query transformation result of the target node embedding, the key transformation result of the source node embedding, and the edge transformation result of the corresponding edge embedding, and then applying the attention vector and activation function. The graph attention network contains two graph attention layers. The first layer uses multi-head attention, and the outputs of each head are concatenated as the input of the second layer. The second layer uses single-head attention to output the final node embedding. The attenuation factor is determined by searching on the validation set based on the process-level anomaly localization accuracy.

[0010] As an optional approach, the step of outputting batch-level anomaly scores and process-level anomaly contribution scores based on the difference vectors includes: performing global average pooling on the difference vectors of all nodes, followed by a fully connected layer and Sigmoid activation to output the batch-level anomaly scores; passing the difference vectors of each node through a fully connected layer with shared parameters to output the original contribution scores of each process; and normalizing the original contribution scores of each process to obtain the process-level anomaly contribution scores.

[0011] As an optional approach, the graph change detection network constructs training data in the following manner: normal samples are derived from historical batch data that has passed final inspection, and corresponding association structure graphs are generated according to the sliding window; abnormal samples are generated by recalculating the association structure graph after applying a systematic offset to the parameters of selected processes in the normal batch data, wherein the magnitude of the systematic offset is within the tolerance range of the corresponding parameters, and the process contribution label of the abnormal samples is marked as the target process to which the offset is applied.

[0012] As an optional approach, the systematic offset includes at least one of the following modes: linear trend offset simulating progressive wear of equipment, step offset simulating batch switching of raw materials, and periodic fluctuation offset simulating periodic vibration of equipment.

[0013] As an optional approach, the training of the graph change detection network employs a composite loss function, which comprises a weighted sum of three terms: a detection loss term for the deviation between the batch-level anomaly score and the true label; a localization loss term for the deviation between the predicted distribution of the process-level anomaly contribution score and the true process contribution distribution; and a regularization loss term for penalizing the portion of the batch-level anomaly score in normal samples that exceeds the tolerance margin. The weight coefficients of each term are selected on the validation set using a grid search based on anomaly detection evaluation metrics.

[0014] According to a second aspect of the embodiments of this application, a multi-process anomaly intelligent detection system for the production of pouch batteries is also provided, comprising: The multi-process parameter construction module is used to align process parameters collected from multiple production processes of the same battery cell according to the process sequence through traceability codes, and construct a battery cell-level multi-process parameter vector. The correlation strength calculation module is used to calculate the partial correlation coefficient matrix between adjacent processes for the multi-process parameter vectors of multiple cells within a sliding window, so as to obtain the direct correlation strength of each pair of adjacent processes after controlling the parameters of the remaining processes. The association structure graph construction module is used to construct the association structure graph of the current window by using the partial correlation coefficient matrix as edge features and the parameter statistics of each process within the window as node features. The association structure graph uses each process as a node and the material flow direction between adjacent processes as directed edges. The graph change detection module is used to input the association structure graph of the current window and the benchmark association structure graph established based on historical normal production data into the graph change detection network. The graph change detection network encodes the two graphs to obtain the embedding of each node, calculates the difference vector of the corresponding node embedding, and outputs the batch-level anomaly score and the process-level anomaly contribution score based on the difference vector. The anomaly determination and location module is used to determine whether there is a multi-process linkage anomaly in the current window based on the batch-level anomaly score, and to locate the source process of the anomaly based on the process-level anomaly contribution score.

[0015] This application aligns process parameters of the same battery cell across multiple production stages and uses partial correlation to characterize the direct relationship between parameters of adjacent stages after excluding the influence of other parameters. Simultaneously, it constructs a correlation structure graph that reflects both the state of the stage itself and the inter-stage relationships. By comparing the changes in the correlation structure between the current production window and the normal production baseline, it achieves a holistic assessment of the production status of multiple stages. This approach is no longer limited to judging whether a single process parameter exceeds the tolerance range; therefore, it can identify linked anomalies caused by factors such as changes in raw materials and progressive equipment wear, where individual parameters are still within the allowable range but the inter-stage relationships have changed. Furthermore, by constraining the direction of attention propagation in the graph based on material flow, it reduces misjudgments of the source caused by the downstream state's reverse influence on the upstream node representation. It also utilizes node differences to simultaneously obtain batch-level anomaly scores and stage-level anomaly contribution scores, unifying anomaly detection and source location into a unified process. This improves the ability to identify and trace hidden cross-stage anomalies in the production of pouch batteries. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this specification and, together with the description, serve to explain the principles of this specification.

[0017] Figure 1 This is a flowchart of a multi-process abnormality intelligent detection method for the production process of soft-pack batteries according to one embodiment of this application.

[0018] Figure 2 This is a schematic diagram of multi-process associated structural change detection in one embodiment of this application.

[0019] Figure 3 This is a schematic diagram of asymmetric graph attention propagation in one embodiment of this application.

[0020] Figure 4 This is a schematic block diagram of a multi-process abnormality intelligent detection system for the production of soft-pack batteries, as described in this application.

[0021] Figure 5This is a structural diagram of an electronic device according to one embodiment of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] This embodiment targets a mass production line for soft-pack lithium-ion batteries used in household 3C cleaning devices. The production line encompasses seven key processes, including coating, rolling, slitting, stacking, tab welding, aluminum-plastic film encapsulation, and electrolyte injection. Each process is equipped with an online parameter acquisition device. Each battery cell carries a unique traceability code throughout the entire process, and production is organized by batch. This method is deployed on the production line's host computer or edge computing device to perform testing after electrolyte injection and before formation. The production line cycle time is 2 to 5 cells per minute, and the testing cycle must match this cycle time.

[0024] The implementation process of the method described in this application will be explained in detail below with reference to specific embodiments.

[0025] Please see Figure 1 , Figure 1 This is a flowchart of a multi-process anomaly intelligent detection method for the production of soft-pack batteries according to one embodiment of this application, such as... Figure 1 As shown, the method includes steps 101-105: In step 101: The process parameters collected from multiple production processes of the same battery cell are aligned according to the process sequence by means of the traceability code, and a battery cell-level multi-process parameter vector is constructed.

[0026] According to an embodiment of this disclosure, the online detection system for each process writes the collected data into the process database after the cell processing is completed, and associates it with the cell's traceability code. In this embodiment, all process parameters of the same cell are extracted from each process database according to the traceability code, and then concatenated into a complete vector according to the process sequence.

[0027] In some embodiments, the key parameters collected for the seven processes are as follows: For the coating process, four parameters are collected: areal density, coating speed, oven temperature, and slurry viscosity; for the rolling process, three parameters are collected: rolling pressure, electrode thickness, and springback rate; for the slitting process, three parameters are collected: cutting speed, burr height, and width deviation; for the lamination process, four parameters are collected: X-direction alignment deviation, Y-direction alignment deviation, lamination tension, and interlayer spacing; for the tab welding process, three parameters are collected: welding current, welding resistance, and tensile strength; for the aluminum-plastic film encapsulation process, three parameters are collected: encapsulation temperature, sealing width, and sealing strength; and for the liquid injection process, three parameters are collected: liquid injection volume, vacuum degree, and settling time. All parameters are scalar values ​​and can be directly output by online sensors or testing instruments used in the processes.

[0028] Let the first The first cell in the first The parameter vector of the process is ,in The sequence is 4, 3, 3, 4, 3, 3, 3, which, when concatenated, form a multi-process parameter vector. Total dimensions Before stitching, each parameter was standardized using z-scores: the average value of each parameter in historical normal batches was used. and standard deviation Based on the standard, the standardized value is ,in These are the original parameter values. and All are dimensionless statistical baseline quantities, which are recalculated synchronously when the baseline correlation structure diagram is updated.

[0029] Step 102: Calculate the partial correlation coefficient matrix between adjacent processes within a sliding window. In some embodiments, the data to be analyzed is organized using a sliding window, with the window size... For example, the number of battery cells can be set to 50, and the window step size can be set to 10 cells. Whenever the electrolyte injection process of 10 new battery cells is completed, the window slides forward one step to include the latest 10 battery cells and remove the oldest 10 battery cells.

[0030] Specifically, within the current window A data matrix is ​​composed of standardized multi-process parameter vectors for each battery cell. Each row corresponds to 23-dimensional parameters of a battery cell.

[0031] calculate Covariance matrix between columns To avoid the matrix becoming nearly singular when the window size is finite, a regularization term is added to obtain... , For example, take 0.01, It is a 23rd order identity matrix. Then... Inverse the matrix to obtain the precision matrix .

[0032] Extract partial correlation coefficients from the accuracy matrix; parameters With parameters The partial correlation coefficient between them, that is, the direct correlation strength between them after controlling for all other parameters, is calculated as follows: ; in For the precision matrix Line number Column elements, and diagonal elements The range of values ​​is The difference between partial correlation coefficient and simple correlation coefficient lies in the fact that the latter reflects the overall linear relationship between two variables, which may be mixed with spurious associations indirectly caused by third-party variables; the former eliminates the indirect influence of all other variables through the accuracy matrix, retaining only the direct linear relationship between the two parameters that cannot be explained by other parameters. In multi-process scenarios, for example, coating surface density and thickness after rolling may show a simple correlation because they are both affected by the slurry solid content, but the partial correlation between the two after controlling the slurry viscosity parameter reflects the direct interaction between the coating and rolling processes.

[0033] From the complete 23×23 dimensional partial correlation coefficient matrix, sub-blocks corresponding to the parameter positions of each pair of adjacent processes are extracted to form 6 adjacent process partial correlation sub-matrices: , , , , , .by For example, its first Line number The column elements are the coating process number. The parameters and the first step of the rolling process The partial correlation coefficients between the parameters.

[0034] As an example, after the above process is used to calculate the values ​​of 50 battery cells within a certain window... The partial correlation coefficient between the surface density of the intermediate coating and the thickness of the electrode after roll forming is 0.72, indicating that there is still a strong positive direct correlation between the two after controlling for the other 21 parameters. This is consistent with physical expectations, that is, electrodes with higher surface density usually have higher thickness after roll forming. However, the partial correlation coefficient between the coating oven temperature and the thickness after roll forming is only 0.05, indicating that there is no significant direct correlation between the two.

[0035] Step 103: Construct the association structure graph of the current window. Using the partial correlation coefficient matrix as edge features and the parameter statistics of each process within the window as node features, construct the association structure graph of the current window. The association structure graph uses each process as a node and the material flow direction between adjacent processes as directed edges.

[0036] Specifically, the aforementioned association structure is represented as a directed graph. The graph contains 7 nodes, each corresponding to a process; it also contains 6 directed edges, each pointing from an upstream process node to an adjacent downstream process node, with the direction consistent with the material processing flow.

[0037] In one embodiment, for node feature construction, node The original feature vector is generated from the first feature vector within the current window. It is composed of the mean and standard deviation of each parameter in each process step: ; This is the first step in this process within the window. The mean of each standardized parameter, To correspond to the standard deviation. The original feature dimensions of each process node are different (8, 6, 6, 8, 6, 6, 6 dimensions respectively), and are mapped to a unified embedding dimension through a process-specific linear projection layer. (For example ): ; For the first The projection weight matrix of each process step. These are the bias vectors, which are learned through backpropagation during model training. Embedded nodes after projection. Node features reflect the central tendency and dispersion of parameter distribution for each process within the current window.

[0038] In one embodiment, regarding the construction of edge features, the edge The original features are the vectors obtained by flattening the corresponding partial correlation submatrices by rows. The original dimensions of the 6 edges are 12, 9, 12, 12, 9, and 9, respectively. They are also mapped to 64 dimensions through an edge-specific projection layer. ; For the first The projected weights of the strip edges, This is the projected edge embedding. The edge features carry information about the direct correlation strength of all parameter pairs between adjacent processes.

[0039] In one embodiment, the baseline association structure diagram The construction method is exactly the same as the current window's association structure diagram; the only difference is the data source, i.e. It is generated using historical normal batch data that has passed final inspection (e.g., the most recent 1000 qualified cells) and calculated according to the same process, representing the relationship structure between each process when production is in a normal state.

[0040] Step 104: Graph Change Detection Network Encoding and Anomaly Score Output. The association structure graph of the current window and the baseline association structure graph established based on historical normal production data are input into the graph change detection network. The graph change detection network encodes the embeddings of each node in the two graphs respectively, calculates the difference vector of the corresponding node embeddings, and outputs the batch-level anomaly score and the process-level anomaly contribution score based on the difference vector.

[0041] According to embodiments of this disclosure, the Graph Change Detection Network (GCDN) receives the associated structure graph of the current window. Related structure diagram with reference As input, both graphs contain 7 64-dimensional node embeddings and 6 64-dimensional edge embeddings; the outputs are batch-level anomaly scores (1-dimensional scalar, with values ​​from 0 to 1) and process-level anomaly contribution scores (7-dimensional vectors, with each component being non-negative and summing to 1).

[0042] Figure 2 This is a schematic diagram illustrating the detection of multi-process related structural changes in one embodiment of this application. For example... Figure 2 As shown, the current window association structure graph constructed from the current sliding window data and the baseline association structure graph constructed from historical normal production data are input into the shared GAT encoder. Both association structure graphs use each production process as a node, process parameter statistics as node features, and the partial correlation coefficient matrix between adjacent processes as edge features, establishing directed connections according to the material flow direction. The shared GAT encoder obtains the node embeddings corresponding to each process in the current graph and the baseline graph, and calculates the difference between the corresponding node embeddings to obtain a node difference vector reflecting the degree of deviation of the current production state from the normal baseline association structure. The node difference vector is processed by global average pooling and a fully connected layer to output a batch-level anomaly score. Simultaneously, each node difference vector is processed by a shared fully connected layer and normalized to obtain the anomaly contribution score for each process, thereby determining the process from which the anomaly originates, thus achieving multi-process linked anomaly detection and anomaly source localization.

[0043] Specifically, in one embodiment, GCDN uses a parameter-shared graph attention network (GAT) as the encoder. and Each node embedding representation is obtained after forward propagation through the encoder. The encoder contains two graph attention layers, with LeakyReLU activation connections between the layers.

[0044] Specifically, the first graph attention layer uses four parallel attention heads. For a single attention head, the nodes... From neighboring nodes When aggregating information, first apply query transformations to the embeddings of the target nodes respectively. Apply key transformation to the source node embedding Apply edge transformation to the edge embedding connecting two nodes. The three transformation results are concatenated and then passed through the attention vector. The original attention score is calculated using LeakyReLU (negative slope 0.2): ; in This represents vector concatenation, with the original fraction pairs at the nodes. Attention weights are obtained by softmax normalization of all neighbors. .

[0045] In one embodiment, during the production of pouch batteries, materials flow from upstream to downstream processes. The impact of each process on the product state exhibits a unidirectional causal relationship; that is, the upstream processing result influences the downstream behavior, and the reverse is not true. GCDN imposes directional constraints on attention weights to encode this physical structure. Specifically, it employs an asymmetric attention mechanism, where information flows from downstream process nodes to upstream process nodes (i.e., nodes...). Aggregation originates from process number greater than Neighbors (Information), attention weight multiplied by a decay factor Attention weights remain unchanged as information flows from upstream to downstream. The value is obtained by searching on the validation set. The selection criterion is determined to be the value with the highest accuracy in locating process-level anomalies. In one embodiment... .

[0046] The purpose of this constraint is that, without directional constraints, parameter fluctuations in downstream processes will negatively affect the embedding representation of upstream nodes during graph attention propagation, leading to the misjudgment of problems that actually originate downstream as being caused by upstream processes when locating the source of anomalies.

[0047] Figure 3 This is a schematic diagram of asymmetric graph attention propagation in one embodiment of this application. For example... Figure 3As shown, the graph attention network applies asymmetric constraints to the information propagation between nodes based on the material flow direction of the pouch battery production process. For the target node, query transformation, key transformation, and edge transformation are performed on the target node embedding, source node embedding, and corresponding edge embedding, respectively. The transformation results are concatenated and LeakyReLU is used to calculate the original attention score, which is then normalized by softmax to obtain the attention weight. When information propagates from upstream process nodes to downstream process nodes, the attention weight remains unchanged; when information propagates from downstream process nodes to upstream process nodes, the attention weight is multiplied by a decay factor to suppress the interference of reverse information on the representation of upstream nodes. The shared GAT encoder adopts a two-layer graph attention structure: the first layer uses multi-head attention, and the second layer uses single-head attention, thereby improving the accuracy of anomaly source location while maintaining the ability to express process-related information.

[0048] After attention-weighted aggregation, the nodes The update embedding in a single header is: ; in and These are the value transformation matrix and the edge information transformation matrix, respectively. For nodes The neighbor set. The outputs of the four heads are concatenated to form a 64-dimensional vector as the output of the first layer.

[0049] The second graph attention layer uses a single attention head, with the input being the 64-dimensional output of the first layer. The dimensions of each transformation matrix are... The attention vector has a dimension of 192, and asymmetric constraints are also applied. The final embedding of each node is output. .

[0050] and The current graph node embedding is obtained after passing through a shared encoder. and baseline graph node embedding The difference vector is obtained by subtracting each process node one by one. , characterizing the The process deviates from the associated structure of the normal baseline in the current window.

[0051] The seven difference vectors are obtained by global average pooling. The data is fed into a fully connected layer FC1 (64-dimensional input, 32-dimensional output, ReLU activation), then into FC2 (32-dimensional input, 1-dimensional output, Sigmoid activation), outputting batch-level anomaly scores. .

[0052] Each difference vector Each of the seven steps is fed into a parameter-shared fully connected network, namely FC3 (64-dimensional input, 32-dimensional output, ReLU activation) and then FC4 (32-dimensional input, 1-dimensional output, no activation function). The raw contribution scores are obtained for each of the seven steps, and then normalized by softmax to obtain the step-level abnormal contribution scores. , .

[0053] In some embodiments, regarding model training, optionally, normal training samples are derived from historical batches that have passed final inspection. For example, the final inspection pass criteria are: capacity ≥ nominal capacity × 95%, internal resistance ≤ nominal value × 120%, thickness ≤ nominal value × 110%, appearance free of scratches, bubbles, and wrinkles, and open-circuit voltage within the range of 3.95V to 4.05V. A total of 100 production batches were collected, each batch containing an average of 200 cells. After dividing the batch into sliding windows with a step size of 10 cells, approximately 2000 normal window samples were obtained.

[0054] After dividing the window into sliding windows, each window generates an association structure diagram according to the aforementioned steps, with the label set to "normal". The process contribution label is set to uniform distribution. .

[0055] Among them, abnormal training samples are generated by applying physically interpretable systematic offsets to normal data. Specifically, one or two target processes are selected from the normal window data, the parameters of the target processes are offset, and the correlation structure diagram is recalculated according to the same process, and marked as abnormal. The process contribution label indicates the target process location for which the offset is applied. The offset mode includes three types: linear trend offset simulates progressive equipment wear, and offsets the first process within the window... The target parameters of each battery cell are superimposed. Step offset simulation of raw material batch switching, uniformly superimposed on the latter half of the battery cells in the window. Periodic fluctuation offset simulates equipment vibration, superimposed ,cycle For example, take 10 to 20 battery cells. Offset amplitude. The target parameter tolerance range is set to 30% to 50%, ensuring that the individual parameter remains within the tolerance while the associated structure exhibits detectable changes. In one embodiment, six abnormal samples are generated for each normal window by randomly selecting the process and offset type, with a normal to abnormal sample ratio of approximately 1:6.

[0056] The tolerance ranges for key parameters in each process are set based on product specifications and process capability benchmarks. Taking the coating process as an example, the areal density tolerance is ±1.5 g / m², the coating speed tolerance is ±2 m / min, the oven temperature tolerance is ±5°C, and the slurry viscosity tolerance is ±200 mPa·s. The tolerance ranges for other process parameters follow the same pattern and are uniformly stored in the production line process parameter configuration table. In practical applications, the specific values ​​of the offset amplitude are read and calculated according to the parameter type in this configuration table.

[0057] In one embodiment, training employs a composite loss function: ; Among them, the detection loss item Binary cross-entropy measures the deviation between batch-level anomaly scores and true labels. ; in For batch label (0 or 1). These are the batch-level anomaly scores output by the model; all terms are dimensionless.

[0058] Location loss item The KL divergence (Kullback-Leibler divergence) measures the deviation between the predicted process contribution distribution and the actual distribution. ; The first tag in the actual process contribution The contribution probability of each process step. The score contributes to the process-level anomalies predicted by the model.

[0059] Regularization loss term This only applies to normal samples; penalties are applied to batch-level outlier scores that exceed the tolerance margin. ;

[0060] Among them, tolerance margin The method for determining this is as follows: during the pre-training phase (the first 20 epochs), the distribution of batch-level outlier scores for normal samples is statistically analyzed, and the 95th percentile is taken as the threshold. ,For example This ensures that the vast majority of normal samples do not trigger the penalty.

[0061] Among them, the weighting coefficient and The search range was determined by a grid search on the validation set. , The harmonic mean of the F1 score for anomaly detection and the localization accuracy is used as the selection criterion in one embodiment. , .

[0062] The optimizer used is Adam, with an initial learning rate of 0.001, adjusted using cosine annealing, a period of 50 epochs, and a minimum decay to 0.0001; the batch size is, for example, 32, and training lasts for 150 to 250 epochs. Training stops early when the F1 score on the validation set does not improve for 20 consecutive epochs, preserving the optimal parameters on the validation set. Dropout (with a dropout rate of 0.3) is applied before the fully connected layers to prevent overfitting.

[0063] According to embodiments of this disclosure, the model input and output are completely consistent in both the training and inference phases. The input consists of two association structure graphs (each containing 7 64-dimensional node embeddings and 6 64-dimensional edge embeddings), and the forward propagation path uses shared GAT encoding, difference vector calculation, and two scorehead outputs. During the training phase, the output is used to calculate the loss and perform backpropagation; during the inference phase, the output is directly used for judgment. The correlation between the model input data (association structure graphs) and the output data (batch-level anomaly scores and process-level anomaly contribution scores) lies in the fact that the association structure graphs encode the direct correlation patterns between multiple process parameters in the current batch. When this pattern deviates from the normal baseline, the difference vector carries deviation information and is quantified as the degree of anomaly by the scorehead.

[0064] Step 105: Determine if a multi-process linkage anomaly exists and locate the source process of the anomaly. Based on the batch-level anomaly score, determine if a multi-process linkage anomaly exists in the current window, and locate the source process of the anomaly based on the process-level anomaly contribution score.

[0065] In one embodiment, the batch-level anomaly score output by GCDN is used. With anomaly detection threshold Comparison. Among them The method for determining this is to calculate the batch-level anomaly score for each window on a validation set containing only normal window samples (e.g., 500 normal windows), and then calculate their mean. and standard deviation ,Pick In one example , ,but The false alarm rate under normal production conditions is around 0.3%.

[0066] The aforementioned multi-process linkage anomaly refers to a systematic deviation of the partial correlation coefficient structure between process parameters from the normal baseline due to common root causes (including but not limited to raw material batch consistency deviation, progressive equipment wear, and environmental temperature and humidity drift). Single-process anomalies are manifested as the mean or variance of a single process parameter significantly exceeding the tolerance range, which can be detected by traditional SPC control charts; while multi-process linkage anomalies are manifested as changes in the inter-process correlation structure, which can only be identified through joint analysis of the aforementioned partial correlation coefficient matrix.

[0067] Specifically, when At that time, it is determined that there is a multi-process linkage anomaly in the current window. Simultaneously, the process-level anomaly contribution score is read. The process with the highest contribution score is selected as the primary target for investigation. For example, if a certain test yields... The sixth process (aluminum-plastic film packaging) contributed a significantly higher score than the others. The system identified the packaging process as an abnormal source and prompted the operator to investigate it first.

[0068] Optionally, multiple alarm levels can be set to differentiate severity, with observation-level thresholds. (Record events without interrupting production), early warning threshold (Note: This is a warning sign.) Alarm threshold. (It is recommended to suspend material feeding for investigation). Optionally, if two or more windows trigger alarms of the same level consecutively, the alarm level will be automatically escalated to the next level.

[0069] In some embodiments, the baseline association structure diagram Recalculate in the following situations: product specification change (e.g., changes in cell capacity leading to changes in the number of lamination layers or electrode formulation), replacement of key equipment components (e.g., replacement of coating die head or rolling mill), and change of raw material batch (e.g., change of negative electrode material or electrolyte supplier). The update method involves collecting the data of the first batch of cells that have passed final inspection after the change (e.g., more than 200 cells), recalculating the node features, edge features, and projection results using the exact same process as constructing the current window's associated structure diagram, and replacing the original... Synchronously update the z-score standardized benchmark value and .

[0070] In some embodiments, optionally, if the production line has not undergone the aforementioned changes but has been running continuously for a long period, a gradual update can be adopted. That is, when several consecutive windows (e.g., 50 windows) are all determined to be normal, the current window's associated structure is adjusted using an exponential moving average method. Fusion, with a fusion coefficient of, for example, 0.01.

[0071] Understandably, in terms of computational complexity, inverting the precision matrix involves a 23rd-order square matrix operation, which has extremely low complexity. The GCDN model has approximately 150,000 parameters (mainly distributed in the transformation matrices of the two GAT layers and the fully connected layer of the scoring head), and a single forward inference operation on an industrial edge device equipped with a general-purpose GPU or NPU accelerator card takes only milliseconds. The detection interval corresponding to a window step size of 10 cells is 2 to 5 minutes (depending on the production line cycle time), and the computation time is much shorter than this interval.

[0072] Therefore, this method detects anomalies by monitoring the stability of the correlation structure between parameters in multiple processes, rather than the absolute values ​​of the parameters. It is capable of detecting changes in cross-process correlation patterns caused by common root causes (such as material batch shifts or progressive equipment wear) even when all parameters are within tolerance. The asymmetric attention mechanism encodes the unidirectional causal structure of material flow between processes into a graph attention network, ensuring that the anomaly localization results are not affected by the back propagation of downstream information. Detection and localization are completed synchronously in a single inference, with an overall processing delay within seconds, making it suitable for embedding into real-time monitoring processes on production lines.

[0073] Please see Figure 4 , Figure 4 This is a schematic block diagram of a multi-stage anomaly intelligent detection system for the production process of soft-pack batteries, as described in this application. Figure 4 As shown, the system includes: The multi-process parameter construction module 401 is used to align the process parameters collected in multiple production processes of the same battery cell according to the process sequence through the traceability code, and construct a battery cell-level multi-process parameter vector. The correlation strength calculation module 402 is used to calculate the partial correlation coefficient matrix between adjacent processes for the multi-process parameter vectors of multiple cells within a sliding window, so as to obtain the direct correlation strength of each pair of adjacent processes after controlling the parameters of the remaining processes. The association structure graph construction module 403 is used to construct the association structure graph of the current window by using the partial correlation coefficient matrix as edge features and the parameter statistics of each process within the window as node features. The association structure graph uses each process as a node and the material flow direction between adjacent processes as directed edges. The graph change detection module 404 is used to input the association structure graph of the current window and the benchmark association structure graph established based on historical normal production data into the graph change detection network. The graph change detection network encodes the two graphs to obtain the embedding of each node, calculates the difference vector of the corresponding node embedding, and outputs the batch-level anomaly score and the process-level anomaly contribution score based on the difference vector. The anomaly determination and location module 405 is used to determine whether there is a multi-process linkage anomaly in the current window based on the batch-level anomaly score, and to locate the source process of the anomaly based on the process-level anomaly contribution score.

[0074] The implementation process of the functions and roles of each module in the above system is detailed in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0075] Figure 5 This is a schematic diagram of the structure of an electronic device according to one embodiment of this application. Figure 5 As shown, at the hardware level, the electronic device includes a processor 1102, an internal bus 1104, a network interface 1106, memory 1108, and non-volatile memory 1110, and may also include other hardware required for business operations. One or more embodiments of this specification can be implemented in software, for example, the processor 1102 reads the corresponding computer program from the non-volatile memory 1110 into the memory 1108 and then runs it. Of course, in addition to software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic module, but can also be hardware or logic devices.

[0076] Based on the same concept as the above-described method, this application also provides a machine-readable storage medium storing a plurality of computer instructions. When executed by a processor, these computer instructions can implement the intelligent detection method for multi-process anomalies in the production of soft-pack batteries disclosed in the examples above. The machine-readable storage medium can be any electronic, magnetic, optical, or other physical storage device, and can contain or store information such as executable instructions, data, etc.

[0077] Based on the same application concept as the above method, this application embodiment also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the intelligent detection method for multiple abnormalities in the production process of soft-pack batteries disclosed in the above examples of this application.

[0078] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A method for intelligent detection of anomalies in multiple processes during the production of soft-pack batteries, characterized in that, include: By using traceability codes, process parameters collected from multiple production processes of the same battery cell are aligned according to the process sequence to construct a multi-process parameter vector at the battery cell level. Within the sliding window, the partial correlation coefficient matrix between adjacent processes is calculated for the multi-process parameter vectors of multiple cells, and the direct correlation strength between each pair of adjacent processes after controlling the parameters of the remaining processes is obtained. Using the partial correlation coefficient matrix as edge features and the parameter statistics of each process within the window as node features, a correlation structure graph of the current window is constructed. The correlation structure graph uses each process as a node and the material flow direction between adjacent processes as directed edges. The association structure graph of the current window and the benchmark association structure graph established based on historical normal production data are input into the graph change detection network. The graph change detection network encodes each node embedding in the two graphs respectively, calculates the difference vector of the corresponding node embedding, and outputs the batch-level anomaly score and the process-level anomaly contribution score based on the difference vector. The batch-level anomaly score determines whether there is a multi-process linkage anomaly in the current window, and the process source of the anomaly is located based on the process-level anomaly contribution score.

2. The method according to claim 1, characterized in that, The calculation of the partial correlation coefficient matrix between adjacent processes includes: calculating the covariance matrix of the multi-process parameter vectors of all cells within the sliding window; adding a regularization term to the covariance matrix and then inverting it to obtain the precision matrix; extracting sub-block elements corresponding to each pair of adjacent process parameters from the precision matrix; calculating the partial correlation coefficient between each pair of adjacent process parameters based on the sub-block elements and forming the partial correlation coefficient matrix.

3. The method according to claim 1, characterized in that, The node features are vectors composed of the mean and standard deviation of each parameter of each process within the sliding window, and the edge features are vectors flattened from the partial correlation coefficient matrix between adjacent processes. Each node feature and each edge feature is mapped to a unified embedding dimension through a process-specific linear projection layer and then used as the input features of the association structure graph.

4. The method according to claim 1, characterized in that, The encoder of the graph change detection network is a graph attention network containing multiple graph attention layers. Each graph attention layer adopts an asymmetric attention mechanism when calculating the attention weights between nodes: for information transmitted from downstream process nodes to upstream process nodes, the attention weights are multiplied by a decay factor to suppress it; for information transmitted from upstream process nodes to downstream process nodes, the attention weights remain unchanged.

5. The method according to claim 4, characterized in that, The attention coefficients of each attention head in the graph attention layer are calculated by concatenating the query transformation result of the target node embedding, the key transformation result of the source node embedding, and the edge transformation result of the corresponding edge embedding, and then applying the attention vector and activation function. The graph attention network contains two graph attention layers. The first layer uses multi-head attention, and the outputs of each head are concatenated as the input of the second layer. The second layer uses single-head attention to output the final node embedding. The attenuation factor is determined by searching on the validation set based on the process-level anomaly localization accuracy.

6. The method according to claim 4, characterized in that, The step of outputting batch-level anomaly scores and process-level anomaly contribution scores based on the difference vectors includes: performing global average pooling on the difference vectors of all nodes, then passing them through a fully connected layer and sigmoid activation to output the batch-level anomaly scores; passing the difference vectors of each node through a fully connected layer with shared parameters to output the original contribution scores of each process, and normalizing the original contribution scores of each process to obtain the process-level anomaly contribution scores.

7. The method according to claim 1, characterized in that, The graph change detection network constructs training data in the following way: normal samples are derived from historical batch data that has passed final inspection, and corresponding association structure graphs are generated according to the sliding window; abnormal samples are generated by recalculating the association structure graph after applying a systematic offset to the parameters of selected processes in the normal batch data, wherein the magnitude of the systematic offset is within the tolerance range of the corresponding parameters, and the process contribution label of the abnormal samples is marked as the target process to which the offset is applied.

8. The method according to claim 7, characterized in that, The systematic offset includes at least one of the following modes: linear trend offset simulating progressive wear of equipment, step offset simulating batch switching of raw materials, and periodic fluctuation offset simulating periodic vibration of equipment.

9. The method according to claim 7, characterized in that, The training of the graph change detection network employs a composite loss function, which comprises three weighted sums: a detection loss term for the deviation between the batch-level anomaly score and the true label; a localization loss term for the deviation between the predicted distribution of the process-level anomaly contribution score and the true process contribution distribution; and a regularization loss term that penalizes the portion of the batch-level anomaly score in normal samples that exceeds the tolerance margin. The weight coefficients of each term are selected on the validation set using a grid search based on anomaly detection evaluation metrics.

10. A multi-process anomaly intelligent detection system for the production of soft-pack batteries, characterized in that, include: The multi-process parameter construction module is used to align process parameters collected from multiple production processes of the same battery cell according to the process sequence through traceability codes, and construct a battery cell-level multi-process parameter vector. The correlation strength calculation module is used to calculate the partial correlation coefficient matrix between adjacent processes for the multi-process parameter vectors of multiple cells within a sliding window, so as to obtain the direct correlation strength of each pair of adjacent processes after controlling the parameters of the remaining processes. The association structure graph construction module is used to construct the association structure graph of the current window by using the partial correlation coefficient matrix as edge features and the parameter statistics of each process within the window as node features. The association structure graph uses each process as a node and the material flow direction between adjacent processes as directed edges. The graph change detection module is used to input the association structure graph of the current window and the benchmark association structure graph established based on historical normal production data into the graph change detection network. The graph change detection network encodes the two graphs to obtain the embedding of each node, calculates the difference vector of the corresponding node embedding, and outputs the batch-level anomaly score and the process-level anomaly contribution score based on the difference vector. The anomaly determination and location module is used to determine whether there is a multi-process linkage anomaly in the current window based on the batch-level anomaly score, and to locate the source process of the anomaly based on the process-level anomaly contribution score.

Images