Electrode production management method and system based on quality traceability

Abstract

This invention discloses an electrode production management method and system based on quality traceability, belonging to the field of electrode production management technology. The method includes: acquiring the electrode processing path and real-time process parameter sequence of the current production batch, constructing a target process parameter-electrode performance correlation graph; collecting the initial performance state of each electrode and mapping it to the initial process parameter node of the graph; performing state propagation prediction using the initial process parameter node as the starting node to obtain the predicted performance state of each electrode; calculating the residual sequence between the predicted performance state and the actual performance state, and back-projecting the residual sequence along the processing path to each process node to obtain the traceability contribution; locating the root cause node of quality anomalies based on the traceability contribution, and using real-time output timestamps to perform time-delay dilation correction on the traceability contribution, generating a quality traceability report. This system achieves automated and accurate traceability of electrode production quality.

Description

Technical Field

[0001] This invention relates to the field of electrode production management technology, specifically to an electrode production management method and system based on quality traceability. Background Technology

[0002] The continuity of electrode production causes quality fluctuations to accumulate and propagate along the process chain. Ultimately, deviations in electrode performance are often caused by parameter anomalies at one or more upstream process nodes. Existing quality traceability methods mostly combine post-production offline sampling with empirical inference, roughly identifying problematic processes by comparing finished product test data with historical process acceptable ranges. This method, based on static acceptance threshold comparisons, cannot express the dynamic transmission effect of process parameters between processes. The traceability granularity remains at the process level, lacking quantitative analysis of the contribution of individual process nodes. When multiple process nodes experience parameter drift simultaneously, it is impossible to decompose the responsibility of each node for the final performance deviation, leading to frequent misjudgments or omissions. Furthermore, the time delay effect caused by changes in production cycle time is ignored in existing solutions. Differences in process duration and fluctuations in output rhythm create a misalignment between the time of parameter anomalies and the time of performance degradation. Static time window matching methods cannot correct this temporal offset, causing traceability bias. The field of electrode production management lacks a technical means to characterize the dynamic propagation of process parameters across processes and quantify the impact of each node on the performance of the finished product, so as to solve the problem of attribution separation of multi-node coupling deviations and the time misalignment problem caused by production delays. Summary of the Invention

[0003] This invention provides an electrode production management method and system based on quality traceability, aiming to achieve accurate decoupling and attribution of coupling deviations at multiple process nodes in electrode production, eliminate traceability time misalignment caused by production delays, and improve the accuracy of locating the root cause of quality anomalies.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an electrode production management method and system based on quality traceability. By performing correlation modeling of process parameters and performance status throughout the entire electrode production process, it achieves accurate traceability from finished product performance anomalies to processing node, and utilizes time delay dilation correction to improve the accuracy of traceability results.

[0005] This method first obtains the electrode processing path and real-time process parameter sequence of the current production batch, and constructs a target process parameter-electrode performance correlation graph. Preferably, during this construction process, multiple process nodes and processing edges connecting adjacent process nodes are extracted along the electrode processing path. Each process node is assigned an attribute vector containing the real-time process parameter sequence and node timestamp, and a corresponding transfer function identifier is set for each processing edge. This transfer function identifier is used to describe the influence relationship between the current process node and the electrode performance of downstream process nodes. Based on the process nodes, processing edges, attribute vectors, and transfer function identifiers, a directed acyclic graph is generated as the aforementioned correlation graph. The transfer function corresponding to the transfer function identifier can be selected as a linear transfer function or a nonlinear transfer function according to the actual process characteristics. When a nonlinear transfer function is used, a radial basis function neural network is preferentially used for fitting to more accurately characterize the performance evolution law in the process.

[0006] After obtaining the correlation map, the initial performance state of each electrode in the production batch is collected and mapped to the initial process parameter node of the correlation map. In a preferred embodiment, the collection of initial performance states includes obtaining the initial resistivity, initial tensile strength, and initial elongation of each electrode at the initial processing station, normalizing each performance index to form a normalized initial performance vector, and then storing each vector into the storage unit corresponding to the initial process parameter node according to the electrode processing sequence, thereby achieving accurate linking of the initial state of all electrodes in the batch with the correlation map.

[0007] Starting with the initial process parameter node, state propagation prediction is performed within the target process parameter-electrode performance correlation graph to obtain the predicted performance state of each electrode. Specifically, starting from the initial process parameter node, the propagation proceeds sequentially along the electrode processing path, reading the attribute vector of the current process node and the transfer function identifier of the processing edge between the current node and the downstream node. A pre-stored state propagation operator is then called to calculate the electrode performance prediction offset. This prediction offset is superimposed on the current node state to generate the predicted node state of the downstream process node. This process is repeated until all process nodes are traversed, and the predicted node state of the last process node is taken as the predicted performance state of the electrode. As an improvement, the state propagation operator can include a process parameter deviation propagation operator and a time decay factor. This time decay factor dynamically adjusts the weight of the prediction offset based on the time interval between the current process node and the downstream process node, thus making the prediction more consistent with the time-dependent impact of production cycle time on performance accumulation.

[0008] After obtaining the predicted performance status, it is compared with the actual performance status stored in the electrode performance nodes within the correlation map to calculate the residual sequence. The process involves reading the actual performance status of the electrodes within a batch from the electrode performance nodes, calculating the difference between the predicted and actual performance status for each electrode as the performance residual, arranging the performance residuals of all electrodes according to the processing order to form an initial residual sequence, and then performing local smoothing on the initial residual sequence to eliminate abnormal fluctuations in residuals caused by measurement noise from individual electrodes, resulting in a residual sequence for traceability. The local smoothing process preferably uses sliding window mean filtering or median filtering. The width of the sliding window is dynamically determined based on the actual number of electrodes in the production batch to achieve a balance between smoothing effect and detail preservation.

[0009] After obtaining the residual sequence, the residuals are back-projected to each process node in reverse order of the electrode processing path to obtain the source contribution of each process node. A preferred back-projection method is as follows: take the terminal residual value of the last electrode in the residual sequence, combine it with the terminal process node identifier, and use a preset back propagation coefficient to distribute the terminal residual value in reverse along the processing edge to all upstream process nodes to obtain the first-level contribution; then take the residual value of the second-to-last electrode and perform reverse distribution for the node one level above the terminal process node to obtain the second-level contribution; and so on, propagating backward from tail to head level by level until all residual values ​​are distributed, and accumulating the contribution obtained by each process node at each level to form the source contribution of that node. This back propagation mechanism enables the performance deviation at the end of the product to be attributed layer by layer to each process node in the entire process, supporting the precise location of the source of the anomaly.

[0010] When locating the root cause of quality anomalies based on contribution scores, the contribution score of each process node is compared with a preset contribution threshold. Process nodes exceeding the threshold are extracted as candidate anomaly nodes. After sorting the candidate nodes by contribution score from highest to lowest, the top-ranked candidate nodes are selected, and their corresponding process parameter adjustment records are retrieved. These adjustment records are then correlated with the residual sequence for verification. Nodes that pass the verification are identified as the root cause of the quality anomalies. By introducing this correlation verification step, nodes with inflated contribution scores that are not actually the source of the anomaly can be excluded, thus improving the confidence of the location results.

[0011] For the identified root cause nodes of quality anomalies, the contribution value of these nodes is corrected for time delay by using the real-time output timestamp of the production batch. Specifically, the time difference between the real-time output timestamp of the production batch and the simulated timestamp when the predicted performance state is achieved is obtained. The time dilation coefficient of the node is determined by combining the ratio of its position number in the electrode processing path to the total number of nodes in the path. The product of the time difference and the time dilation coefficient is used as the dilation correction amount. This correction amount is then added to the original contribution value to obtain the contribution value after time delay dilation correction. This correction fully considers the amplification effect of actual production delays on process nodes, making the contribution value more consistent with the actual situation on site.

[0012] Finally, a quality traceability report for this production batch is generated based on the above information. The batch identifier, the name and location of the root cause node of the anomaly are obtained. The contribution after time dilation correction, the real-time process parameter sequence of the node, and the residual sequence are bound together to form a quality traceability data package. A preset report template is called to fill in the corresponding fields of the data package and output the report. This report provides intuitive and traceable data support for process improvement and quality management.

[0013] Correspondingly, the electrode production management system based on quality traceability provided by the present invention includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it can realize each step of the electrode production management method based on quality traceability described above, thereby automatically completing the entire process management from spectrum construction, performance prediction, residual calculation, contribution back projection, anomaly root cause location to time delay correction and report generation on the production site.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0015] Based on the construction of a directed acyclic graph to express the transmission relationship between process parameters and electrode performance, starting from the initial process parameter node, the state transmission operator based on attribute vectors and transfer function identifiers is called node by node along the forward sequence of the processing path. The electrode performance prediction offset is calculated and the state is superimposed to obtain the predicted performance state of each electrode, thereby generating a residual sequence of predicted and actual performance. This forward propagation prediction mechanism explicitly models the offset contribution of each process to performance in the graph transmission process, so that the final performance residual naturally carries the cumulative influence information of each process. It realizes the complete influence link tracing from the source parameters to the final performance, and solves the defect of existing technologies that rely solely on the comparison of finished product qualification thresholds and cannot analyze the dynamic transmission effect of multiple processes. After obtaining the residual sequence, the terminal residual value is distributed layer by layer to all upstream process nodes according to the reverse sequence of the electrode processing path based on the preset reverse propagation coefficient. This reverse distribution operation is performed on the residual value corresponding to each position in the residual sequence. The contribution of each node at each level is accumulated to form the source contribution of that node. This graph-based residual inverse backprojection method deconstructs the overall performance deviation along the processing edge and attributes it to each process node. It quantifies and separates the independent responsibility share of each node in multi-node coupled deviations, avoiding misjudgments caused by the inability to distinguish the differences in deviation contributions between nodes in traditional empirical inference. For the root cause node of quality anomalies, the time difference between the real-time output timestamp and the predicted simulation timestamp of the production batch is obtained. Combined with the relative position index of the node in the processing path, a time dilation coefficient is determined. The time difference is multiplied by the dilation coefficient to obtain the dilation correction amount, which is then superimposed on the original traceability contribution. This time delay dilation correction operation compensates for the time series offset caused by production rhythm fluctuations according to the node position weight, eliminating the misalignment caused by static time matching and further correcting the reliability of the node traceability contribution. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of a quality traceability method based on the correlation graph of process parameters and electrode performance.

[0018] Figure 2 This is a flowchart for constructing the correlation graph between target process parameters and electrode performance;

[0019] Figure 3 It is a flowchart of state propagation prediction based on the correlation graph of process parameters and electrode performance;

[0020] Figure 4 This is a flowchart of the source contribution calculation based on the backward propagation of the residual sequence;

[0021] Figure 5 This is a flowchart for tracing the root causes of quality anomalies and correcting time delay inflation;

[0022] Figure 6 This is a flowchart of the initial performance state sequence generation and graph node mapping process;

[0023] Figure 7 This is a schematic diagram of fitting the nonlinear transfer function of a radial basis function neural network;

[0024] Figure 8 This is a comparison chart of the measured value of the time decay factor and the theoretical fitting curve;

[0025] Figure 9 It consists of the electrode performance residual sequence and the sliding window mid-range filtering smoothing curve;

[0026] Figure 10 It is the normalized initial performance index distribution curve. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] See Figure 1This invention provides an electrode production management method based on quality traceability. This method aims to trace quality deviations occurring during electrode production process step-by-step, decomposing the final performance differences of the electrodes backward along the processing path to each process node, thereby locating the root cause of quality anomalies. The method first obtains the electrode processing path and real-time process parameter sequence for the current production batch. Based on the electrode processing path and real-time process parameter sequence, a target process parameter-electrode performance correlation graph for the production batch is constructed. Subsequently, the initial performance state of each electrode in the production batch is collected and mapped to the initial process parameter node in the target process parameter-electrode performance correlation graph. Using the initial process parameter node as the starting node, state propagation prediction is performed within the target process parameter-electrode performance correlation graph to obtain the predicted performance state of each electrode in the production batch. The residual sequence between the predicted performance state and the actual performance state stored in the electrode performance node within the target process parameter-electrode performance correlation graph is calculated. The residual sequence is then back-projected to each process node of the production batch in reverse order of the electrode processing path to obtain the traceability contribution of each process node. Next, the root cause node of the quality anomaly is located based on the traceability contribution, and the traceability contribution of the root cause node of the quality anomaly is corrected by time delay expansion using the real-time output timestamp of the production batch, thereby generating a quality traceability report for the production batch.

[0029] Example 1:

[0030] In practice, the process of constructing the correlation map between target process parameters and electrode performance includes the following:

[0031] See Figure 2 This function retrieves the electrode processing path for the current production batch. The electrode processing path includes multiple process nodes arranged in the electrode processing sequence, as well as processing edges connecting adjacent process nodes. A process node corresponds to an independent processing step in the electrode production process, such as mixing, pressing, sintering, or cooling. A processing edge represents the processing flow of the electrode from an upstream process node to a downstream process node.

[0032] For each process node, a corresponding attribute vector is set. The attribute vector is a multi-dimensional vector, where one component is the real-time process parameter sequence for this process node, and the other component is the node timestamp. The real-time process parameter sequence consists of multiple process parameter values ​​collected in real-time at this process node. These process parameter values ​​include at least one of sintering temperature, pressing pressure, slurry viscosity, coating thickness, drying time, and sintering atmosphere concentration. These parameter values ​​are collected by sensors arranged on this process node at a predetermined sampling frequency. The node timestamp records the start time of the electrode entering this process node and the end time of its departure from this process node.

[0033] For each processing edge, a transfer function identifier is set for that edge. The transfer function identifier is a code used to uniquely determine the type of transfer function, which includes at least linear and nonlinear transfer function types. The transfer function describes the electrode performance influence relationship from the current process node to the downstream process node; that is, it takes the process parameters contained in the attribute vector of the current process node as input and outputs the predicted offset of the electrode performance of the downstream process node.

[0034] In some implementations, when the transfer function identifier of a certain processing edge corresponds to a linear transfer function type, the corresponding linear transfer function form is that the predicted offset of the electrode performance of the downstream process node is equal to the product of the linear weighting coefficient and the characteristic value of the process parameter of the current process node, and then summed with the bias term. The linear weighting coefficient and the bias term are obtained by least-squares fitting of historical process parameter data and historical electrode performance change data collected before this production batch. The characteristic value of the process parameter of the current process node is a statistical quantity that can characterize the process state extracted from the real-time process parameter sequence of this process node, including the mean or standard deviation.

[0035] In other implementations, when the transfer function identifier of a certain processing edge corresponds to a nonlinear transfer function type, a radial basis function neural network (RBN) is used to fit the nonlinear transfer function. The RBN structure includes an input layer, a hidden layer, and an output layer. The input vector received by the input layer is obtained by normalizing the real-time process parameter sequence from the attribute vector of the current process node; each dimension of the input vector corresponds to a normalized process parameter value. The hidden layer contains J neurons, each using a radial basis function as its activation function. The output layer contains one neuron, outputting the predicted offset of the electrode performance of the downstream process node.

[0036] When constructing the nonlinear transfer function, the output expression of the radial basis function neural network is:

[0037]

[0038] in, This indicates that the radial basis function neural network is applied to the input vector. The output electrode performance affects the predicted value; The input vector consists of the normalized real-time process parameter values ​​of the current process node; The number of neurons in the hidden layer. The value depends on the number of training samples. Dynamically determined, take That is, the square root of the sample size rounded up; For the first The connection weights between hidden layer neurons and output layer neurons; For the first The center vector of the radial basis function corresponding to each hidden layer neuron; For the first The width parameter of the radial basis function corresponding to each hidden layer neuron; Represents the input vector With the center vector The Euclidean distance between them.

[0039] The training process of the radial basis function neural network is as follows: Process parameter data and electrode performance variation data of upstream and downstream process nodes in historical production batches are collected as training samples. Each training sample uses the normalized real-time process parameter vector of the upstream process node as input and the actual electrode performance offset of the downstream process node as the output label. The K-means clustering algorithm is used to cluster the input vectors of the training samples, with the number of clusters set to... The clustering results Each cluster center serves as the center vector of the radial basis functions. For each cluster, calculate the input vector of all samples within the cluster to the cluster center. The maximum distance is used, and half of the maximum distance is taken as the corresponding width parameter. Substituting the input vectors of all training samples into the hidden layer activation function yields the hidden layer output matrix. ,matrix The Line number Column elements are , For the first The input vector of each training sample. The output label vector of the training samples. By solving the system of linear equations The least squares solution yields the connection weight vector. Each component This refers to the corresponding connection weight.

[0040] After setting the attribute vectors of all process nodes and the transfer function identifiers of all processing edges, the process nodes are used as graph nodes, and the processing edges are used as directed edges of the graph. The direction of the directed edges of the graph is consistent with the forward direction of the electrode processing path. The attribute vectors of the process nodes are stored in the data structure of the graph nodes, and the transfer function identifiers of the processing edges are stored in the attributes of the graph edges. This generates a directed acyclic graph for the current production batch. This directed acyclic graph is used as the target process parameter-electrode performance correlation graph.

[0041] See Figure 7In the figure, the horizontal axis represents the normalized process parameter values, ranging from 0 to 1, and the vertical axis represents the electrode performance prediction offset, ranging from approximately -1.0 to 1.0. The scatter points in the figure represent the actual observed data points of the historical training samples, and the curve represents the electrode performance prediction offset output by the radial basis function neural network model.

[0042] The curves exhibit an overall fluctuating nonlinear trend, reflecting the complex nonlinear relationship between electrode performance offset and process parameters. Specifically, as the normalized process parameter value gradually increases from 0 to approximately 0.2, the predicted electrode performance offset shows a slow upward trend, rising from near 0 to approximately 0.3. Subsequently, in the 0.2 to 0.4 range, the predicted performance offset rapidly decreases, reaching a minimum value of approximately -0.8, demonstrating a significant negative offset. Entering the 0.4 to 0.7 range, the predicted performance offset rapidly rebounds, reaching a peak of approximately 0.9, showing a strong positive performance improvement effect. In the 0.7 to 0.85 range, the predicted performance offset decreases again to approximately -0.2 to 0. Finally, from the 0.85 to 1.0 range, the predicted performance offset shows a slow upward trend, approaching 0.2.

[0043] The dense scatter plots fluctuate around the curve, indicating that the radial basis function neural network (RBN) can effectively fit the nonlinear mapping relationship between process parameters collected in historical production batches and electrode performance offsets. The output curve of this RBN is the nonlinear transfer function constructed in Example 1, used to describe the impact of real-time process parameters at the current process node on the predicted offset of electrode performance at downstream process nodes. The non-monotonic variation of the curve reflects the actual production situation where multiple process parameters interact complexly with electrode performance.

[0044] In summary, this figure demonstrates that by training a radial basis function neural network on historical production batch process parameters and electrode performance variation data, a nonlinear transfer function describing the relationship between process parameters and performance deviations was successfully obtained. This effectively constructs the nonlinear transfer function part of the target process parameter-electrode performance correlation graph, laying the foundation for subsequent state propagation prediction and residual backpropagation.

[0045] Example 2:

[0046] In specific implementation, please refer to Figure 3 The initial process parameter node is located within the target process parameter-electrode performance correlation graph. The node state of the initial process parameter node is represented by the normalized initial performance vector mapped to this node. The normalized initial performance vector is a multi-dimensional vector with dimensions consistent with the number of electrode performance indicators. Each dimension corresponds to the normalized value of one initial performance indicator of the electrode, such as normalized initial resistivity, normalized initial tensile strength, and normalized initial elongation.

[0047] Using the initial process parameter node as the current process node, state propagation prediction is performed sequentially along the electrode machining path. The attribute vector of the current process node, along with a machining edge connecting the current process node to its immediate downstream process node, is obtained from the target process parameter-electrode performance correlation graph. The attribute vector of the current process node contains the real-time process parameter sequence and the node timestamp, which includes the start time of the electrode entering the current process node and the end time of its departure.

[0048] The transfer function identifier stored on the aforementioned processing edge is read, and the pre-stored state transfer operator is called according to the transfer function identifier. The state transfer operator consists of a process parameter deviation transfer operator and a time decay factor. The process parameter deviation transfer operator is used to calculate the original offset of electrode performance from the process parameters of the current process node, and the time decay factor is used to adjust the weight of the original offset according to the time interval between the current process node and the downstream process node.

[0049] The process parameter deviation transfer operator in the state transfer operator corresponds one-to-one with the transfer function identifier. When the transfer function identifier corresponds to a linear transfer function type, the process parameter deviation transfer operator uses a pre-stored linear regression coefficient vector and bias value. The linear regression coefficient vector is inner-producted with the statistical feature vector extracted from the real-time process parameter sequence of the current process node, and then the bias value is added to obtain the original offset. The statistical feature vector can include the mean, standard deviation, and range of the real-time process parameter sequence of the current process node. The linear regression coefficient vector and bias value are least-squares solutions obtained through training with historical production batch data. During training, the statistical characteristics of the process parameters of the current process node in the historical production batches are used as independent variables, and the actual measured changes in electrode performance at downstream process nodes are used as dependent variables.

[0050] When the transfer function identifier corresponds to a nonlinear transfer function type, the process parameter deviation transfer operator uses a pre-trained radial basis function (RBF) neural network model. The input of the RBF neural network model is the normalized real-time process parameter vector of the current process node, and the output is the original offset. The construction method of the RBF neural network model is consistent with the fitting method of the nonlinear transfer function in the aforementioned target process parameter-electrode performance correlation graph. Its core architecture includes an input layer, a hidden layer, and an output layer. The hidden layer uses the radial basis function as the activation function, and the model parameters include the center vector, width parameter, and connection weight of each neuron in the hidden layer. These parameters have been determined and stored during the graph construction.

[0051] The time decay factor is calculated analytically based on the time interval between the current process node and the downstream process node. The calculation formula is as follows:

[0052]

[0053] in, This represents the time decay factor, and its output range is... ; This is the time interval between the current process node and the direct downstream process node. The time interval is obtained by subtracting the end time of the current process node's timestamp from the start time in the downstream process node's timestamp, and the unit is minutes. The time decay coefficient is a preset value. The range of values ​​is Time decay coefficient The following method was used to determine this: Collect full-process data and final electrode performance data from multiple historical production batches, aiming to minimize the mean square error between the predicted state of the last process node and the actual performance state of the electrode, within a preset range. Inner step length right A grid search is performed to select the value that minimizes the mean square error. In this embodiment, the value obtained through the grid search is... The optimal value is Therefore, Set as .

[0054] When calling the state propagation operator to calculate the electrode performance prediction offset propagated from the current process node to the downstream process node, the real-time process parameter sequence in the attribute vector of the current process node is first input into the process parameter deviation propagation operator to obtain the original offset; then the original offset is multiplied by the time decay factor, and the product is used as the electrode performance prediction offset.

[0055] The predicted offset of electrode performance is superimposed on the node state of the current process node to generate the predicted node state of the downstream process node. That is, the predicted node state of the downstream process node is equal to the sum of the node state vector of the current process node and the electrode performance predicted offset vector. The dimension of the node state vector of the current process node is the same as the dimension of the electrode performance predicted offset of the downstream process node, both being the number of electrode performance indicators. If the electrode performance indicators include resistivity, tensile strength, and elongation, then all the above vectors are three-dimensional vectors, and the superposition operation adds the vector elements one by one.

[0056] Set the downstream process node as the new current process node and repeat the above process: obtain the attribute vector of the new current process node, obtain the transfer function identifier on the machining edge connecting the new current process node and the next direct downstream process node, call the corresponding state transfer operator to calculate the new electrode performance prediction offset, and superimpose it with the node state of the new current process node to obtain the predicted node state of the next process node. Propagate step by step along the forward sequence of the electrode machining path in this way until all process nodes in the target process parameter-electrode performance correlation graph have been traversed.

[0057] When the propagation process reaches the last process node in the electrode processing path, the predicted node state of the last process node is the predicted performance state of each electrode in that production batch. The predicted performance state is a multi-dimensional vector containing predicted values ​​for the final resistivity, final tensile strength, and final elongation of the electrode. Since the entire prediction process is performed independently for different electrodes within the same production batch, the initial performance vector of each electrode may be different, thus enabling the acquisition of the predicted performance state of each electrode in that production batch.

[0058] See Figure 8 In the figure, the horizontal axis represents the time interval Δt in minutes, and the vertical axis represents the time decay factor Γ, indicating the degree to which the influence of process parameters on electrode performance decreases as the time interval increases. The scatter points in the figure are measured decay factor data collected from historical production batches. The measured decay factor is obtained by calculating the performance impact weight corresponding to the time interval between adjacent process nodes. The measured data points show a monotonically decreasing trend along the vertical axis, indicating that the transmission weight of performance impact gradually decreases as the time interval increases.

[0059] The solid line in the figure represents the theoretical decay curve, expressed as Γ(Δt) = exp(-0.03 × Δt), where the exponential decay coefficient 0.03 is the time decay coefficient η obtained by performing a grid search optimal fit on multiple historical batches of data. The overall distribution of the measured scatter points closely follows the theoretical decay curve, demonstrating the good fit of the theoretical model to the actual decay characteristics. Specifically, when the time interval Δt is close to 0, the decay factor Γ is close to 1, indicating that the performance impact is basically unaffected; as Δt increases to about 50 minutes, the decay factor rapidly decreases to about 0.2, indicating that the performance impact is significantly weakened; when Δt reaches 100 minutes or more, the decay factor approaches 0, indicating that the performance impact caused by the longer time interval almost disappears.

[0060] This curve reflects the analytical calculation formula and parameter settings of the time decay factor in Example 2. It quantitatively describes the weighting effect of the time interval between process nodes on performance transfer using an exponential function model, providing a theoretical basis and empirical support for time decay correction in state propagation prediction. The figure effectively verifies the rationality and applicability of selecting η=0.03 as the time decay coefficient, ensuring that the state propagation process accurately reflects the impact of time factors on electrode performance deviation.

[0061] Example 3:

[0062] In specific implementation, please refer to Figure 4 The process of calculating the residual sequence between the predicted performance state and the actual performance state and backpropagating it to obtain the source contribution includes the following steps.

[0063] Within the target process parameter-electrode performance correlation graph, each electrode performance node stores the actual electrode performance state of each electrode in that production batch. The actual electrode performance state is a multi-dimensional vector, with the dimensions matching the number of electrode performance indicators. Electrode performance indicators include at least resistivity, tensile strength, and elongation. The actual electrode performance state is collected by testing equipment located at the end of the production batch's testing station.

[0064] For each electrode in the production batch, the predicted performance state vector and the actual performance state vector of the electrode are obtained. The difference between the predicted and actual performance state vectors is calculated, and the difference is determined as follows: each component of the predicted performance state vector is subtracted from the corresponding component of the actual performance state vector to obtain the electrode performance residual vector. The dimension of the electrode performance residual vector is the same as the dimension of the predicted performance state vector, and each component of the electrode performance residual vector represents the deviation between the predicted and actual values ​​of the corresponding performance index. The magnitude of the electrode performance residual vector of an electrode, or the component value of a specific performance index in the electrode performance residual vector, can be used as the performance residual of that electrode. In this embodiment, the component value of the electrode performance residual vector in terms of resistivity is used as the performance residual of the electrode.

[0065] Following the processing sequence of the electrode processing path, i.e., the time sequence in which the electrode passes through each process node, the performance residuals of all electrodes in this production batch are arranged to generate an initial residual sequence. The initial residual sequence is a one-dimensional sequence, where the k-th element corresponds to the performance residual of the k-th processed electrode.

[0066] Local smoothing is performed on the initial residual sequence to eliminate abrupt changes in residuals caused by measurement noise from individual electrodes. Local smoothing employs a sliding window midpoint filter, specifically implemented as follows: a sliding window width W is set, dynamically determined based on the number of electrodes M in the production batch. The rules for determining the sliding window width W are as follows: when the number of electrodes M is less than or equal to 50, W is set to 3; when the number of electrodes M is greater than 50 and less than or equal to 200, W is set to 5; and when the number of electrodes M is greater than 200, W is set to 7. This determination rule is based on the following: when the electrode production batch is small, abnormal fluctuations caused by measurement noise have a significant impact, but an excessively wide window will over-smooth the true changes; a smaller fixed window is selected to retain most of the fluctuation information. As the batch size increases, the statistical characteristics tend to stabilize, and the window size is appropriately increased to more effectively suppress noise. For each position i in the initial residual sequence, W consecutive residual values ​​centered at i are taken (if the position is close to the sequence boundary, the number of elements in the window may be less than W), and the median of these W residual values ​​is taken as the smoothed residual value at that position. The residual values ​​after smoothing all positions constitute a residual sequence.

[0067] In other implementations, local smoothing can be achieved using sliding window mean filtering. The width of the sliding window for mean filtering is determined in the same way as for median filtering, i.e., the arithmetic mean of W residuals is used as the smoothed residual value.

[0068] After obtaining the residual sequence, reverse projection is performed. The terminal residual value corresponding to the last electrode in the residual sequence is obtained. This terminal residual value is the last element of the smoothed residual sequence. At the same time, the node identifier of the terminal process node of the electrode processing path is obtained. The terminal process node is the process node with an out-degree of 0 in the target process parameter-electrode performance correlation graph.

[0069] Based on the node identifier of the final process node, all backpropagation coefficients associated with the final process node are retrieved from a pre-set backpropagation coefficient table. Each record in the pre-set backpropagation coefficient table corresponds to the reverse direction of a directed processing edge, i.e., the direction from the downstream process node to the upstream process node. The backpropagation coefficients are pre-set based on the structure of the target process parameter-electrode performance correlation graph and the transfer function identifier of the processing edge. For processing edges corresponding to linear transfer function identifiers, the backpropagation coefficient uses the normalized weight of the linear regression coefficient; for processing edges corresponding to nonlinear transfer function identifiers, the backpropagation coefficient uses the average absolute value of the input contribution to the output in the radial basis function neural network, calculated by the integral gradient method and normalized. The final residual value is distributed to each upstream process node of the final process node according to these backpropagation coefficients. The distribution method is: the final residual value is multiplied by the backpropagation coefficient corresponding to each reverse edge to obtain the first-level contribution received by the upstream process node. Therefore, each upstream process node obtains a first-level contribution scalar.

[0070] The residual value corresponding to the second-to-last electrode in the residual sequence is used as the current propagation residual. Simultaneously, the node identifier of the previous process node of the final process node is obtained. The previous process node of the final process node is the upstream process node directly connected to the final process node in the electrode processing path. For this previous process node, its associated backpropagation coefficient is called, and the current propagation residual is distributed to all directly upstream process nodes of this previous process node according to the backpropagation coefficient. The distribution method is the same as described above, and each assigned upstream process node receives a second-level contribution. It should be noted that if an upstream process node receives contributions in multiple levels, each received contribution is retained separately.

[0071] This reverse allocation process is repeated, moving forward one electrode position each time, corresponding to the residual value of the previous electrode in the residual sequence, while simultaneously backtracking one process node upstream of the electrode processing path, until the residual value corresponding to the first electrode in the residual sequence has been allocated. In each allocation, the current residual value is multiplied by the backpropagation coefficient associated with the current process node, and then allocated to all directly upstream process nodes of the current process node.

[0072] After all reverse allocation processes are completed, each process node may receive multiple contribution values ​​from multiple levels (i.e., multiple different currently propagating process nodes). The contribution values ​​received by each process node from all levels are summed, and the sum is the traceability contribution of that process node. The traceability contribution indicates the degree to which that process node contributes to the final electrode performance deviation.

[0073] See Figure 9The figure shows the initial residual point distribution (gray scattered points) and the curve trend of the residual sequence (black solid line) after sliding window midpoint filtering smoothing in Example 3, for the electrode resistivity performance residual sequence within the current production batch. The horizontal axis represents the electrode serial number, ranging from 1 to approximately 500, reflecting the actual processing sequence of the electrodes; the vertical axis represents the electrode resistivity performance residual, with a value range of approximately -1.0 to 3.0, and the residual is based on the difference between the predicted and actual values ​​of the normalized resistivity index.

[0074] The initial residual sequence exhibits relatively discrete characteristics, specifically large fluctuations and isolated outliers at each processing step. While the residual values ​​fluctuate around 0, significant anomalies exceeding 1.5 appear in local intervals, and sporadic residual points also exist in the negative range, indicating that the individual electrode residuals are subject to strong random fluctuations due to measurement noise. After sliding window midpoint filtering, the smoothed residual sequence significantly suppresses sharp anomalies and noise fluctuations in the initial sequence, resulting in a smoother and more continuous curve that shows a clear phased change trend.

[0075] Within the range of electrode numbers from approximately 0 to 200, the smoothed residual values ​​generally show a slow upward trend, with the residual values ​​mainly distributed between 0 and 1, indicating that the electrode resistivity deviation is generally small and gradually increases during this stage.

[0076] In the range of serial numbers from approximately 200 to 350, the residual sequence shows a significant jump, with the residual values ​​stabilizing between 1 and 2. In particular, a brief peak occurs near serial number 270, reaching a local maximum of approximately 2.0, reflecting a significant increase in electrode resistivity deviation during this stage, which may correspond to the influence range of the root cause node of the quality anomaly.

[0077] After number 350, the residual value gradually dropped below 1 and fluctuated between 0.5 and 1.5 in the range of 400 to 500, indicating that the resistivity residual gradually stabilized and the degree of abnormality weakened.

[0078] This figure clearly demonstrates the temporal characteristics and abrupt change ranges of the electrode resistivity residuals through the residual changes along the electrode sequence dimension, providing a quantitative analytical basis for the inverse projection and root cause node localization of quality anomalies in subsequent embodiments. The application of sliding window midpoint filtering effectively eliminates residual anomalies caused by measurement noise from a single electrode, ensuring the continuity and trend of the residual sequence and meeting the technical requirements for local smoothing of the residual sequence in Embodiment 3.

[0079] Example 4:

[0080] In specific implementation, please refer to Figure 5The traceability contribution of each process node is compared one by one with a preset contribution threshold. The contribution threshold is determined based on the statistical distribution of the traceability contributions of all process nodes, and is calculated by obtaining the traceability contribution set of all process nodes and calculating the average value of the traceability contribution set. with standard deviation Contribution threshold Set as The reason for using the average plus three standard deviations as the threshold is that, assuming process fluctuations follow a normal distribution, deviations exceeding three standard deviations are low-probability events and have a high degree of anomaly indicativeness. Process nodes whose contribution exceeds the contribution threshold are extracted as candidate anomaly nodes.

[0081] All candidate anomaly nodes are sorted from largest to smallest according to their contribution to source tracing, resulting in an ordered sequence of candidate anomaly nodes. A predetermined number of candidate anomaly nodes, ranked first in the sequence, are then selected; this predetermined number is set to three. The rationale for setting this number to three is that the probability of multiple unrelated independent anomaly sources occurring simultaneously in the electrode production process is extremely low; retaining the top three candidate nodes is sufficient to cover the main sources of anomalies while also ensuring investigation efficiency.

[0082] Obtain the process parameter adjustment records corresponding to each of the three candidate anomaly nodes. These records are generated by the Manufacturing Execution System (MES) during production and include the time the process parameter was modified, the parameter value before modification, the parameter value after modification, and the modification operation type. Then, correlate and verify the process parameter adjustment records of each candidate anomaly node with the residual sequence.

[0083] The specific implementation of correlation verification is as follows: For a candidate anomaly node, the set of time points when the process parameters were adjusted is extracted from the process parameter adjustment records of the candidate anomaly node. In the residual sequence, each electrode has a processing sequence index, and each electrode has a processing timestamp when it passes through the corresponding process of the candidate anomaly node. The process parameter adjustment time points are matched with the processing timestamps of each electrode at the candidate anomaly node to determine the range of electrodes affected by the process parameter adjustment event. The mean residual values ​​of U electrodes processed before the process parameter adjustment event and V electrodes processed after the process parameter adjustment event are calculated, with U and V both preset to 5. The absolute value of the difference between the two mean residual values ​​is calculated. If the absolute value exceeds a preset change threshold, it is determined that there is a causal relationship between the process parameter adjustment event of the candidate anomaly node and the performance fluctuation in the residual sequence, and the correlation verification is passed. The change threshold is set to twice the standard deviation of the residuals of all electrodes in the residual sequence to distinguish between significant changes and random fluctuations caused by process parameter adjustments. The candidate anomaly node that passes the correlation verification is located as the root cause node of the quality anomaly. If the first three candidate abnormal nodes fail the association verification, then candidate abnormal nodes are added in order of the candidate abnormal node sequence and the association verification is repeated until a node or sequence that has passed the association verification is exhausted.

[0084] After locating the root cause of the quality anomaly, a time-delay inflation correction is applied to the contribution of that root cause to the tracing process. The real-time output timestamp of the current production batch is obtained from the production management system. This timestamp is the clock time recorded by the system after the last electrode of the current production batch completes all performance index acquisition at the end-of-line inspection station. Simultaneously, the simulated timestamp when the predicted performance state of all electrodes within the production batch is completed is obtained. This simulated timestamp is the system clock time when the state propagation prediction process is completed and the predicted performance state of the last electrode is generated. The time difference between the real-time output timestamp and the simulated timestamp is calculated, in minutes.

[0085] Obtain the position number of the root cause node of the quality anomaly in the electrode machining path. The position numbering starts from 1 and increments sequentially from the starting process node to the ending process node. Obtain the total number of nodes in the electrode machining path. Determine the time dilation coefficient of the root cause node of the quality anomaly based on the ratio of the position number to the total number of nodes, while introducing a pre-calibrated base dilation coefficient. The formula for calculating the contribution of the time dilation correction is as follows:

[0086]

[0087] in, This indicates the contribution of source tracing after time-delay inflation correction; This indicates the contribution of the root cause node of the quality anomaly before correction. This indicates the real-time output timestamp of the current production batch; Indicates the simulation timestamp when the predicted performance state is completed; This represents the time difference between the real-time output timestamp and the simulated timestamp, in minutes. This represents the pre-calibrated basic expansion coefficient. The range of values ​​is , This implementation method is determined through historical case fitting. Set as ; This indicates the position number of the root cause node of the quality defect in the electrode processing path; Indicates the total number of nodes in the electrode machining path. Basic expansion coefficient. Set as The basis for this approach is as follows: Multiple production batches with confirmed root causes of quality anomalies are selected from the historical production database. The time difference between the real-time output timestamp and the simulated timestamp, the anomaly root cause node position number, and the verified anomaly contribution are extracted for each batch. The above formula is used to minimize the mean square error between the corrected contribution and the verified anomaly contribution, within a specified range. Inner The step size is used to perform parameter search to obtain the parameter that minimizes the mean square error. Value Therefore, adopt As the basic expansion coefficient.

[0088] The inflation correction is obtained by multiplying the time difference between the real-time output timestamp and the simulated timestamp by the base inflation coefficient and then by the ratio of the position number to the total number of nodes. This inflation correction is then added to the contribution of the root cause node of the quality anomaly to obtain the contribution after inflation correction. .

[0089] Retrieve the batch identifier of the current production batch. The batch identifier is a unique coded string issued by the production planning system. Retrieve the node name and node location of the root cause node of the quality anomaly. The node name is the process name of this process node in the electrode processing path, and the node location is the location sequence number. Bind the contribution value after time dilation correction, the real-time process parameter sequence of the root cause node of the quality anomaly, and the residual sequence with the batch identifier, node name, and node location to form a structured quality traceability data package. The quality traceability data package is organized in key-value pair format and includes a batch identifier field, a node name field, a node location field, a corrected contribution value field, a real-time process parameter sequence field, and a residual sequence field.

[0090] The system invokes a pre-defined report template, which is a file in Extensible Markup Language (XML) format and contains placeholder tags corresponding to the fields mentioned above. The data values ​​from the quality traceability data package are sequentially filled into the placeholder positions of the corresponding fields in the report template, completing template instantiation. The completed report template is output in text format as the quality traceability report for the current production batch. The quality traceability report fully records the root cause information of anomalies, contribution metrics, relevant process parameter curves, and residual fluctuations, allowing quality engineers to directly view and perform traceability analysis.

[0091] Example 5:

[0092] In specific implementation, please refer to Figure 6 The initial performance status of each electrode in the current production batch is collected, and initial performance testing equipment is deployed at the starting processing station of the current production batch. The starting processing station is the first process in the electrode production process after the electrode raw material or original blank is formed. The initial performance testing equipment includes a resistivity meter, a universal testing machine, and an extensometer. For each electrode in the current production batch, before the electrode leaves the starting processing station, initial resistivity, initial tensile strength, and initial elongation are collected sequentially. Initial resistivity is collected by measuring and reading the resistivity value at a designated location on the surface of the electrode blank using a four-probe resistivity meter; initial tensile strength is collected by performing a tensile test on an accompanying sample from the same batch using a universal testing machine to obtain the tensile strength value; initial elongation is collected by recording the elongation within the gauge length at tensile fracture using an extensometer and calculating the elongation value. The initial resistivity value, initial tensile strength value, and initial elongation value obtained above constitute the initial performance ternary set of an electrode.

[0093] The initial performance triplet for each electrode is normalized to obtain the normalized initial performance vector for that electrode. The normalization process uses a maximum-minimum normalization method, and the specific normalization calculation formula is as follows:

[0094]

[0095] in, This indicates that all electrodes in the current production batch are in the [number]th [stage]. Normalized values ​​for each performance metric Choose 1, 2, and 3, which correspond to resistivity performance index, tensile strength performance index, and elongation performance index, respectively. This indicates the current normalized electrode performance index. The original collected values; This indicates that all electrodes in the current production batch meet the performance specifications. The minimum value of the original collected values; This indicates that all electrodes in the current production batch meet the performance specifications. The maximum value of the original collected values. The basis for using the maximum and minimum values ​​of the batch values ​​as the normalization boundary is that electrode raw materials from the same production batch are consistent, and process settings are similar. The performance fluctuation range within a batch is more concentrated than across batches. Using the extreme values ​​within a batch for normalization can more accurately reflect the relative differences between electrodes within the batch. According to the above formula, for an electrode, three normalized values ​​are calculated for the resistivity, tensile strength, and elongation performance indicators respectively. These three normalized values ​​are combined into a three-dimensional vector, which is the normalized initial performance vector of the electrode.

[0096] The normalized initial performance vectors of all electrodes in the current production batch are arranged according to the actual processing order of the electrodes to generate an initial performance state sequence. The actual processing order of the electrodes is the order in which they are processed at the starting processing station, and this order is recorded and distributed by the production execution system. The initial performance state sequence is an ordered list, where the first element corresponds to the normalized initial performance vector of the first processed electrode, the second element corresponds to the normalized initial performance vector of the second processed electrode, and so on, until the last electrode.

[0097] Each normalized initial performance vector in the initial performance state sequence is sequentially mapped to the storage unit corresponding to the initial process parameter node of the target process parameter-electrode performance correlation map. The initial process parameter node of the target process parameter-electrode performance correlation map is a data node pre-created during the map construction phase. This initial process parameter node contains an array of storage units, the length of which is equal to the total number of electrodes in the current production batch. The mapping operation writes the i-th normalized initial performance vector in the initial performance state sequence to the i-th storage location in the storage unit array, where i is numbered starting from 1. This establishes the correspondence between the initial performance state of each electrode and the map structure.

[0098] The aforementioned electrode production management method based on quality traceability operates within a quality traceability-based electrode production management system. This system includes a memory, a processor, and a computer program stored in the memory and running on the processor. The memory stores the target process parameter-electrode performance correlation graph, a pre-set backpropagation coefficient table, radial basis function neural network model parameters, state transfer operators, historical production data, and intermediate data generated in each embodiment. The processor calls the computer program in the memory. When executing the computer program, the processor implements the steps of the quality traceability-based electrode production management method described above, including constructing the target process parameter-electrode performance correlation graph, state propagation prediction, residual sequence calculation and back projection, locating the root cause node of quality anomalies, time delay dilation correction, and all processing logic for generating quality traceability reports.

[0099] See Figure 10 The horizontal axis represents the normalized initial performance value (0~1), and the vertical axis represents the probability density. The peak probability density values ​​of the three performance indicators are concentrated in the range of 2.0~2.75. The figure contains three curves, representing the normalized initial resistivity distribution (solid line), the normalized initial tensile strength distribution (dashed line), and the normalized initial elongation distribution (dotted line), respectively, corresponding to the statistical distribution of the initial performance state of each electrode in the current production batch collected at the starting processing station.

[0100] As can be seen from the curve shape, the normalized initial resistivity distribution curve has a bimodal shape, with the peak value located in the range of approximately 1.75 to 2.25 of the normalized performance value. This indicates that the initial resistivity of most electrodes is relatively concentrated in this range, with a relatively symmetrical distribution and gentle fluctuations. The peak value of the normalized initial tensile strength distribution curve is high and narrow, located at approximately 2.25, indicating that the initial state of tensile strength exhibits less dispersion. The peak value of the normalized initial elongation distribution curve is slightly lower than the peak value of tensile strength, at approximately 2.25, and the distribution range is slightly wider, indicating that the initial performance fluctuation of elongation is slightly larger than that of resistivity.

[0101] Overall, the three normalized initial performance indices are concentrated in the range of 2.0 to 2.75, with probability densities all above 2. This indicates that the initial performance states of most electrodes in the current production batch are relatively consistent, showing a certain concentration trend. This normalized initial performance distribution reflects the effect of the initial performance acquisition and normalization process in Example 5, providing basic data support for subsequently mapping the normalized initial performance states to the initial process parameter nodes of the target process parameter-electrode performance correlation graph. This ensures that the initial performance vector of each electrode accurately reflects its relative differences within the batch, thereby guaranteeing the accuracy of state propagation prediction and the precision of quality traceability.

[0102] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An electrode production management method based on quality traceability, characterized in that, include: Obtain the electrode processing path and real-time process parameter sequence for the current production batch, and construct the target process parameter-electrode performance correlation map for the production batch based on the electrode processing path and real-time process parameter sequence. The initial performance status of each electrode in the production batch is collected, and the initial performance status is mapped to the initial process parameter node of the target process parameter-electrode performance correlation graph. Within the target process parameter-electrode performance correlation graph, state propagation prediction is performed using the initial process parameter node as the starting node to obtain the predicted performance state of each electrode in the production batch. Calculate the residual sequence between the predicted performance state and the actual performance state stored in the electrode performance node of the target process parameter-electrode performance correlation graph. Then, back-project the residual sequence to each process node of the production batch in the reverse order of the electrode processing path to obtain the traceability contribution of each process node. Based on the traceability contribution, the root cause node of the quality anomaly is located, and the traceability contribution of the root cause node of the quality anomaly is corrected by time delay expansion using the real-time output timestamp of the production batch, thereby generating a quality traceability report for the production batch. The specific steps of back-projecting the residual sequence to each process node of the production batch in reverse order of the electrode processing path to obtain the traceability contribution of each process node include: Obtain the terminal residual value corresponding to the last electrode in the residual sequence, and obtain the node identifier of the terminal process node of the electrode processing path; Based on the node identifier of the end process node, a preset reverse propagation coefficient is invoked to distribute the end residual value to all upstream process nodes of the end process node in the reverse direction of the processing edge, thereby obtaining the first-level contribution of each upstream process node. The residual value corresponding to the second-to-last electrode in the residual sequence is taken as the current propagation residual, and the node identifier of the previous process node of the end process node is obtained. The reverse allocation is then performed to obtain the second-level contribution of each upstream process node. Repeat the reverse allocation until the residual value corresponding to the first electrode in the residual sequence is allocated. Then, accumulate all the hierarchical contributions received by each process node to obtain the traceability contribution of each process node. The traceability contribution represents the degree of contribution of the process node to the final electrode performance deviation. The time delay dilation correction involves obtaining the time difference between the real-time output timestamp of the production batch and the simulated timestamp when the predicted performance state is completed. The time dilation coefficient of the node is determined by combining the position number of the root cause node of the quality anomaly in the electrode processing path with the total number of nodes in the path. The product of the time difference and the time dilation coefficient is used as the dilation correction amount. This correction amount is superimposed with the original source tracing contribution to obtain the contribution after time delay dilation correction.

2. The electrode production management method based on quality traceability according to claim 1, characterized in that, The specific steps for constructing the target process parameters-electrode performance correlation graph for this production batch include: Obtain multiple process nodes on the electrode processing path of this production batch, as well as the processing edges connecting the multiple process nodes; A corresponding attribute vector is set for each process node, and the attribute vector contains the real-time process parameter sequence and node timestamp of the process node; A corresponding transfer function identifier is set for each processing edge. The transfer function identifier is used to describe the electrode performance influence relationship between the current process node and the downstream process node. Based on the multiple process nodes, processing edges, attribute vectors and transfer function identifiers, a directed acyclic graph of the production batch is generated as the target process parameter-electrode performance correlation graph.

3. The electrode production management method based on quality traceability according to claim 2, characterized in that, The transfer function identifier indicates whether the transfer function is a linear transfer function or a nonlinear transfer function. The nonlinear transfer function is fitted using a radial basis function neural network.

4. The electrode production management method based on quality traceability according to claim 2, characterized in that, Within the target process parameter-electrode performance correlation graph, the steps of performing state propagation prediction using the initial process parameter node as the starting node to obtain the predicted performance state of each electrode in this production batch specifically include: Starting from the initial process parameter node, according to the forward order of the electrode processing path, the attribute vector of the current process node and the transfer function identifier of the processing edge connecting the current process node and the downstream process node are obtained sequentially. Based on the attribute vector of the current process node and the transfer function identifier, the pre-stored state transfer operator is called to calculate the electrode performance prediction offset propagated from the current process node to the downstream process node. The electrode performance prediction offset is superimposed with the node status of the current process node to generate the predicted node status of the downstream process node. The above steps are repeated until all process nodes in the target process parameter-electrode performance correlation graph are traversed. The predicted node status of the last process node is used as the predicted performance status of each electrode in this production batch.

5. The electrode production management method based on quality traceability according to claim 4, characterized in that, The state transfer operator includes a process parameter deviation transfer operator and a time decay factor. The time decay factor is used to adjust the weight of the prediction offset according to the time interval between the current process node and the downstream process node.

6. The electrode production management method based on quality traceability according to claim 4, characterized in that, The steps for calculating the residual sequence between the predicted performance state and the actual performance state stored in the electrode performance node of the target process parameter-electrode performance correlation graph specifically include: Locate the electrode performance node within the target process parameter-electrode performance correlation graph, and read the actual electrode performance status of the production batch stored within the electrode performance node; For each electrode, the difference between the predicted performance state and the actual performance state of the electrode is calculated, and the difference is used as the performance residual of the electrode. According to the processing sequence of the electrode processing path, the performance residuals of all electrodes in the production batch are arranged to generate the initial residual sequence of the production batch. The initial residual sequence is locally smoothed to eliminate residual abrupt changes caused by noise from a single electrode measurement, thus obtaining the residual sequence.

7. The electrode production management method based on quality traceability according to claim 6, characterized in that, The initial residual sequence is locally smoothed by means filtering or median filtering using a sliding window, wherein the width of the sliding window is dynamically determined according to the number of electrodes in the production batch.

8. The electrode production management method based on quality traceability according to claim 7, characterized in that, The steps for locating the root cause node of quality anomalies based on the aforementioned source tracing contribution specifically include: The traceability contribution of each process node is compared with a preset contribution threshold. Process nodes whose traceability contribution exceeds the contribution threshold are extracted as candidate abnormal nodes. The candidate abnormal nodes are sorted in descending order of traceability contribution to obtain a candidate abnormal node sequence. Take out the first preset number of candidate abnormal nodes from the candidate abnormal node sequence, obtain the process parameter adjustment records corresponding to each of the first preset number of candidate abnormal nodes, perform correlation verification between the process parameter adjustment records and the residual sequence, and locate the candidate abnormal nodes that pass the correlation verification as the root cause nodes of quality abnormalities.

9. An electrode production management system based on quality traceability, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the electrode production management method based on quality traceability as described in any one of claims 1 to 8.

Images