Battery cell K value anomaly detection method and system

Abstract

The invention discloses a cell K value anomaly detection method and system, and relates to the technical field of lithium ion battery manufacturing and detection.The method comprises the steps that production process data of a target cell and open-circuit voltage data at multiple set time points in the standing process are collected; reconstructing the open-circuit voltage data into a time sequence, and splicing the time sequence with the production process data to form a multi-dimensional input feature vector; inputting the feature vector into a pre-trained GRU-BiLSTM hybrid prediction model, and outputting a prediction K value and a dichotomy judgment result; according to the invention, abnormity traceability is realized through an SHAP algorithm. According to the method, accurate prediction and early abnormality recognition of the K value of the battery cell are realized, the detection period is shortened to 3-5 days, the abnormality classification accuracy is improved to 95% or above, meanwhile, an abnormality root can be quickly positioned, and the lithium battery production efficiency, quality control and traceability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery manufacturing and testing technology, and more specifically, to a method and system for detecting abnormal K-values ​​in battery cells. Background Technology

[0002] In the lithium battery manufacturing industry, the cell's K-value (self-discharge rate) is a core indicator for measuring its storage performance and cycle life. An abnormal K-value (such as a sudden increase in K-value) can directly lead to rapid capacity decay, poor consistency, and even safety accidents. The current mainstream K-value detection method in the industry is "static-open circuit voltage (OCV) testing," which involves placing the cell in a constant temperature environment of 25°C for 7-14 days, then calculating the K-value (K=ΔOCV / Δt) by the difference in OCV before and after static storage, and comparing it with a fixed threshold (≤0.05mV / h) to determine whether it is abnormal.

[0003] To improve testing efficiency, some companies have attempted to shorten the settling time or increase the testing frequency, but several technical limitations remain: First, traditional methods rely on manual reading of OCV data and calculation of K values, which is not only inefficient but also prone to misjudgment due to human error. Second, relying solely on the "OCV difference at a single time point" to determine anomalies does not consider the dynamic decay trend of OCV during settling or correlate it with cell production process data, resulting in an anomaly detection accuracy of only 75%-85%. This allows a large number of potentially abnormal cells to flow into downstream processes, increasing after-sales risks. Third, data such as electrode compaction density, liquid injection volume, and formation temperature generated during production, as well as OCV data from multiple settling time points, are only archived without establishing a correlation analysis model with K values. This results in low data utilization and fails to provide support for anomaly tracing. Fourth, the testing cycle is long, requiring a full settling period before the K value status can be determined. The 7-14 day settling period cannot meet the large-scale production needs of the lithium battery industry. If batch-related K value anomalies occur, it will lead to a large backlog of cells, severely impacting delivery cycles. Summary of the Invention

[0004] To address the issues of low detection efficiency, low anomaly detection accuracy, low data utilization, and long detection cycle in existing battery cell K-value detection technologies, this invention proposes a battery cell K-value anomaly detection method and system. By integrating multi-dimensional production process data and OCV dynamic time-series data, and using a GRU-BiLSTM hybrid model for prediction, this method can significantly improve detection efficiency and anomaly detection accuracy while shortening the detection cycle, thereby increasing data utilization.

[0005] According to one aspect of the present invention, a method for detecting abnormal K-values ​​in battery cells is provided, comprising the following steps: S1, collecting production process data of the target battery cell and open-circuit voltage data at multiple set time points within a preset resting time, wherein the production process data includes at least one of the following: electrode compaction density, liquid injection volume, formation voltage, formation time, electrode baking temperature, and electrode tab welding strength; S2, reconstructing the open-circuit voltage data into an open-circuit voltage time series and concatenating it with the production process data to form a multi-dimensional input feature vector; S3, inputting the multi-dimensional input feature vector into a pre-trained hybrid prediction model, wherein the hybrid prediction model is used to extract the temporal features of the open-circuit voltage time series, fuse the temporal features with the production process data, and output the predicted K-value of the target battery cell and a binary classification judgment result regarding whether the K-value is abnormal based on the fused features.

[0006] More specifically, the hybrid prediction model in step S3 includes an input layer, a GRU layer, a BiLSTM layer, a feature fusion layer, and an output layer connected in sequence. The input layer receives the multidimensional input feature vector and inputs the open-circuit voltage time series sequence into the GRU layer, while transmitting the production process data to the feature fusion layer. The GRU layer uses the ReLU activation function to extract short-term time series features of the open-circuit voltage time series. The BiLSTM layer uses the tanh activation function to extract long-term time series features of the open-circuit voltage time series from both forward and reverse directions based on the short-term time series features. The feature fusion layer includes two fully connected networks, both using the ReLU activation function, to concatenate and fuse the long-term time series features with the production process data. The output layer includes a first output unit and a second output unit. The first output unit uses the Linear activation function to generate the predicted K value based on the fused features output by the feature fusion layer. The second output unit uses the Sigmoid activation function to obtain the binary classification result based on the predicted K value through a preset classification probability threshold.

[0007] Furthermore, a joint loss function based on combining MSE loss and cross-entropy loss is proposed. The hybrid prediction model is trained using training data, which is a set of labeled multidimensional input feature vectors. The labels include the true K-value label and the true classification label. The joint loss function... The formula is as follows:

[0008] in, Weight the loss for MSE; The MSE loss function is... Where N is the total number of battery cells to be tested. Labels for the true K-value. To predict the K value; Let cross-entropy be the loss function. Where N is the total number of battery cells to be tested. For real category tags, To predict classification probabilities.

[0009] Furthermore, the binary classification result is deemed abnormal based on at least one of the following: the predicted K value output by the output layer of the hybrid prediction model exceeds a preset dynamic threshold; or the classification probability output by the output layer of the hybrid prediction model is greater than the preset classification probability threshold.

[0010] Furthermore, the dynamic threshold is set through the following steps: Based on the characteristics affecting the K-value decay law of the battery cell, the battery cells are divided into different type groups, wherein the characteristics include the battery cell chemical system; for each battery cell type group, open-circuit voltage data and corresponding K-value data of historical normal battery cells in that group are collected at multiple set time points within a preset resting time; based on the K-value data corresponding to each battery cell type group, the K-value decay threshold at different set time points is analyzed; a mapping relationship between the preset resting time and the K-value decay threshold is established to generate the dynamic threshold for that type group.

[0011] Furthermore, the cell K-value anomaly detection method of the present invention further includes: when the binary classification judgment result is abnormal, extracting the multi-dimensional input feature vector corresponding to the abnormal cell; calculating the SHAP value of each dimension of the data in the multi-dimensional input feature vector using the SHAP interpretability algorithm; determining the data with the highest weight in terms of influence on the K-value anomaly based on the absolute value sorting of the SHAP values; and generating a source tracing report indicating the root cause of the anomaly by combining the production process standards corresponding to each data.

[0012] Based on the above, further, in step S1, the preset settling time is from 1 hour to 168 hours after the start of settling, and there are no fewer than three set time points.

[0013] Based on the above aspects, further, before step S2, the collected production process data and open circuit voltage data are cleaned, and the cleaned production process data is standardized.

[0014] According to another aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the above-described method for detecting abnormal K-values ​​of battery cells when executing the computer program.

[0015] According to another aspect of the present invention, a battery cell K-value anomaly detection system is provided, comprising a cloud layer, an edge layer, and a terminal layer. The terminal layer is deployed in a battery cell production workshop and configured to collect raw data, including production process data of the target battery cell and open-circuit voltage data at multiple set time points within a preset resting period. The edge layer is deployed on a local server in the workshop and includes the aforementioned electronic device. The electronic device is used to execute a battery cell K-value anomaly detection method based on the raw data from the terminal layer to generate a predicted K-value and anomaly determination result, and to output an early warning signal and a source tracing report when the battery cell K-value is abnormal. The cloud layer is deployed on a cloud platform and configured to retrain and update a hybrid prediction model based on historical data received from the electronic device, and to send the updated hybrid prediction model to the electronic device.

[0016] The beneficial effects of this invention are as follows: This invention uses a GRU-BiLSTM hybrid model to analyze early static data, such as the first 72 hours, which integrates multi-dimensional production process data and OCV dynamic time-series data. It can complete the K-value anomaly determination within 3-5 days of static storage, greatly improving detection efficiency and significantly shortening the traditional 7-14 day detection cycle, avoiding cell backlog and adapting to the needs of large-scale production and delivery.

[0017] This invention uses a joint loss function combining MSE loss and cross-entropy loss to train the model, which can control the K-value prediction error within ±3% and improve the anomaly classification accuracy to over 95%, significantly reducing the probability of qualified cells being misjudged and scrapped and defective cells entering the market.

[0018] Based on historical data of different cell types, this invention establishes a dynamic threshold for K-value decay by type. Compared with the traditional fixed threshold, it avoids misjudgment caused by differences in cell material and further improves the accuracy of early warning.

[0019] This invention introduces SHAP interpretability analysis to quantify the influence weight of various production process data on K-value anomalies. This allows for rapid location of specific production links after an anomaly occurs, significantly improving traceability efficiency and providing direct data support for process optimization.

[0020] The three-layer architecture of the cell K-value anomaly detection system of the present invention, namely cloud layer-edge layer-terminal layer, enables real-time data acquisition, local processing and cloud-based iterative optimization. This ensures timely anomaly warnings and allows for continuous model updates based on historical data, adapting to the testing needs of multiple models and batches of lithium batteries. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the cell K-value anomaly detection method of the present invention.

[0022] Figure 2 This is a schematic diagram of the network structure of an embodiment of the hybrid prediction model of the present invention.

[0023] Figure 3 This is a comparison chart of the dynamic threshold curve of the K-value of the battery cell of the present invention and the measured data.

[0024] Figure 4 This is a structural diagram of the electronic device of the present invention. Detailed Implementation

[0025] like Figure 1 As shown, according to an embodiment of the present invention, a method for detecting abnormal K-values ​​in battery cells is provided, comprising the following steps: S1, collecting production process data of the target battery cell and open-circuit voltage data at multiple set time points within a preset resting time, wherein the production process data includes at least one of the following: electrode compaction density, liquid injection volume, formation voltage, formation time, electrode baking temperature, and electrode tab welding strength; S2, reconstructing the open-circuit voltage data into an open-circuit voltage time series and concatenating it with the production process data to form a multi-dimensional input feature vector; S3, inputting the multi-dimensional input feature vector into a pre-trained hybrid prediction model, wherein the hybrid prediction model is used to extract the temporal features of the open-circuit voltage time series, fuse the temporal features with the production process data, and output the predicted K-value of the target battery cell and a binary classification judgment result regarding whether the K-value is abnormal based on the fused features. The cell K-value anomaly detection method of this invention breaks away from the traditional detection mode of single OCV difference and long-term static placement. It constructs a hybrid model for prediction by acquiring time-series features from multiple sources, simultaneously outputting the predicted K-value and binary classification results. This solves the shortcomings of traditional methods, such as long detection cycles and only post-event judgment, achieving high efficiency in K-value detection and early anomaly identification. This invention integrates static data from the production process and dynamic OCV time-series data from the static placement process, overcoming the limitations of traditional methods that rely on only a single data dimension.

[0026] Preferably, the cell K-value anomaly detection method of the present invention includes cleaning the collected production process data and open-circuit voltage data before step S2, and standardizing the cleaned production process data. Specifically, the production process data is standardized using Min-Max (mapping the data to the [0,1] interval), as shown in the following formula:

[0027] Where x is the original data, , These are the minimum and maximum values ​​of the data in this dimension, respectively. This is the standardized data.

[0028] Specifically, the production process data of the present invention includes one or more of the following: electrode compaction density (0.3-0.8 g / cm³), liquid injection volume (5-20 g), formation voltage (3.0-4.2 V), formation time (2-8 h), electrode baking temperature (80-120 °C), and electrode tab welding strength (50-150 N).

[0029] In addition, this invention can collect OCV data during the resting process using a high-precision voltage acquisition instrument (accuracy 0.001mV), recording the OCV values ​​of the battery cells after resting at a constant temperature of 25℃ for 1h, 6h, 12h, 24h, 36h, 48h, 60h, 72h, 96h, 120h, 144h, and 168h. The actual K value calculated after 168h (7 days) (K=ΔOCV / 168h) is used as the label of the hybrid prediction model of this invention, where K≤0.05mV / h is "normal" and K>0.05mV / h is "abnormal".

[0030] like Figure 2 As shown, specifically, the hybrid prediction model of the present invention includes an input layer, a GRU layer, a BiLSTM layer, a feature fusion layer, and an output layer connected in sequence, with the following specific structure: Input layer: Receives 18-dimensional input data (6-dimensional static data + 12-dimensional time-series data), where the time-series data is input to the GRU layer and the static data is directly passed to the fully connected layer; GRU layer: Set to 1 layer with 32 neurons and ReLU activation function, used to capture short-term dependencies in OCV time series data (such as the OCV decay pattern in the first 24 hours of rest). BiLSTM layer: Set to 1 layer with 64 neurons, activation function is tanh, receives the output of GRU layer, analyzes the long-term correlation of OCV time series from two directions: "forward (time increment)" and "backward (time decrement)", output dimension is 128 (64×2). Fully connected layer: There are 2 layers. The first layer has 64 neurons (ReLU activation) and the second layer has 32 neurons (ReLU activation). The output of the BiLSTM layer is concatenated with 6-dimensional static production data and then feature fusion is performed. Output layer: 1 neuron, using the Linear activation function, outputs the predicted K value; at the same time, 1 classification output node (Sigmoid activation) is added to output the judgment result of "normal / abnormal" (0 for normal, 1 for abnormal).

[0031] The GRU layer is used to perform nonlinear transformation on the open-circuit voltage time series to accelerate gradient convergence during model training and capture the rapid nonlinear change characteristics of short-term open-circuit voltage decay. The output range of the BiLSTM layer is constrained within the (-1,1) interval to prevent gradient explosion during bidirectional transmission of the time series features and to stably capture the long-term, bidirectional dependency of open-circuit voltage decay. The feature fusion layer includes two fully connected network layers to introduce sparse activation when fusing time series features with production data, suppressing interference from irrelevant features and enhancing the nonlinear expressive power of the model. The output layer includes a first output unit and a second output unit. The first output unit uses a Linear activation function to output the continuous value of the predicted K value without distortion. The second output unit uses a Sigmoid activation function to map the fused features to a probability space of (0,1) to output the classification probability representing the possibility of anomalies in the K value, and then obtains a binary classification result through a preset classification probability threshold (e.g., 0.8).

[0032] As an innovation of this invention, it employs a joint loss function that combines MSE loss and cross-entropy loss. The hybrid prediction model described above is trained using training data, which is a set of labeled multidimensional input feature vectors. The labels include the true K-value label and the true classification label. The joint loss function... The formula is as follows:

[0033] in, Weight the loss for MSE; The MSE loss function is... ,in, For the true K value, To predict the K value; Let cross-entropy be the loss function. ,in, For real category tags, To predict classification probabilities.

[0034] The training parameters of the hybrid prediction model are as follows: optimizer is Adam; learning rate is 0.001; number of iterations is 200; batch size is 32; training set to test set ratio is 8:2. The training process is as follows: Data input: The preprocessed multi-source data is input into the model in batches, the time-series OCV data is passed into the GRU layer, and the static production data is temporarily stored. Feature extraction: The GRU layer outputs short-term temporal features (32 dimensions), which are then fed into the BiLSTM layer to output bidirectional long-term temporal features (128 dimensions). Feature fusion: 128-dimensional time-series features are concatenated with 6-dimensional static production data to form 134-dimensional fused features, which are then fed into a fully connected layer for nonlinear transformation; The gradient is calculated using the backpropagation algorithm, the weights and biases of each layer of the model are updated, and the joint loss is minimized. When the prediction error of the K value on the test set is ≤3% and the anomaly classification accuracy is ≥95%, the model training is completed and the parameters are saved.

[0035] According to the hybrid prediction model of the present invention, the binary classification result is determined to be abnormal based on at least one of the following: the predicted K value output by the output layer of the hybrid prediction model exceeds a preset dynamic threshold; or the classification probability output by the output layer of the hybrid prediction model is greater than the preset classification probability threshold.

[0036] The dynamic threshold is set through the following steps: Based on the characteristics affecting the K-value decay law of the battery cell, the battery cells are divided into different type groups, where the characteristics include the battery cell chemical system; for each battery cell type group, open-circuit voltage data and corresponding K-value data of historical normal battery cells in that group are collected at multiple set time points within a preset resting time; based on the K-value data corresponding to each battery cell type group, the K-value decay threshold at different set time points is analyzed; a mapping relationship between the preset resting time and the K-value decay threshold is established to generate the dynamic threshold for that type group.

[0037] like Figure 3 As shown, as an example of dynamic thresholds, for ternary lithium batteries, the OCV decay threshold after 3 days (72h) of rest is 0.15mV / h, and the threshold after 5 days (120h) of rest is 0.2mV / h; for lithium iron phosphate batteries, the OCV decay threshold after 3 days of rest is 0.1mV / h, and the threshold after 5 days of rest is 0.12mV / h. If the classification probability is ≥0.8 (indicating an abnormal tendency), or the predicted K value exceeds the dynamic threshold corresponding to the resting time, an abnormality warning is triggered; when the resting time reaches 72h (3 days), if the model determines an abnormality, the warning result is directly output; if it determines normality, monitoring continues until 168h (7 days) to finally confirm whether the K value is normal.

[0038] The cell K-value anomaly detection method of the present invention further includes: when the binary classification judgment result is abnormal, extracting the multi-dimensional input feature vector corresponding to the abnormal cell; calculating the SHAP value of each dimension of the data in the multi-dimensional input feature vector using the SHAP interpretability algorithm; determining the data with the highest weight in terms of influence on the K-value anomaly based on the absolute value sorting of the SHAP values; and generating a source tracing report indicating the root cause of the anomaly by combining the production process standards corresponding to each data.

[0039] Specifically, when the system triggers an anomaly warning, the SHAP value analysis module is activated to quantify the impact weight of each input data on the anomaly. The steps are as follows: extract the 18-dimensional input data of the abnormal battery cell and input it into the SHAP interpreter; calculate the SHAP value of each input dimension (positive values ​​indicate that the data in that dimension drives the K value to increase, and negative values ​​indicate inhibition); sort the data by the absolute value of the SHAP values ​​and output the top 3 influencing factors with the highest contribution; combine the production process standard range to determine whether the top 3 contributing factors exceed the standard and generate a traceability report.

[0040] According to another aspect of the present invention, an electronic device is provided. Figure 4 A schematic diagram of the physical structure of the electronic device of the present invention is shown, such as... Figure 4 As shown, the electronic device may include: a processor 110, a communication interface 120, a memory 130, and a communication bus 110, wherein the processor 110, the communication interface 120, and the memory 130 communicate with each other through the communication bus 110. The processor 110 can call the computer program in the memory 130 to execute the cell K-value anomaly detection method of the present invention. The method includes: S1, collecting production process data of the target cell and open circuit voltage data at multiple set time points within a preset resting time, wherein the production process data includes at least one of the following: electrode compaction density, liquid injection volume, formation voltage, formation time, electrode baking temperature, and electrode tab welding strength; S2, reconstructing the open circuit voltage data into an open circuit voltage time series and splicing it with the production process data to form a multi-dimensional input feature vector; S3, inputting the multi-dimensional input feature vector into a pre-trained hybrid prediction model, wherein the hybrid prediction model is used to extract the time series features of the open circuit voltage time series, fuse the time series features with the production process data, and output the predicted K-value of the target cell and a binary classification judgment result on whether the K-value is abnormal based on the fused features.

[0041] Furthermore, when the computer program in memory 130 can be implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0042] According to another aspect of the present invention, a battery cell K-value anomaly detection system is provided, comprising: a cloud layer, an edge layer, and a terminal layer, wherein the terminal layer is deployed in a battery cell production workshop and configured to collect raw data, the raw data including production process data of the target battery cell and open-circuit voltage data at multiple set time points within a preset resting time; the edge layer is deployed on a local server in the workshop and includes the electronic device of the present invention, the electronic device being used to execute a battery cell K-value anomaly detection method based on the raw data of the terminal layer to generate a predicted K-value and anomaly determination result, and to output an early warning signal and a source tracing report when the battery cell K-value is abnormal; the cloud layer is deployed on a cloud platform and configured to retrain and update a hybrid prediction model based on historical data received from the electronic device, and to send the updated hybrid prediction model to the electronic device.

[0043] Specifically, the terminal layer includes a MES system (for collecting production data), a high-precision voltage acquisition instrument (for collecting OCV data), and a constant-temperature static chamber (for controlling the static environment). Data is uploaded to the edge layer in real time via industrial Ethernet with a transmission delay of ≤1s. The edge layer is responsible for data preprocessing (cleaning and standardization) and real-time analysis. After OCV data is collected, it is immediately input into the GRU-BiLSTM model to generate real-time prediction results. If an anomaly warning is triggered, a local audible and visual alarm is activated. The cloud layer stores historical data (production data, OCV data, and anomaly records) and periodically (weekly) supplements the model training set with historical data to retrain the model and improve accuracy. It also provides a web-based visualization interface, allowing engineers to view abnormal cell lists, traceability reports, model accuracy trends, and other data, and supports data export and report generation.

[0044] In summary, this invention, through a multi-source data fusion system, a GRU-BiLSTM hybrid prediction model, a SHAP value traceability mechanism, and a dynamic threshold strategy, can shorten the cell K-value detection cycle from the traditional 7-14 days to 3-5 days, improving single-batch turnover efficiency and anomaly detection accuracy, and reducing material waste and after-sales losses. By leveraging SHAP value analysis to compress anomaly traceability time, it helps to quickly adjust process parameters to reduce production line downtime. At the same time, by continuously iterating the model using historical data, it can further predict K-values ​​and extend the early warning time, forming a positive cycle of "data-model-process optimization," which efficiently adapts to the production needs of multiple lithium battery models and batches.

[0045] To make the technical solution of the present invention clearer and easier to understand, the present invention will be described in detail below with reference to specific embodiments.

[0046] First specific embodiment: Detection of abnormal K-value in ternary lithium battery cells A ternary lithium battery cell (model: 3.7V / 2000mAh) produced by a lithium battery company was selected as the test object. Production process data of this batch of cells were collected through the MES system: electrode compaction density 0.65g / cm³, liquid injection volume 15.2g, formation voltage 3.7V, formation time 6h, electrode baking temperature 105℃, and electrode tab welding strength 120N. The OCV values ​​of the cells were collected at 1h, 6h, 12h, 24h, 36h, 48h, 60h, 72h, 96h, 120h, 144h, and 168h at a constant temperature of 25℃ using a voltage acquisition instrument with an accuracy of 0.001mV. Abnormal cell data is removed; MinMax standardization is performed on the production process data, and the OCV data at 12 time points is reconstructed into a time series, which is then concatenated with the standardized production process data to form an 18-dimensional input feature vector. A hybrid prediction model was constructed, with α=0.6 in the joint loss function, Adam as the optimizer, a learning rate of 0.001, 200 iterations, and a batch size of 32. The training and test sets were divided in an 8:2 ratio. After training, the K-value prediction error on the test set was 2.3%, and the anomaly classification accuracy was 96.8%, meeting the training criteria. The 18-dimensional feature vector of the battery cell under test was input into the model to obtain the predicted K-value and classification probability: the predicted OCV decay rate of the battery cell after 72 hours (3 days) of rest was 0.17mV / h, exceeding the dynamic threshold of 0.15mV / h for ternary batteries, and the classification probability was 0.86, which was greater than the preset classification probability threshold of 0.8, thus indicating an abnormal K-value. The SHAP interpretability algorithm was used to analyze the abnormal cell data, and the SHAP values ​​for each dimension were calculated: electrolyte injection deviation SHAP value 0.43, formation temperature deviation SHAP value 0.30, and electrode compaction density deviation SHAP value 0.20, which were the top three contributing factors. A review of production records revealed that the actual electrolyte injection volume of the cell was 14.3g, lower than the standard range of 15±0.5g. A traceability report was generated: "Insufficient electrolyte injection volume caused an abnormal K value; it is recommended to adjust the electrolyte injection valve flow rate parameter to 15.2±0.1g / cell."

[0047] Second specific embodiment: Application of the lithium iron phosphate battery cell K-value anomaly detection system The cell K-value anomaly detection system of this invention was deployed in a lithium battery production plant: The terminal layer is equipped with an MES system, a voltage acquisition instrument with an accuracy of 0.001mV, and a 25℃ constant temperature static chamber; The edge layer uses local servers with Intel Core i7-12700 CPUs and 16GB of memory to deploy various functional modules; The cloud platform was built to store nearly two years of production and testing data for the enterprise.

[0048] After the system was operational, a batch of lithium iron phosphate cells (model: 3.2V / 3000mAh) was tested. After 72 hours of rest, 18 abnormal cells were identified. An anomaly tracing analysis module revealed that the formation temperature of these 18 cells exceeded the standard range of 80-120℃ (actually 128℃). Based on the tracing report, technicians adjusted the formation process temperature parameters to 100℃. Subsequent batches saw their K-value defect rate decrease from 3.9% to 0.7%, resulting in a 20% increase in production efficiency.

Claims

1. A method for detecting abnormal K-values ​​in battery cells, characterized in that, Includes the following steps: S1. Collect production process data of the target cell and open circuit voltage data at multiple set time points within a preset resting time. The production process data includes at least one of the following: electrode compaction density, liquid injection volume, formation voltage, formation time, electrode baking temperature, and electrode tab welding strength. S2. Reconstruct the open-circuit voltage data into an open-circuit voltage time series, and concatenate it with the production process data to form a multi-dimensional input feature vector; S3. Input the multidimensional input feature vector into the pre-trained hybrid prediction model, wherein the hybrid prediction model is used to extract the time series features of the open circuit voltage time series, fuse the time series features with the production process data, and output the predicted K value of the target cell and the binary classification judgment result on whether the K value is abnormal based on the fused features.

2. The method for detecting abnormal K-value of a battery cell according to claim 1, characterized in that, The hybrid prediction model in step S3 includes an input layer, a GRU layer, a BiLSTM layer, a feature fusion layer, and an output layer connected in sequence. The input layer is used to receive the multidimensional input feature vector, input the open-circuit voltage time sequence therein into the GRU layer, and send the production process data to the feature fusion layer; The GRU layer uses the ReLU activation function to extract short-term time-series features of the open-circuit voltage time series. The BiLSTM layer employs the tanh activation function to extract long-term time-series features of the open-circuit voltage time-series sequence from both the forward and reverse directions based on the short-term time-series features. The feature fusion layer comprises two fully connected networks, both employing the ReLU activation function, used to concatenate and fuse the long-term time-series features with the production process data; The output layer includes a first output unit and a second output unit. The first output unit uses a Linear activation function to generate the predicted K value based on the fused features output by the feature fusion layer. The second output unit uses a Sigmoid activation function to obtain the binary classification result based on the predicted K value through a preset classification probability threshold.

3. The cell K-value anomaly detection method according to claim 2, characterized in that, Joint loss function based on MSE loss and cross-entropy loss The hybrid prediction model is trained using training data, which is a set of labeled multidimensional input feature vectors. The labels include the true K-value label and the true classification label. The joint loss function... The formula is as follows: in, Weight the loss for MSE; The MSE loss function is... Where N is the total number of battery cells to be tested. Labels for the true K-value. To predict the K value; Let cross-entropy be the loss function. Where N is the total number of battery cells to be tested. For real category tags, To predict classification probabilities.

4. The method for detecting abnormal K-values ​​in battery cells according to claim 1, characterized in that, The binary classification result is deemed abnormal based on at least one of the following: The predicted K value output by the output layer of the hybrid prediction model exceeds a preset dynamic threshold; The classification probability output by the output layer of the hybrid prediction model is greater than the preset classification probability threshold.

5. The cell K-value anomaly detection method according to claim 4, characterized in that, The dynamic threshold is set through the following steps: Based on the characteristics that affect the decay law of the K value of the battery cell, the battery cells are divided into different types of groups, wherein the characteristics include the battery cell chemical system; For each cell type group, collect open-circuit voltage data and corresponding K-value data of the historical normal cells in that group at multiple set time points within a preset resting time. Based on the K-value data corresponding to each cell type group, the K-value decay threshold at different set time points is analyzed and obtained. Establish a mapping relationship between the preset resting time and the K-value decay threshold to generate the dynamic threshold for this type group.

6. The method for detecting abnormal K-value of a battery cell according to claim 1, characterized in that, Further includes: When the binary classification result is abnormal, the multi-dimensional input feature vector corresponding to the abnormal battery cell is extracted; The SHAP interpretability algorithm is used to calculate the SHAP value of each dimension of the multidimensional input feature vector. Based on the absolute value of the SHAP value, determine the data with the highest weight in terms of the impact of abnormal K value; By combining the production process standards corresponding to each data point, a source tracing report is generated that identifies the root cause of the anomaly.

7. The method for detecting abnormal K-value of a battery cell according to any one of claims 1 to 6, characterized in that, In step S1, the preset settling time is from 1 hour to 168 hours after the start of settling, and there are no fewer than three set time points.

8. The cell K-value anomaly detection method according to claim 7, characterized in that, Before step S2, the collected production process data and open-circuit voltage data are cleaned, and the cleaned production process data is standardized.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, characterized in that, When the processor executes the computer program, it implements the cell K-value anomaly detection method as described in any one of claims 1 to 8.

10. A battery cell K-value anomaly detection system, characterized in that, It includes the cloud layer, edge layer, and terminal layer, among which, The terminal layer is deployed in the cell production workshop and is configured to collect raw data, including production process data of the target cell and open circuit voltage data at multiple set time points within a preset resting time. The edge layer is deployed on a local server in the workshop and includes the electronic device as described in claim 9. The electronic device is used to execute a cell K-value anomaly detection method based on the raw data of the terminal layer to generate a predicted K-value and anomaly judgment result, and output an early warning signal and a source tracing report when the cell K-value is abnormal. The cloud layer is deployed on a cloud platform and is configured to retrain and update the hybrid prediction model based on historical data received from the electronic device, and send the updated hybrid prediction model to the electronic device.

Images