Lithium battery production process abnormity diagnosis method and system based on predicted capacity deviation

Abstract

The invention discloses a lithium battery production process anomaly diagnosis method and system based on predicted capacity deviation, and belongs to the technical field of new energy batteries. The method comprises the following steps: acquiring multi-process data in a battery production process; preprocessing the process data; training a deep learning model based on an FT-Transform architecture by using the preprocessed data, so that the model can predict the discharge capacity of the battery according to the data in the production process; inputting production process data of a to-be-diagnosed battery into the trained model to obtain a'predicted capacity 'of the to-be-diagnosed battery; and comparing the'predicted capacity 'with the'actual capacity' obtained by the capacity grading test of the battery, and calculating the capacity deviation between the'predicted capacity 'and the'actual capacity'. According to the method, the abnormal battery with problems in the performance and production process can be efficiently and accurately identified by utilizing the existing production data without adding extra tests, and a new technical approach is provided for improving the overall quality and safety of a battery system.

Description

Technical Field

[0001] This invention relates to the field of new energy battery technology, specifically to a method and system for diagnosing abnormalities in the lithium battery production process based on predicted capacity deviation. Background Technology

[0002] In the mass production of lithium-ion batteries, ensuring high consistency among each individual cell is crucial for the safe and long-life operation of the final battery system. Currently, the industry commonly uses static indicators such as capacity, voltage, and self-discharge (K-value) for screening and grading. However, these methods only ensure that the selected batteries are similar in static parameters, but cannot effectively identify "hidden" abnormal batteries that may have potential performance degradation or safety risks due to minor, unrecorded fluctuations in the production process (such as uneven electrolyte wetting, minor wrinkles in the separator, or momentary equipment malfunctions). Although these batteries initially meet the static parameters, they may fail prematurely during long-term cycling, seriously affecting the overall reliability of the battery system. Adding additional physical testing procedures (such as rate discharge and DC internal resistance) can improve the screening effect, but it significantly increases production time and equipment costs. Therefore, there is an urgent need for a new method that can utilize existing production data to perform in-depth diagnosis of batteries and identify hidden anomalies in a low-cost and high-efficiency manner. Summary of the Invention

[0003] This invention provides a method and system for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation, in order to solve the problems mentioned in the background art.

[0004] A method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation is provided, including: Step A: Production process data acquisition, obtaining multi-dimensional production process data covering at least two processes in lithium battery production, forming a process data vector corresponding to the lithium battery; Step B: Prediction model training. Using historical process data vectors corresponding to multiple lithium batteries and the actual capacity values ​​corresponding to each lithium battery, a capacity prediction model based on the FT-Transformer (a deep learning architecture based on multi-head self-attention mechanism) architecture is trained. Step C: Predict the capacity of the battery to be diagnosed. Input the process data vector of the lithium battery to be diagnosed into the trained capacity prediction model to obtain the predicted capacity value of the lithium battery to be diagnosed. Step D: Anomaly diagnosis. Obtain the actual capacity value of the lithium battery to be diagnosed after the capacity grading process, calculate the absolute error between the predicted capacity value and the actual capacity value, and compare the absolute error with the preset anomaly judgment threshold. If the absolute error is greater than the anomaly judgment threshold, the lithium battery to be diagnosed is determined to be an abnormal battery. The capacity prediction model of the FT-Transformer architecture in step B has the following structure: Feature Tokenizer (feature tokenization module): Converts numerical and categorical features in process data vectors into feature embedding vectors of preset dimensions. Transformer encoder: It consists of at least one stacked Transformer layer. The Transformer layer contains a multi-head self-attention mechanism to handle the information interaction of feature embedding vectors. Prediction Head: CLSTokens (classification terms) are concatenated at the front of the feature embedding sequence to form the input of the Transformer encoder, and its final output is used for capacity prediction.

[0005] As a further technical solution of the present invention: the production process data of step A includes at least one or more of the following: material batch information of the assembly process, weight information of the liquid injection process, process voltage and capacity information of the formation process, charging capacity and equipment information of the capacity testing process, and open circuit voltage and internal resistance information of each stage.

[0006] As a further technical solution of the present invention: the production process data of step A also includes battery type, material batch number, weight before and after liquid injection, voltage and capacity data of the formation process, charging capacity of the capacity grading process, capacity grading equipment and channel number, environmental data and time data.

[0007] As a further technical solution of the present invention: the process data vector refers to the core process data used to characterize a single lithium battery in the entire production process, and is composed of two parts: original process parameters and feature engineering derived parameters, specifically as follows: 1. Original process parameters: a. Assembly process characteristics: including battery type and material batch number of positive and negative electrode materials; b. Characteristics of the electrolyte injection process: including the battery weight before electrolyte injection, the amount of electrolyte injected at the first time, the amount of electrolyte injected at the second time, the final weight of the finished battery, and the weight of the battery reduced during the formation process; c. Characteristics of the formation process: including the battery's initial voltage before formation, the battery's termination voltage after formation, the battery's charge and discharge capacity during each constant current / constant voltage stage of formation, the battery's temperature after formation, and the station number where the battery is located during formation. d. Characteristics of the capacity grading process: including the charge and discharge capacity of the battery during each constant current / constant voltage stage of the capacity grading process, the initial voltage and temperature of the battery before capacity grading, and the station number of the battery during capacity grading; e. Other electrochemical performance characteristics: including open-circuit voltage, AC internal resistance, and aging K value at each stage; Feature engineering derived parameters: a. Material batch inconsistency index (Stds): For cases where a single battery contains multiple cores, the material batch number of each core is converted into a specific date and then mapped to a time frame value, and its standard deviation is calculated; the specific calculation process is as follows: ; This feature is used to quantify the differences in material activity and coating uniformity between different cores within the same battery due to batch fluctuations in raw materials. b. Formation rest voltage drop: Calculate the difference between the end voltage of the first stage constant current charging and the start voltage of the second stage constant current charging in the formation process; used to capture the polarization recovery capability and interface stability of the cell during the short rest period; c. Equipment aging degree: Let d be the difference in the number of days between the battery's production date and the production line's official start-up date. Then:

[0008] This feature is used to capture the impact of latent environmental factors that accumulate over time, such as equipment wear and tear and fine-tuning of process parameters, on battery capacity. d. Number of batteries processed in the same batch: Considering that the number of cells processed in the same batch may affect the uniformity of the heat field distribution and current distribution inside the cabinet, thus affecting the performance of the cells, the number of batteries processed by each device in the same time period is calculated and bound to the corresponding batteries. As a further technical solution of the present invention: the step of uniformly converting the numerical features and categorical features in the process data vector into a feature embedding vector of a preset dimension includes: Numerical feature transformation: Each numerical feature By using a learnable weight vector Perform scalar multiplication expansion and add a bias vector. To generate embedding vectors The specific calculation formula is as follows:

[0009] Categorical feature transformation: Each categorical feature First, one-hot encoding is performed to obtain the vector. Then through a learnable embedding matrix Perform a search and add a bias vector. To generate embedding vectors The specific calculation formula is as follows.

[0010]

[0011] Ultimately, the converted Numerical features and Category features Stack them together to obtain the embedding matrix. The details are as follows: .

[0012] As a further technical solution of the present invention: the input of the Transformer encoder is a learnable [CLS] token concatenated at the front end of the feature embedding vector sequence, which is used to aggregate global information.

[0013] As a further technical solution of the present invention: the model training in step B takes minimizing the mean square error as the optimization objective, and the threshold for minimizing the mean square error is ±1%.

[0014] As a further technical solution of the present invention: the anomaly judgment threshold in step D is determined based on the statistical characteristics of the prediction error distribution of the capacity prediction model on the validation dataset, and the anomaly judgment threshold is ±1%.

[0015] As a further technical solution of the present invention: after step D determines that the lithium battery to be diagnosed is an abnormal battery, the method further includes an abnormal attribution analysis step: using the SHAP analysis tool, the feature importance of the process data vector of the lithium battery to be diagnosed is calculated to locate the key process or key process parameter that causes the absolute error between the predicted capacity value and the actual capacity value. The key process is selected from at least one of the assembly process, liquid injection process, formation process, and capacity testing process. The key process parameter is selected from at least one of the liquid injection volume, formation voltage, formation capacity, and capacity testing charging capacity.

[0016] A lithium battery manufacturing process anomaly diagnosis system, comprising: Data acquisition module: used for production process data acquisition, to obtain multi-dimensional production process data covering at least two processes in lithium battery production, and to form a process data vector corresponding to the lithium battery; Model storage module: Stores the capacity prediction model trained in step B; Calculation module: Performs the predicted capacity calculation in step C, and the absolute error calculation between the predicted capacity and the actual capacity in step D; Diagnostic module: Performs a comparison between the absolute error and the anomaly judgment threshold in step D, and determines the abnormal battery.

[0017] A lithium battery manufacturing process anomaly diagnosis system, comprising: Data acquisition module: Execute step A to collect lithium battery production process parameters and form a process data vector; Model storage module: Stores the capacity prediction model trained in step B; Calculation module: Performs the predicted capacity calculation in step C, and the absolute error calculation between the predicted capacity and the actual capacity in step D; Diagnostic module: Performs a comparison between the absolute error and the anomaly judgment threshold in step D, and determines the abnormal battery.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. Achieved early, proactive anomaly diagnosis based on capacity deviation: This method shifts the perspective of quality monitoring from "end-of-line inspection" to "process diagnosis." Even if a battery's final capacity is qualified, a significant deviation between its predicted and actual capacity suggests potential undetected fluctuations or compensation effects in its production process. This is of great value in identifying potential process consistency issues.

[0019] 2. Improved efficiency in problem tracing: By combining interpretability analysis tools such as SHAP, attribution analysis can be performed on batteries judged as "abnormal," making it easier to locate key processes or parameters that cause prediction deviations, and significantly improving the accuracy and efficiency of process problem investigation. 3. Strong technical adaptability: This method is based on existing production data for modeling, requiring no additional hardware or testing procedures, thus exhibiting good economic efficiency and feasibility. Furthermore, the powerful feature interaction learning capability of the FT-Transformer model allows it to adapt well to complex industrial data environments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the present invention.

[0022] Figure 2 This is a schematic diagram of the FT-Transformer model architecture of the present invention.

[0023] Figure 3 This is a SHAP summary diagram.

[0024] Figure 4 This is a schematic diagram of the analysis results using the SHAP tool.

[0025] Figure 5 This is a schematic diagram of the analysis results for a single battery. Detailed Implementation

[0026] 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, and 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.

[0027] Please see Figure 1 and Figure 2 This invention discloses a method for diagnosing anomalies in the lithium battery production process based on predicted capacity deviation, comprising the following steps: Step A: Production Process Data Acquisition. Obtain process parameters for a single lithium battery throughout the entire production process from the Manufacturing Execution System (MES) or relevant databases. These parameters cover multiple core processes from assembly, electrolyte injection, formation to capacity testing, and include both numerical data (such as weight, voltage, temperature, and time) and categorical data (such as equipment number, material batch, and battery model). These multi-dimensional parameters together constitute a process data vector that comprehensively describes the battery's production history.

[0028] Step B: Predictive Model Training. A large number of historically produced lithium-ion batteries with normal performance are used as training samples. Each sample consists of its process data vector and the actual capacity value obtained through capacity testing. Using these samples, a deep learning model based on the Feature Tokenizer-Transformer (FT-Transformer) architecture is trained. The core of this model lies in learning the complex nonlinear mapping relationship between high-dimensional, heterogeneous process data vectors and a single capacity scalar value.

[0029] Step C: Battery Capacity Prediction. For a newly produced lithium battery on the production line that needs diagnostics, first, its complete production process data vector is collected. Then, this vector is input into the capacity prediction model trained in Step B, and the model will output a predicted capacity value. This predicted value can be understood as the theoretical capacity that the battery "should" achieve based on its production process record.

[0030] Step D: Anomaly Diagnosis. After the battery completes the standard capacity testing process, its actual physical capacity value is measured. Next, the absolute error between the predicted capacity value given by the calculation model and the actual capacity value is calculated. Finally, this absolute error is compared with a pre-set anomaly judgment threshold. If the error exceeds the threshold, it indicates that the battery's actual performance significantly deviates from its "theoretical performance," and it can be judged as an "abnormal battery," triggering further engineering analysis or isolation treatment. Conversely, the battery's production process is considered stable and controllable. It should be noted that the anomaly judgment threshold can also be dynamically adjusted according to the quality requirements of different production batches and battery models, or determined in combination with other statistical methods (such as percentile method) to ensure the flexibility and accuracy of the diagnosis.

[0031] Example: This embodiment aims to diagnose whether there are any abnormalities in a batch of 280Ah lithium batteries during the production process, including the following steps: Step 1: Data Collection; Thirty-four key characteristic parameters of each battery cell during the production process were collected to form a process data vector. Specific parameters include: 1) Assembly process: battery type, A / B core positive / negative electrode batch number.

[0032] 2) Liquid injection process: pre-weighing, first / secondary liquid injection volume, and finished battery weight.

[0033] 3) Formation process: voltage before and after formation, charging capacity at each stage, voltage drop during formation rest, temperature at the end of formation, and formation equipment and channel number.

[0034] 4) Capacity grading process: capacity grading start time, capacity grading voltage, capacity grading charging capacity, capacity grading equipment and channel number.

[0035] 5) Others: production line running time, production season, working days, etc.

[0036] Step 2: Model training; Model training configuration a. Loss Function: Capacity prediction is a regression problem. We choose Mean Squared Error (MSE) as the loss function, as shown in the following formula:

[0037] in, It refers to the batch size. This is the actual capacity value. This is the capacity value predicted by the model. The prediction goal is to make the predicted capacity as close as possible to the actual capacity, that is, to find a set of parameters that... Minimum.

[0038] b. Optimizer selection: The AdamW optimizer was selected for model training.

[0039] An FT-Transformer model was trained using over 90,000 historical normal battery samples (process data vectors and their corresponding actual capacities). The model configuration used was as follows: 1) Model architecture: A 2-layer Transformer encoder is used, with each layer containing 8 attention heads.

[0040] 2) Embedding dimension: All input features are converted into 128-dimensional embedding vectors.

[0041] 3) Training configuration: The AdamW optimizer was used with a learning rate of 5e-5 and a batch size of 256. The goal was to minimize the mean squared error (MSELoss) and the training was conducted for a total of 150 epochs.

[0042] Step 3: Capacity prediction; A new battery to be diagnosed was selected, and its 34 key feature parameters were collected and input into the trained FT-Transformer model. The predicted capacity value output by the model was 298.5 Ah.

[0043] Step 4: Abnormal diagnosis; The battery under diagnosis underwent a standard capacity test, yielding an actual capacity of 292.2 Ah. The calculated absolute error was |298.5 Ah - 292.2 Ah| = 6.3 Ah. Assuming anomaly threshold (generally considered abnormal if the predicted capacity differs from the actual capacity by more than 1%) is set at 5.0 Ah, then 6.3 Ah > 5.0 Ah, and the battery is deemed abnormal. Although the battery's final capacity (292.2 Ah) may still be within the acceptable range, it is significantly lower than the theoretical value predicted by the model based on its production process data. This suggests that the battery may have experienced a negative, unrecorded disturbance at some stage of production (e.g., a minor internal short circuit, insufficient electrolyte wetting, or incomplete discharge during capacity testing), preventing it from reaching its full performance potential. This battery will be flagged, and its detailed process data will be submitted to process engineers for in-depth attribution analysis.

[0044] Use the SHAP tool for consistency diagnosis and anomaly attribution analysis. In the lithium battery production process, differences in the capacity grading station, fluctuations in ambient temperature, and changes in production time can all affect the capacity after grading. Quantitatively assessing the impact of various process parameters on battery capacity, and thus optimizing them accordingly, is a core challenge in quality control. These influencing factors are often intertwined and overlapping, making it difficult to effectively identify and quantify them using traditional univariate statistical methods such as linear regression. Therefore, based on the predictive model in this example, the SHAP (SHapley Additive exPlanations) interpretability analysis tool can be used to further analyze the factors affecting capacity.

[0045] Example 1: Global Feature Importance and Correlation Analysis To gain a macroscopic understanding of the impact of various process parameters on battery capacity prediction, this embodiment selects 500 random samples to calculate SHAP values ​​and generates a SHAP summary plot (Beeswarm Plot, e.g.) Figure 3 As shown in the figure, each point represents a feature of a sample, and the color indicates the level of the feature value (red represents a high value, and blue represents a low value). The analysis results show that the feature values ​​(color) of "Constant Current Constant Voltage Charging 1" and "OCV2" are significantly positively correlated with the SHAP value (horizontal axis position), that is, the higher the value of Constant Current Constant Voltage Charging 1 and OCV2, the higher the predicted discharge capacity.

[0046] Example 2: Single Feature Impact Analysis Taking the capacity grading equipment as an example, due to slight differences in temperature field, contact resistance, or calibration parameters between different devices or channels, the capacity of the same battery may differ after processing at different stations. Analysis using the SHAP tool yields the following results: Figure 4 The results are shown. The analysis reveals that the capacity prediction equipment in the "1-4-1 storage location" generally tends to produce a negative capacity prediction bias (lower SHAP value), with extreme biases of approximately 5Ah (ranging from -3Ah to +2Ah) caused by different workstations / channels. This analysis helps process engineers quickly locate abnormal machines exhibiting "systematic bias."

[0047] Example 3: Single Battery Analysis For specific samples encountered during production (such as batteries with abnormally high or low capacity), analysis can be performed on a per-cell basis. Taking a high-capacity battery sample as an example (e.g....) Figure 5As shown in the figure, the sample predicted capacity is 335.2A, compared to a baseline capacity of 298.951Ah. The charging capacity of 263.906Ah contributed +17.61Ah to the SHAP value, which is the main reason for the predicted value being much higher than the baseline (298.951Ah). This SHAP analysis for a single battery can help engineers determine whether the high capacity performance of the battery is due to sufficient actual active material (such as a large charging amount) or other measurement errors.

[0048] The method described in this application can be used not only for lithium iron phosphate batteries, but also for batteries based on similar principles such as sodium-ion batteries and ternary lithium-ion batteries. This application is not limited to the above-described embodiments.

[0049] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0050] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation, characterized in that, include; Step A: Production process data acquisition, obtaining multi-dimensional production process data covering at least two processes in lithium battery production, forming a process data vector corresponding to the lithium battery; Step B: Prediction model training. Using historical process data vectors corresponding to multiple lithium batteries and the actual capacity values ​​corresponding to each lithium battery, a capacity prediction model based on the FT-Transformer architecture is trained. Step C: Predict the capacity of the battery to be diagnosed. Input the process data vector of the lithium battery to be diagnosed into the trained capacity prediction model to obtain the predicted capacity value of the lithium battery to be diagnosed. Step D: Anomaly diagnosis. Obtain the actual capacity value of the lithium battery to be diagnosed after the capacity grading process, calculate the absolute error between the predicted capacity value and the actual capacity value, and compare the absolute error with the preset anomaly judgment threshold. If the absolute error is greater than the anomaly judgment threshold, the lithium battery to be diagnosed is determined to be an abnormal battery. The capacity prediction model of the FT-Transformer architecture in step B has the following structure: Feature word meta-conversion module: Converts numerical and categorical features in process data vectors into feature embedding vectors of preset dimensions; Transformer encoder: It consists of at least one stacked Transformer layer. The Transformer layer contains a multi-head self-attention mechanism to handle the information interaction of feature embedding vectors. Prediction Head: The CLSToken is concatenated at the front of the feature embedding sequence to form the input of the Transformer encoder, and its final output is used for capacity prediction.

2. The method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation according to claim 1, characterized in that, The production process data in step A includes at least one or more of the following: material batch information for the assembly process, weight information for the liquid injection process, process voltage and capacity information for the formation process, charging capacity and equipment information for the capacity testing process, and open circuit voltage and internal resistance information for each stage.

3. The method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation according to claim 1, characterized in that, The production process data in step A also includes battery type, material batch number, weight before and after liquid injection, voltage and capacity data during formation, charging capacity during capacity testing, capacity testing equipment and channel number, environmental data, and time data.

4. The method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation according to claim 1, characterized in that, The process data vector refers to the core process data used to characterize a single lithium battery in the entire production process. It is composed of two parts: original process parameters and feature engineering derived parameters, and its specific structure is as follows: Original process parameters: a. Assembly process characteristics: including battery type and material batch number of positive and negative electrode materials; b. Characteristics of the electrolyte injection process: including the battery weight before electrolyte injection, the amount of electrolyte injected at the first time, the amount of electrolyte injected at the second time, the final weight of the finished battery, and the weight of the battery reduced during the formation process; c. Characteristics of the formation process: including the battery's initial voltage before formation, the battery's termination voltage after formation, the battery's charge and discharge capacity during each constant current / constant voltage stage of formation, the battery's temperature after formation, and the station number where the battery is located during formation. d. Characteristics of the capacity grading process: including the charge and discharge capacity of the battery during each constant current / constant voltage stage of the capacity grading process, the initial voltage and temperature of the battery before capacity grading, and the station number of the battery during capacity grading; e. Other electrochemical performance characteristics: This includes the open-circuit voltage, AC internal resistance, and aging K-value at each stage; Feature engineering derived parameters: a. Material batch inconsistency index (Stds): For cases where a single battery contains multiple cores, the material batch number of each core is converted into a specific date and then mapped to a time frame value, and its standard deviation is calculated; the specific calculation process is as follows: ； This is used to quantify the differences in material activity and coating uniformity between different cores within the same battery due to batch fluctuations in raw materials. b. Formation rest voltage drop: Calculate the difference between the end voltage of the first stage constant current charging and the start voltage of the second stage constant current charging in the formation process; used to capture the polarization recovery capability and interface stability of the cell during the short rest period; c. Equipment aging degree: Let d be the difference in the number of days between the battery's production date and the production line's official start-up date. Then: Used to capture the impact of latent environmental factors that accumulate over time, such as equipment wear and tear and fine-tuning of process parameters, on battery capacity. d. Number of batteries processed in the same batch: Considering that the number of cells processed in the same batch may affect the uniformity of the heat field distribution and current distribution inside the cabinet, thus affecting the performance of the cells, the number of batteries processed by each device at the same time is calculated and bound to the corresponding batteries.

5. The method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation according to claim 4, characterized in that, The process data vector is transformed into a feature embedding vector of a preset dimension by uniformly converting both numerical and categorical features into such vectors. include: Numerical feature transformation: Each numerical feature By using a learnable weight vector Perform scalar multiplication expansion and add a bias vector. To generate embedding vectors The specific calculation formula is as follows: Categorical feature transformation: Each categorical feature First, one-hot encoding is performed to obtain the vector. Then through a learnable embedding matrix Perform a search and add a bias vector. To generate embedding vectors The specific calculation formula is as follows: Ultimately, the converted Numerical features and Category features Stack them together to obtain the embedding matrix. The details are as follows: 。 6. The method for diagnosing anomalies in lithium battery production process based on predicted capacity deviation according to claim 4, characterized in that, The input to the Transformer encoder is a learnable [CLS] token appended to the front of the feature embedding vector sequence, used to aggregate global information.

7. The method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation according to claim 1, characterized in that, The model training in step B aims to minimize the mean squared error, with a threshold of ±1%.

8. The method for diagnosing anomalies in lithium battery production process based on predicted capacity deviation according to claim 1, characterized in that, The anomaly detection threshold in step D is determined based on the statistical characteristics of the prediction error distribution of the capacity prediction model on the validation dataset, and the anomaly detection threshold is ±1%.

9. The method for diagnosing anomalies in lithium battery production processes based on predicted capacity deviation according to claim 1, characterized in that, After step D determines that the lithium battery to be diagnosed is an abnormal battery, the step further includes an abnormal attribution analysis step: using the SHAP analysis tool, the feature importance of the process data vector of the lithium battery to be diagnosed is calculated to locate the key process or key process parameter that causes the absolute error between the predicted capacity value and the actual capacity value. The key process is selected from at least one of the assembly process, liquid injection process, formation process, and capacity testing process. The key process parameter is selected from at least one of the liquid injection volume, formation voltage, formation capacity, and capacity testing charging capacity.

10. A lithium battery production process anomaly diagnosis system, characterized in that, include: Data acquisition module: used for production process data acquisition, to obtain multi-dimensional production process data covering at least two processes in lithium battery production, and to form a process data vector corresponding to the lithium battery; Model storage module: stores the capacity prediction model trained in step B of claim 1; Calculation module: performs the predicted capacity calculation in step C of claim 1, and the absolute error calculation between the predicted capacity and the actual capacity in step D; Diagnostic module: Performs the comparison between the absolute error and the anomaly judgment threshold in step D of claim 1, and determines the abnormal battery.

Images