Early warning and closed-loop control method and system based on deep learning

Abstract

The application provides a warning and closed-loop control method and system based on deep learning, and belongs to the field of new energy power batteries. Lithium precipitation and SEI film failure are detected by deep learning, physical constraints are combined with a deep learning model, so that the interpretability and detection accuracy are improved, and battery charging and discharging control and lithium precipitation pulse repair are realized. The method comprises the following steps: step (1), data acquisition and preprocessing; step (2), designing a self-encoder model; step (3), training the self-encoder model; step (4), data reconstruction prediction, and generating a warning signal; and step (5), adjusting operating parameters and implementing pulse repair.

Description

Technical Field

[0001] This application relates to the field of new energy power batteries, specifically proposing a deep learning-based early warning and closed-loop control method for improving the lifespan of lithium batteries, as well as a control system applying such a method. Background Technology

[0002] Currently, new energy vehicles have entered a stage of large-scale popularization globally, becoming a core track for addressing energy transition and "dual-carbon" goals. New energy vehicles are gaining an increasingly larger market share, with pure electric vehicles accounting for over 70%. As the power source of automobiles, the power battery plays an indispensable role; its safety, lifespan, and energy density directly affect the performance and driving experience of new energy vehicles. According to relevant regulations on the access of new energy vehicle manufacturers and products in China, power batteries must have fault warning and thermal runaway protection functions. However, the deficiencies of existing power battery management systems (BMS) in fault warning accuracy, lifespan prediction capabilities, and active control logic have become a key bottleneck restricting the high-quality development of the new energy vehicle industry.

[0003] A power battery is a multi-physics coupled system, making early warning and control extremely difficult. Some faults are difficult to detect, while others occur in milliseconds, often posing significant risks to life and property. During charging and discharging, chemical reactions easily occur at the electrode-electrolyte contact points. The solid electrolyte interphase (SEI) film and lithium deposition (lithium plating) are two core factors affecting the stability of the contact surface. Damage to the SEI and lithium plating can become major factors leading to battery life degradation and even serious safety accidents.

[0004] Existing technological solutions include: one is machine learning-based fault diagnosis technology. First, a large amount of battery operating data is collected and preprocessed, including cleaning, denoising, and normalization. Then, the data is divided into training and testing sets. The training set is used to train a machine learning model, including Support Vector Machines (SVM), Random Forests (RF), and Neural Networks (NN). Model parameters are adjusted to achieve optimal performance. The trained model is then written into the new energy vehicle processor, which collects battery operating parameters in real time and outputs a battery health status assessment result. For example, a previously published domestic patent, application number CN120142959A, uses a dual-cross-physical guidance framework to estimate SOC-SOH, but only focuses on state prediction and does not integrate fault diagnosis functionality. Another example is patent application number CN116087790A, which identifies faulty batteries based on autoencoder reconstruction errors, but lacks physical constraints and is prone to misjudgment due to data noise. Secondly, a fault mitigation scheme based on active balancing and optimized charging strategies specifically delays SEI film damage and lithium plating by dynamically adjusting charging and discharging parameters, thereby reducing fault causes from the operating condition perspective. The SEI film protection strategy dynamically adjusts the charging cut-off voltage based on battery temperature and SOH, activating staged charging when the battery temperature is too high. The lithium plating suppression strategy employs a low-temperature preheating and segmented stepped current charging mode.

[0005] The existing technologies described above have numerous shortcomings and deficiencies. Firstly, both traditional monitoring of electrochemical parameters and the use of machine learning or deep learning for battery monitoring and early warning lack a physical explanation. This results in insufficient reliability and low accuracy in fault diagnosis. Secondly, monitoring battery chemical parameters leads to delayed early warning responses. Electrochemical parameters only change after internal battery reactions have occurred. Furthermore, as battery life decreases, chemical parameters exhibit significant drift, increasing the false alarm rate for battery faults. Most technologies only provide fault warning functions and lack a closed-loop control logic encompassing warning, control, and repair.

[0006] In view of the above, this application is hereby submitted. Summary of the Invention

[0007] The early warning and closed-loop control method and system based on deep learning described in this application aim to solve the problems existing in the prior art. It uses deep learning to detect lithium plating and SEI film failure, combines physical constraints with deep learning models, thereby increasing interpretability and detection accuracy, and realizing battery charge and discharge control and pulse repair of lithium plating.

[0008] To achieve the above design objectives, the deep learning-based early warning and closed-loop control method includes the following steps:

[0009] Step (1): Data acquisition and preprocessing;

[0010] Battery operation data and gas data are collected by multi-source sensors, and derived parameters are calculated and the data is preprocessed.

[0011] Step (2): Design a physically constrained dual-decoder autoencoder model;

[0012] The design is capable of reconstructing and predicting data to obtain data reconstruction results, future state prediction results, and physical constraint dual-decoder autoencoder models of latent space vectors.

[0013] Step (2.1), design the encoder;

[0014] A dual-decoder structure including a reconstruction decoder and a prediction decoder is adopted;

[0015] Step (2.1.1): Design the parameters of the encoder's convolutional layer, LSTM layer, and fully connected layer;

[0016] Step (2.1.2), design of the physical constraint layer;

[0017] After passing through the LSTM layer, a high-dimensional feature vector is obtained, including the negative electrode overpotential. SEI film impedance change rate Internal resistance R i and single-unit voltage V c The instantaneous estimate is obtained; the abstract features above are mapped to intermediate physical quantities through a fully connected layer, and the physical constraint layer calculates based on the data output by the fully connected layer. and These two parameters are penalized through physical residual constraints;

[0018] Step (2.2), Dual decoder design;

[0019] This includes a dual decoder for design reconstruction and prediction, which can predict the future state of the battery while monitoring for battery anomalies;

[0020] Step (2.3): Design the loss function;

[0021] The loss function includes reconstruction error, prediction error, and physical residual loss;

[0022] Step (3), training the autoencoder model;

[0023] The Adam adaptive learning rate optimization algorithm is used to train the model. First, the reconstruction decoder is trained, and then the data prediction decoder is trained. If the conditions are not met within a specified period, training is stopped immediately and the parameters are adjusted. The parameters of the model are iteratively optimized using the training dataset.

[0024] Step (4): Data reconstruction and prediction to generate early warning signals;

[0025] A pre-trained physical constraint dual-decoder autoencoder model is used to reconstruct and predict the data to obtain the data reconstruction results, future state prediction results, and latent space vectors. Based on the gas data obtained in the above steps, the loss function of the autoencoder model, the prediction results, and the derived parameters, multi-level thresholds are set.

[0026] Step (5): Adjust operating parameters and perform pulse repair;

[0027] Based on the warning signal, the battery's operating parameters are adaptively adjusted or a pulse repair procedure is triggered.

[0028] Furthermore, step (1) includes the following steps:

[0029] Step (1.1): Design the acquisition frequency and calculate the derived data;

[0030] The collected parameters include the individual cell voltage V. c Total voltage V t Total current I t Battery temperature T, battery life (SOH), state of charge (SOC), SEI film impedance K s ;

[0031] Estimate the battery internal resistance R using the AC method. i And calculate the lithium plating sensitivity factor. The calculation formula is as follows:

[0032]

[0033] in, This refers to the capacity corresponding to a voltage exceeding 4.2V during charging. The total capacity during the entire charging process;

[0034] The formula for calculating SEI film growth parameters is as follows:

[0035]

[0036] in, for The slope of the curve, as described in the above formula for calculating SEI film growth indicators, characterizes... The rate of change of the slope of the curve;

[0037] Step (1.2): Use a gas sensor to detect the concentrations of hydrogen, carbon monoxide, and carbon dioxide around the battery, with a sampling frequency lower than that of the electrochemical signal;

[0038] Step (1.3), data preprocessing;

[0039] The data is aligned along the timeline, the dataset is divided, and standardization is performed.

[0040] The time axis of data from different acquisition frequencies is processed using interpolation or averaging methods;

[0041] Divide all data into training and validation sets;

[0042] The Z-score method is used to eliminate the dimensions of the data;

[0043] The data is divided using a sliding window method to adapt to the data processing method of the autoencoder.

[0044] Furthermore, in step (2.1.2), soft constraints of physical residuals are added to the loss function, and hard constraints of physical constraint layer are added to the decoder, so that the parameters in the latent space vector can be interpreted at the physical level.

[0045] Latent variables [ , [Including lithium plating risk index] and SEI membrane health ;

[0046] The formula for calculating the lithium plating risk index is as follows:

[0047]

[0048] in, This is the sensitivity coefficient; This represents the critical threshold for lithium plating. The overpotential at the negative electrode is calculated using the formula shown below.

[0049]

[0050] in, This refers to the voltage of a single unit; This is the open-circuit voltage; This is the operating current; This refers to the battery's internal resistance.

[0051] The formula for calculating the SEI health index is as follows:

[0052]

[0053] in, This is the time scaling factor; This represents the impedance value of the SEI film.

[0054] Furthermore, in step (2.2), the trained dual decoders perform collaborative early warning, with the reconstruction decoder providing rapid early warning and the prediction decoder serving as a supplement to the reconstruction decoder;

[0055] Step (2.2.1), design of the reconstruction decoder;

[0056] The reconstruction decoder receives data from the LSTM layer of the decoder for processing, sets the number of neurons in the fully connected layer, and expands the data structure after passing through the fully connected layer;

[0057] The data is reshaped to conform to the input structure of the deconvolution layer.

[0058] By deconvolution, the original data structure can be restored from the original data structure.

[0059] Finally, the original data dimensions are reconstructed through another convolutional layer;

[0060] Step (2.2.2), Design of the predictive decoder;

[0061] The predictive decoder processes the data in the latent space, using a repeating vector layer, an LSTM layer, and a time-distributed fully connected layer to predict battery voltage, temperature, and lithium plating risk index. And SEI membrane health.

[0062] Furthermore, in step (2.3), the expressions for the physical residual L1 and the prediction residual L2 in the loss function are as follows:

[0063]

[0064]

[0065] in, This represents the reconstruction error; This represents the prediction error; This is the physical residual loss;

[0066] Step 2.3.1, Reconstruction Error;

[0067] Reconstruction error includes data reconstruction error and regularization terms to prevent model overfitting;

[0068] The total reconstruction error is shown in the following formula:

[0069]

[0070] in, Mean square error, The encoder's raw input data, The data reconstructed from the reconstructed model, For regularization terms, The regularization coefficient is . For the parameters of the model, To reconstruct the weight coefficients corresponding to different parameters in the autoencoder;

[0071] Step (2.3.2), prediction error;

[0072] The prediction error is shown in the following formula:

[0073]

[0074] in, To predict the weight coefficients corresponding to different parameters in an autoencoder, The encoder's raw input data, To predict the data predicted by the autoencoder model;

[0075] Step (2.3.3), physical residual;

[0076] The physical residuals include lithium plating residuals and SEI film residuals, as shown in the following formula:

[0077]

[0078] in, These are lithium plating residual and SEI film residual, respectively. As a time factor, This is the negative overpotential. For SEI residual coefficients, The change rate of the SEI film impedance.

[0079] Furthermore, step (4) includes,

[0080] Step (4.1), safety protection based on gas sensors;

[0081] The gas sensor transmits its data directly to the server without going through a self-encoder;

[0082] Based on the chemical mechanisms of lithium plating and SEI film failure, hydrogen (H2) sensors, carbon monoxide (CO) sensors, and carbon dioxide (CO2) sensors are installed near the battery pack; concentration thresholds for hydrogen and carbon monoxide and warning thresholds for carbon dioxide are set; when the concentration of any gas exceeds the limit, the main circuit is immediately cut off and emergency heat dissipation is triggered.

[0083] Step (4.2): Setting early warning thresholds based on loss function and predicted data;

[0084] The loss function includes reconstruction residual, prediction residual, and physical residual. The prediction data output includes voltage, temperature, lithium plating risk index, and SEI film health. The original data includes negative electrode potential, SEI film growth index, and lithium plating sensitivity factor. Parameters related to lithium plating and SEI film failure are selected from these to set the warning threshold.

[0085] Step (4.2.1), Lithium plating threshold setting;

[0086] Prioritize reconstruction error and lithium plating sensitivity factors as early warning indicators, and then combine them with lithium plating residual and negative electrode overpotential for lithium plating verification;

[0087] Predicting future lithium plating risk using a lithium plating risk index involves selecting the following key parameters: reconfiguration error (L...). r ), lithium plating sensitive factor (a), negative electrode overpotential ( Lithium plating residual (L) p1 Lithium plating risk index (Z) l );

[0088] Step (4.2.2), SEI membrane failure threshold setting;

[0089] Reconstruction error and SEI film growth indicators are prioritized as early warning indicators, and then combined with SEI film residual and SEI film health to verify SEI film failure. The main parameters selected are: reconstruction error (L... r SEI film growth parameters, SEI film impedance change rate, SEI film residual (L) p2 SEI membrane health (Z) s ).

[0090] Furthermore, step (5) includes,

[0091] Step (5.1), Battery parameter control;

[0092] Lithium plating can be mitigated by controlling battery parameters during charging, and the damage to the SEI film caused by battery load can be reduced by controlling battery parameters during discharging.

[0093] Step (5.1.1): Control of battery parameters during charging;

[0094] During charging, the charging speed and battery health are coordinated and adjusted to control the cutoff voltage at different temperatures. The cutoff voltage is lowered at high temperatures and further lowered at low temperatures. The formula for calculating the cutoff voltage is:

[0095]

[0096] in, The reference cutoff voltage; This is the temperature compensation coefficient; This is the lithium plating risk compensation coefficient; This is the lithium plating sensitivity compensation coefficient; Battery temperature;

[0097] The charging current is controlled based on the battery capacity, gradually decreasing as the battery capacity increases; the cutoff voltage and charging current are dynamically adjusted according to the lithium plating index. The charging current calculation formula is as follows:

[0098]

[0099] in, This represents the initial charging current; SOC represents the battery capacity. The descending flow index; This is the current attenuation coefficient; Characterized the current decay factor, when As the current increases, the charging current decreases further;

[0100] Step (5.1.2): Control of battery parameters during discharge;

[0101] The formula for calculating the discharge current is:

[0102]

[0103] in, This is the rated discharge current; This is a weighting coefficient for health status; This is the weighting coefficient for impedance change; It is a constant to prevent division by zero errors;

[0104] Step (5.2), Pulse lithium plating repair control during charging;

[0105] When lithium plating is detected in the battery, pulse repair is performed when the battery capacity is between 60% and 100%; when the negative electrode overpotential ( If the value is greater than 0.1V, the lithium plating residual (L) p1 The lithium plating risk index (Z) is greater than 0.05. l Repair should be performed when the value is greater than 0.8.

[0106] Step (5.2.1), Lithium plating repair pulse settings;

[0107] The pulse period is set to 30 seconds, the pulse current to -0.5C, the discharge time to 200ms, and a 1-second rest period after each discharge. The chemical reaction formula is: ;

[0108] Step (5.2.2): Pulse amplitude adaptive adjustment;

[0109] The pulse period, pulse current magnitude, discharge time, and rest period are dynamically adjusted based on changes in gas concentration.

[0110] This application proposes a deep learning-based early warning and closed-loop control system, which includes the following execution modules:

[0111] The data acquisition and preprocessing module is used to acquire the individual cell voltage, total voltage, total current, temperature, SOH, SOC, and SEI film impedance of the battery, and to estimate the battery internal resistance, calculate the negative electrode overpotential, lithium plating sensitivity factor, and SEI film growth index. At the same time, it acquires the concentrations of hydrogen, carbon monoxide, and carbon dioxide through gas sensors; and performs time alignment, standardization, and sliding window division processing on the acquired multi-source heterogeneous data.

[0112] A physically constrained dual-decoder autoencoder model, comprising:

[0113] The encoder, consisting of a convolutional layer, an LSTM layer, and a physical constraint layer connected in sequence, is used to receive preprocessed data, extract features, and compress them to generate a latent space vector containing a lithium plating risk index and SEI film health.

[0114] A reconstruction decoder is used to reconstruct the latent space vector back into battery state data with the same dimension as the encoder input.

[0115] A predictive decoder is used to predict battery state parameters over a future period of time based on the latent space vector.

[0116] The loss function calculation module is used to calculate the total loss of the model, which includes the reconstruction error of the reconstruction decoder, the prediction error of the prediction decoder, and the physical residual loss based on the electrochemical mechanism.

[0117] The multi-level early warning module is used to determine the graded threshold based on gas concentration data, the loss function output value of the model, the predicted output value, and the calculated derived parameters, and to trigger early warning signals of different levels.

[0118] The closed-loop control module is used to respond to the warning signal, adaptively adjust the battery's charging and discharging parameters, and / or trigger the lithium plating pulse repair program.

[0119] This application proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the deep learning-based early warning and closed-loop control method.

[0120] This application proposes a computer-readable storage medium storing a computer program that, when executed, implements the deep learning-based early warning and closed-loop control method.

[0121] In summary, the deep learning-based early warning and closed-loop control method and system described above have the following advantages:

[0122] 1. This application provides a closed-loop detection and control scheme for improving the lifespan of lithium batteries. It utilizes deep learning to detect lithium plating and SEI film failure, combining physical constraints with a deep learning model, thereby significantly increasing interpretability and the accuracy and effectiveness of detection.

[0123] 2. This application uses a deep learning framework with dual decoders to perform fusion analysis of multiple battery parameters, thereby enabling non-destructive monitoring and early warning of SEI film damage and lithium plating level, while simultaneously realizing battery charge and discharge control and pulse repair of lithium plating.

[0124] 3. By adding a gas sensor, this application can solve the problem that existing technologies rely on electrochemical parameters for early warning, which lags behind chemical processes. This not only enables timely detection of the risk of thermal runaway and ensures battery safety, but also extends battery life and reduces the cost of battery replacement and recycling. Attached Figure Description

[0125] Figure 1 This is a flowchart of the deep learning-based early warning and closed-loop control method.

[0126] Figure 2 This is a schematic diagram of the early warning and closed-loop control system.

[0127] Figure 3 A schematic diagram of the physical residual soft constraints for reconstructing the decoder;

[0128] Figure 4 A schematic diagram of the hard constraints of the physical constraint layer and the soft constraints of the physical residual for the predictive decoder;

[0129] Figure 5 This is a schematic diagram of the model training results;

[0130] Figure 6 A schematic diagram of a closed-loop architecture for early warning, prediction, control, and repair; Detailed Implementation

[0131] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of this application; however, this application may be implemented in other ways than those described herein, and therefore, this application is not limited to the specific embodiments disclosed below.

[0132] Example 1, such as Figure 1As shown, this application proposes a deep learning-based early warning and closed-loop control method, constructing a battery state sensing center based on a "Physical Embedding-Deep Learning Fusion Model (PCAE)". This center uses a unique "hardware and software combined" physical constraint mechanism to ensure high-precision and interpretable monitoring of battery lithium plating and SEI film failure states, deeply integrating deep learning with electrochemical mechanisms, demonstrating strong cross-disciplinary innovation. Multi-source data fusion (electrochemical parameters + gas sensors) is introduced to improve the real-time performance and accuracy of early warning. A hierarchical early warning mechanism and closed-loop control strategy are designed, including pulse repair and adaptive charge-discharge control, forming a complete "monitoring-early warning-control-repair" closed-loop system. Based on this, a dual-decoder architecture is used to achieve collaborative early warning of "real-time diagnosis" and "future prediction", and multi-source fusion decision-making greatly reduces the false alarm rate. Specifically, the method includes the following implementation steps:

[0133] Step (1): Data acquisition and preprocessing;

[0134] Battery operation data and gas data are collected by multi-source sensors, and derived parameters are calculated and the data is preprocessed.

[0135] Step (1.1): Design the acquisition frequency and calculate the derived data;

[0136] The collected parameters include the individual cell voltage V. c Total voltage V t Total current I t Battery temperature T, battery life (SOH), state of charge (SOC), SEI film impedance K s ;

[0137] For data such as individual voltage that changes rapidly, a higher frequency is used for acquisition; for data such as total current and total voltage, a medium frequency is used for acquisition; and for data such as temperature that changes slowly, a lower frequency is used for acquisition.

[0138] Estimate the battery internal resistance R using the AC method. i And calculate the lithium plating sensitivity factor. The calculation formula is as follows:

[0139]

[0140] in, This refers to the capacity corresponding to a voltage exceeding 4.2V during charging. The total capacity during the entire charging process;

[0141] The formula for calculating SEI film growth parameters is as follows:

[0142]

[0143] in, for The slope of the curve, as described in the above formula for calculating SEI film growth indicators, characterizes... The rate of change of the slope of the curve;

[0144] Step (1.2): Use a gas sensor to detect the concentrations of hydrogen, carbon monoxide, and carbon dioxide around the battery, with a sampling frequency lower than that of the electrochemical signal;

[0145] Step (1.3), data preprocessing;

[0146] The data is aligned along the timeline, the dataset is divided, and standardization is performed.

[0147] The time axis of data from different acquisition frequencies is processed using interpolation or averaging methods;

[0148] Divide all data into a training set and a validation set in a 6:4 ratio;

[0149] The Z-score method is used to eliminate the dimensions of the data;

[0150] A sliding window method is used to divide the data to adapt to the data processing method of the autoencoder; for example, the window size is set to 500, the stride is 10, and aggregation processing is performed; the window covers 500 seconds of data, and slides 10 data points per second; the final data is... , where m is the number of features.

[0151] Step (2): Design the autoencoder model;

[0152] The design is capable of reconstructing and predicting data to obtain data reconstruction results, future state prediction results, and physical constraint dual-decoder autoencoder models of latent space vectors. Specifically, by adding a physical constraint layer to the existing autoencoder, the parameters in the latent space vectors are explained by electrochemical physical laws. Furthermore, lithium plating residuals and SEI film residuals are added to the loss function of the autoencoder, so that the autoencoder is deeply integrated with the physical constraints.

[0153] Step (2.1), design the encoder;

[0154] A dual-decoder structure including a reconstruction decoder and a prediction decoder is adopted;

[0155] Step (2.1.1): Design the parameters of the encoder's convolutional layer, LSTM layer, and fully connected layer;

[0156] The single-unit voltage V c Total voltage V t Total current I tThe battery temperature (T), battery life (SOH), and battery charge (SOC) are input into the convolutional layer. The number of filters is 64, the filter size is 3, the stride is 1, and the same padding is used.

[0157] The input data structure of the convolutional layer is (64, 300, 6), and the output structure is (64, 300, 64).

[0158] Define the number of neurons in the LSTM layer as 128, the structure of the input data of the LSTM layer as (64, 300, 64), and the structure of the output data of the LSTM layer as (64, 128).

[0159] The output of the LSTM layer is mapped to a two-dimensional vector through a fully connected layer. , The output data structure is (64, 2);

[0160] Step (2.1.2), design of the physical constraint layer;

[0161] After passing through the LSTM layer, a high-dimensional feature vector with a structure of (64, 128) is obtained. This vector represents the data characteristics of the battery parameters, including the negative electrode overpotential. SEI film impedance change rate Internal resistance R i and single-unit voltage V c Instantaneous estimates, etc.;

[0162] The abstracted features are mapped to intermediate physical quantities through a fully connected layer, and the physical constraint layer calculates based on the data output by the fully connected layer. and The two parameters are penalized by physical residual constraints; specifically, soft constraints of physical residuals are added to the loss function, and hard constraints of physical constraint layers are added to the decoder, so that the parameters in the latent space vector can be interpreted at the physical level.

[0163] Latent variables [ , [Including lithium plating risk index] and SEI membrane health ;

[0164] The formula for calculating the lithium plating risk index is as follows:

[0165]

[0166] in, This is the sensitivity coefficient; This represents the critical threshold for lithium plating. The overpotential at the negative electrode is calculated using the formula shown below.

[0167]

[0168] in, This refers to the voltage of a single unit; This is the open-circuit voltage; This is the operating current; This refers to the battery's internal resistance.

[0169] The formula for calculating the SEI health index is as follows:

[0170]

[0171] in, This is the time scaling factor; This refers to the impedance value of the SEI film.

[0172] Step (2.2), Dual decoder design;

[0173] This includes a dual decoder for design reconstruction and prediction, which can predict the future state of the battery while monitoring for battery anomalies;

[0174] The trained dual decoders perform collaborative early warning; the reconstruction decoder can provide rapid early warning, while the prediction decoder can supplement the reconstruction decoder.

[0175] Step (2.2.1), design of the reconstruction decoder;

[0176] The reconstruction decoder receives data from the LSTM layer of the decoder and processes it. The input data structure is (64, 128). Through reconstruction, the data is restored to (64, 300, 6).

[0177] The number of neurons in the fully connected layer is set to 512. After passing through the fully connected layer, the data structure is expanded to (64, 512).

[0178] The data is reshaped to conform to the input structure of the deconvolution layer; the reshape layer transforms the data structure into (64, 16, 32). The stride of the deconvolution layer is set to 2, the kernel size to 5, and the number of filters to 64.

[0179] By deconvolution, the data structure is restored to (64, 32, 64).

[0180] Finally, the original data dimensions are reconstructed by another convolutional layer. The kernel size of the convolutional layer is set to 3, the number of filters is 6, and the data structure is restored to (64, 300, 6).

[0181] Step (2.2.2), Design of the predictive decoder;

[0182] The predictive decoder processes the data in the latent space, using a repeating vector layer, an LSTM layer, and a time-distributed fully connected layer to predict the battery's voltage, temperature, and lithium plating risk index over 300 seconds. and SEI membrane health The four features of the output predicted value are [U, T, ... , ];

[0183] The input data structure is (64, 2), and after being repeated 30 times through a repeating vector layer, the output data structure is (64, 30, 2).

[0184] Define the number of neurons in the LSTM layer as 128. After passing through the LSTM layer, the data structure becomes (64, 30, 128).

[0185] The time-distributed fully connected layer has 4 neurons, with an input data structure of (64, 30, 128) and an output data structure of (64, 30, 4). The four features of the output predicted value are [U, T, ...]. , ];

[0186] Step (2.3): Design the loss function;

[0187] Existing battery early warning methods can only learn the data features of electrochemical parameters through a loss function penalty to achieve early warning. This application, by adding physical constraints and a dual-decoder design, enables the entire training process to capture the physical laws of the battery and achieves a deep integration of monitoring and prediction. Therefore, the loss function includes reconstruction error, prediction error, and physical residuals related to lithium plating and SEI film failure.

[0188] The loss function is an evaluation mechanism for autoencoders, used to measure the degree to which autoencoders learn data features, and serves as an important indicator for lithium plating and SEI film failure early warning. It mainly includes reconstruction error, prediction error, and physical residual loss.

[0189] Step (2.3.1), reconstruction error;

[0190] The reconstruction error comprises two parts: the data reconstruction error itself and a regularization term to prevent model overfitting. The reconstruction error involves calculating the interpolation between the original input data and the reconstructed data, including the unit voltage V. c Total voltage V t Total current I t The battery has six parameters: battery temperature (T), battery life (SOH), and battery capacity (SOC).

[0191] The total reconstruction error is shown in the following formula:

[0192]

[0193] in, Mean square error, The encoder's raw input data, The data reconstructed from the reconstructed model, For regularization terms, The regularization coefficient is . For the parameters of the model, To reconstruct the weight coefficients corresponding to different parameters in the autoencoder;

[0194] Step (2.3.2), prediction error;

[0195] The predictor decoder obtains a total of [U, T, ...] after prediction. , The values ​​of the four parameters, specifically, , and Calculate U and T respectively from the original data used for training data in the next 300 seconds of data. In latent space vectors , Perform calculations;

[0196] The prediction error is shown in the following formula:

[0197]

[0198] in, To predict the weight coefficients corresponding to different parameters in an autoencoder, The encoder's raw input data, To predict the data predicted by the autoencoder model;

[0199] Step (2.3.3), physical residual loss;

[0200] When designing the physical residuals, the rate of change of the negative electrode overpotential and SEI film impedance due to lithium plating and SEI film failure is considered, and a penalty latent space vector is applied. , The goal is to achieve self-encoder conforming to physical constraints;

[0201] The physical residuals include lithium plating residuals and SEI film residuals, as shown in the following formula:

[0202]

[0203] in, These are lithium plating residual and SEI film residual, respectively. As a time factor, This is the negative overpotential. For SEI residual coefficients, The rate of change of SEI film impedance;

[0204] Step (2.3.4), Reconstruction Loss and Predicted Loss

[0205] The reconstruction decoder and the prediction decoder have different roles, and different loss functions are used for training the model. The reconstruction loss is used instead of the prediction loss. Train the reconstruction decoder using prediction loss. The reconstruction loss and prediction loss are trained as follows:

[0206]

[0207]

[0208] in, This represents the reconstruction error; This represents the prediction error; This is the physical residual loss;

[0209] Step (3), training the autoencoder model;

[0210] The Adam adaptive learning rate optimization algorithm is used to train the model. First, the reconstruction decoder is trained, and then the data prediction decoder is trained. If the condition is not met within 50 cycles, training is stopped immediately and the parameters are adjusted.

[0211] Specifically, the parameters of the model are iteratively optimized using the training dataset;

[0212] The reconstruction decoder minimizes the reconstruction loss. During training, the prediction decoder is trained by minimizing the prediction loss. Conduct training; For reconstruction error, For prediction error, This is the physical residual loss;

[0213] The dataset divided in step (1.3) is used; when the value of the loss function meets the required value c, the model training ends immediately.

[0214] like Figure 5 As shown, this application can verify the training results of the autoencoder of the proposed physical constraint dual decoder using an open-source dataset. The curve in the figure is the loss function change curve obtained during the training process.

[0215] The training process of the physical constraint dual-decoder autoencoder model is as follows:

[0216] Step (4): Data reconstruction and prediction to generate early warning signals;

[0217] A pre-trained physical constraint dual decoder autoencoder model is used to reconstruct and predict the data, so as to obtain the data reconstruction results, future state prediction results and latent space vector.

[0218] Based on the gas data obtained from the above steps, the loss function of the autoencoder model, the prediction results, and the derived parameters, multi-level thresholds are set.

[0219] Existing battery warning technologies often target obvious anomalies in battery parameters or extreme situations such as accidents, failing to clearly display the abnormal state of the battery. To address this, this application establishes a tiered warning mechanism through multi-source data fusion. This includes...

[0220] The first level is the extreme case of battery thermal runaway. It is judged based on the data of gas sensors. When the gas concentration is abnormal, measures are taken immediately to prevent or delay the occurrence of the accident.

[0221] The second level uses the traditional threshold method to control the overpotential of the negative electrode. Monitoring is performed on the lithium plating sensitivity factor α, and data-driven compensation is used. Data obtained directly or indirectly can be quickly screened, and verified by combining data obtained from deep learning, thereby further improving the accuracy of early warning.

[0222] The third level is prediction and early warning, which achieves early detection and control by analyzing prediction data.

[0223] A tiered early warning mechanism can quantify risk levels, organically integrate battery parameters, and uncover complex relationships between them. It can handle not only sudden risks but also potential risks, achieving accurate early warning. Since changes in chemical gas concentration are more pronounced than changes in electrochemical parameters during lithium plating and SEI film failure, early warning is initially based on abnormal gas concentration, supplemented by raw data, loss function, and predicted data. The protection mechanism is unconditionally activated when an abnormal gas concentration occurs. Thresholds for lithium plating and SEI film failure warnings are set based on raw data, loss function, and the output of the predictive decoder to protect against non-emergency conditions. The specific steps are as follows:

[0224] Step (4.1), safety protection based on gas sensors;

[0225] The gas sensor transmits its data directly to the server without going through a self-encoder;

[0226] Based on the chemical mechanisms of lithium plating and SEI film failure, hydrogen (H2), carbon monoxide (CO), and carbon dioxide (CO2) sensors are installed near the battery pack. Concentration thresholds for hydrogen and carbon monoxide, and a warning threshold for carbon dioxide, are set; when the concentration of any gas exceeds the limit, the main circuit is immediately cut off and emergency cooling is triggered.

[0227] Step (4.2): Setting early warning thresholds based on loss function and predicted data;

[0228] The loss function includes reconstruction residual, prediction residual, and physical residual. The prediction data output includes voltage, temperature, lithium plating risk index, and SEI film health. The original data includes negative electrode potential, SEI film growth index, lithium plating sensitivity factor, etc. From these, parameters related to lithium plating and SEI film failure are selected to set the warning threshold.

[0229] Step (4.2.1), Lithium plating threshold setting;

[0230] Prioritize reconstruction error and lithium plating sensitivity factors as early warning indicators, and then combine them with lithium plating residual and negative electrode overpotential for lithium plating verification;

[0231] Predicting future lithium plating risk using a lithium plating risk index involves selecting the following key parameters: reconfiguration error (L...). r ), lithium plating sensitive factor (a), negative electrode overpotential ( Lithium plating residual (L) p1 Lithium plating risk index (Z) l );

[0232] Step (4.2.2), SEI membrane failure threshold setting;

[0233] Reconstruction error and SEI film growth indicators are prioritized as early warning indicators, and then combined with SEI film residual and SEI film health to verify SEI film failure. The main parameters selected are: reconstruction error (L... r SEI film growth parameters, SEI film impedance change rate, SEI film residual (L) p2 SEI membrane health (Z) s );

[0234] Step (5): Adjust operating parameters and perform pulse repair;

[0235] Based on the warning signal, the battery's operating parameters are adaptively adjusted or a pulse repair program is triggered;

[0236] Unlike existing technologies, this application directly applies the parameters characterizing the battery state to the control or even repair of battery parameters, that is, to control battery parameters based on early warning information and repair lithium plating when necessary.

[0237] Therefore, using early warning data for battery parameter control and lithium plating repair, and adaptively adjusting control parameters based on collected data, enables adaptive proactive control and repair. This achieves a complete closed-loop system encompassing parameter monitoring, early warning, control, and repair. It not only provides early warning but also preventative measures, and extends battery life through pulse repair technology, thereby reducing costs.

[0238] Step (5.1), Battery parameter control;

[0239] Lithium plating can be mitigated by controlling battery parameters during charging, and the damage to the SEI film caused by battery load can be reduced by controlling battery parameters during discharging.

[0240] Step (5.1.1): Control of battery parameters during charging;

[0241] The charging process coordinates the charging speed and battery health, thus adjusting the lithium plating risk index. Lithium plating sensitive factor As a reference for controlling charging parameters, it enables more precise control of parameters for high and low risks;

[0242] Specifically, the cutoff voltage is controlled at different temperatures; the cutoff voltage is reduced at high temperatures and further reduced at low temperatures. The formula for calculating the cutoff voltage is:

[0243]

[0244] in, The reference cutoff voltage; This is the temperature compensation coefficient; This is the lithium plating risk compensation coefficient; This is the lithium plating sensitivity compensation coefficient; Battery temperature;

[0245] The charging current is controlled based on the battery capacity, gradually decreasing as the battery capacity increases; the cutoff voltage and charging current are dynamically adjusted according to the lithium plating index. The charging current calculation formula is as follows:

[0246]

[0247] in, This represents the initial charging current; SOC represents the battery capacity. The descending flow index; This is the current attenuation coefficient; Characterized the current decay factor, when As the current increases, the charging current decreases further;

[0248] Step (5.1.2): Control of battery parameters during discharge;

[0249] During discharge, the primary focus is on mitigating the damage to the SEI film caused by the battery load. The cutoff voltage and discharge current are adjusted based on the SEI film's health. By incorporating SEI film health and impedance change rate, the discharge current is dynamically regulated. When the system detects accelerated SEI film aging or instability, it automatically limits the discharge power to reduce the impact of high-current discharge on the SEI film, thus providing a "protective discharge." The discharge current calculation formula is as follows:

[0250]

[0251] in, This is the rated discharge current; This is a weighting coefficient for health status; This is the weighting coefficient for impedance change; It is a constant to prevent division by zero errors;

[0252] Step (5.2), Pulse lithium plating repair control during charging;

[0253] When lithium plating is detected in the battery, pulse repair is performed when the battery capacity is between 60% and 100%; when the negative electrode overpotential ( If the value is greater than 0.1V, the lithium plating residual (L) p1 The lithium plating risk index (Z) is greater than 0.05. l Repair should be performed when the value is greater than 0.8.

[0254] Step (5.2.1), Lithium plating repair pulse settings;

[0255] The pulse period is set to 30 seconds, the pulse current to -0.5C, the discharge time to 200ms, and a 1-second rest period after each discharge. The chemical reaction formula is: ;

[0256] Step (5.2.2): Pulse amplitude adaptive adjustment;

[0257] The pulse period, pulse current magnitude, discharge time, and rest period are dynamically adjusted according to changes in gas concentration.

[0258] If the H2 or CO concentration shows an upward trend after the application of the repair pulse, it indicates that the repair process may have produced side reactions or is not effective. The system should reduce the pulse current or extend the settling time. If the gas concentration remains stable or decreases, it indicates that the repair measures are effective and can be continued with the current parameters or slightly increased intensity.

[0259] Based on the aforementioned deep learning-based early warning and closed-loop control method, this application proposes the following deep learning-based early warning and closed-loop control system. This system transforms early warning information into adaptive control commands and intelligent repair strategies, thereby forming a complete ecological closed loop of "monitoring-early warning-control-repair". Therefore, it can effectively and significantly extend battery life and generate huge economic and environmental benefits.

[0260] The system includes:

[0261] The data acquisition and preprocessing module is used to acquire the individual cell voltage, total voltage, total current, temperature, SOH, SOC, and SEI film impedance of the battery, and to estimate the battery internal resistance, calculate the negative electrode overpotential, lithium plating sensitivity factor, and SEI film growth index. At the same time, it acquires the concentrations of hydrogen, carbon monoxide, and carbon dioxide through gas sensors; and performs time alignment, standardization, and sliding window division processing on the acquired multi-source heterogeneous data.

[0262] A physically constrained dual-decoder autoencoder model, comprising:

[0263] The encoder, consisting of a convolutional layer, an LSTM layer, and a physical constraint layer connected in sequence, is used to receive preprocessed data, extract features, and compress them to generate a latent space vector containing a lithium plating risk index and SEI film health.

[0264] A reconstruction decoder is used to reconstruct the latent space vector back into battery state data with the same dimension as the encoder input.

[0265] A predictive decoder is used to predict battery state parameters over a future period of time based on the latent space vector.

[0266] The loss function calculation module is used to calculate the total loss of the model, which includes the reconstruction error of the reconstruction decoder, the prediction error of the prediction decoder, and the physical residual loss based on the electrochemical mechanism.

[0267] The multi-level early warning module is used to determine the graded threshold based on gas concentration data, the loss function output value of the model, the predicted output value, and the calculated derived parameters, and to trigger early warning signals of different levels.

[0268] The closed-loop control module is used to respond to the warning signal, adaptively adjust the battery's charging and discharging parameters, and / or trigger the lithium plating pulse repair program.

[0269] Furthermore, the physical constraint layer maps the output of the LSTM layer into a two-dimensional vector. , ],in The lithium plating risk index is calculated using the following formula: , The formula for calculating SEI membrane health is as follows: ;in, This is the sensitivity coefficient; This represents the critical threshold for lithium plating. This is the negative overpotential. This is the time scaling factor; The impedance value of the SEI film. The change rate of the SEI film impedance.

[0270] The loss function calculation module mentioned above, its physical residual loss Including lithium plating residual and SEI membrane residual The calculation formulas are as follows: and ,in and These are the weighting coefficients.

[0271] The multi-level early warning module includes:

[0272] The first-level early warning unit is used to immediately trigger the highest level alarm and cut off the main circuit when the concentration of hydrogen or carbon monoxide exceeds 50 ppm, or the change in carbon dioxide concentration exceeds 50% VOL.

[0273] The second-level early warning unit is used to make threshold judgments based on reconstruction error, lithium plating sensitivity factor, negative electrode overpotential, lithium plating residual, SEI film growth index, SEI film impedance change rate, and SEI film residual.

[0274] The third-level early warning unit is used to make threshold judgments based on the future lithium plating risk index and SEI film health output by the predictive decoder.

[0275] The closed-loop control module includes:

[0276] The charging parameter control unit dynamically adjusts the charging cut-off voltage based on battery temperature and lithium plating risk index. and charging current The calculation formulas are as follows: and ,in The reference cutoff voltage; This is the temperature compensation coefficient; For battery temperature, This represents the initial charging current; SOC represents the battery capacity. The descending flow index;

[0277] The pulse repair control unit is used to apply a negative pulse current for repair when lithium plating is detected and the battery SOC is between 60% and 100%, and can adaptively adjust the pulse parameters according to changes in gas concentration.

[0278] This application also proposes a novel electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned deep learning-based early warning and closed-loop control method.

[0279] This application also proposes a novel computer-readable storage medium storing a computer program that, when executed, enables the implementation of the aforementioned deep learning-based early warning and closed-loop control method.

[0280] As described above, similar technical solutions can be derived from the solutions presented in the accompanying drawings and description, and all of them still fall within the scope of the claims of this application.

Claims

1. A deep learning-based early warning and closed-loop control method, characterized in that: Includes the following steps, Step (1): Data acquisition and preprocessing; Battery operation data and gas data are collected by multi-source sensors, and derived parameters are calculated and the data is preprocessed. Step (2): Design the autoencoder model; The design is capable of reconstructing and predicting data to obtain data reconstruction results, future state prediction results, and physical constraint dual-decoder autoencoder models of latent space vectors. Step (2.1), design the encoder; A dual-decoder structure including a reconstruction decoder and a prediction decoder is adopted; Step (2.1.1): Design the parameters of the encoder's convolutional layer, LSTM layer, and fully connected layer; Step (2.1.2), design of the physical constraint layer; After passing through the LSTM layer, a high-dimensional feature vector is obtained, including the negative electrode overpotential. SEI film impedance change rate Internal resistance R i and single-unit voltage V c The instantaneous estimate is obtained; the intermediate physical quantity is mapped from the high-dimensional feature vectors through a fully connected layer, and the physical constraint layer calculates the value based on the data output by the fully connected layer. and These two parameters are penalized through physical residual constraints; Soft constraints of physical residuals are added to the loss function, and hard constraints of physical constraint layers are added to the decoder, so that the parameters in the latent space vector can be interpreted at the physical level. Latent variables [ , [Including lithium plating risk index] and SEI membrane health ; The formula for calculating the lithium plating risk index is as follows: in, This is the sensitivity coefficient; This represents the critical threshold for lithium plating. The overpotential at the negative electrode is calculated using the formula shown below. in, This refers to the voltage of a single unit; This is the open-circuit voltage; This is the operating current; This refers to the battery's internal resistance. The formula for calculating the SEI health index is as follows: in, This is the time scaling factor; This refers to the impedance value of the SEI film. Step (2.2), Dual decoder design; This includes a dual decoder for design reconstruction and prediction, which can predict the future state of the battery while monitoring for battery anomalies; Step (2.3): Design the loss function; The loss function includes reconstruction error, prediction error, and physical residual loss; Step (3), training the autoencoder model; The Adam adaptive learning rate optimization algorithm is used to train the model. First, the reconstruction decoder is trained, and then the data prediction decoder is trained. If the conditions are not met within a specified period, training is stopped immediately and the parameters are adjusted. The parameters of the model are iteratively optimized using the training dataset. Step (4): Data reconstruction and prediction to generate early warning signals; A pre-trained physical constraint dual decoder autoencoder model is used to reconstruct and predict the data, so as to obtain the data reconstruction results, future state prediction results and latent space vector. Based on the gas data obtained from the above steps, the loss function of the autoencoder model, the prediction results, and the derived parameters, multi-level thresholds are set. Step (5): Adjust operating parameters and perform pulse repair; Based on the warning signal, the battery's operating parameters are adaptively adjusted or a pulse repair procedure is triggered.

2. The early warning and closed-loop control method based on deep learning according to claim 1, characterized in that: The steps described Step (1) includes the following steps: Step (1.1): Design the acquisition frequency and calculate the derived data; The collected parameters include the individual cell voltage V. c Total voltage V t Total current I t Battery temperature T, battery life SOH, state of charge SOC, SEI film impedance K s ; Estimate the battery internal resistance R using the AC method. i And calculate the lithium plating sensitivity factor. The calculation formula is as follows: in, This refers to the capacity corresponding to a voltage exceeding 4.2V during charging. The total capacity during the entire charging process; The formula for calculating SEI film growth parameters is as follows: in, for The slope of the curve, as described in the above formula for calculating SEI film growth indicators, characterizes... The rate of change of the slope of the curve; Step (1.2): Use a gas sensor to detect the concentrations of hydrogen, carbon monoxide, and carbon dioxide around the battery, with a sampling frequency lower than that of the electrochemical signal; Step (1.3), data preprocessing; The data is aligned along the timeline, the dataset is divided, and standardization is performed. The time axis of data from different acquisition frequencies is processed using interpolation or averaging methods; Divide all data into training and validation sets; The Z-score method is used to eliminate the dimensions of the data; The data is divided using a sliding window method to adapt to the data processing method of the autoencoder.

3. The deep learning-based early warning and closed-loop control method according to claim 1, characterized in that: The steps described Step (2.2): The trained dual decoders perform collaborative early warning, with the reconstruction decoder providing rapid early warning and the prediction decoder serving as a supplement to the reconstruction decoder; Step (2.2.1), design of the reconstruction decoder; The reconstruction decoder receives data from the LSTM layer of the decoder for processing, sets the number of neurons in the fully connected layer, and expands the data structure after passing through the fully connected layer; The data is reshaped to conform to the input structure of the deconvolution layer. By deconvolution, the original data structure can be restored from the original data structure. Finally, the original data dimensions are reconstructed through another convolutional layer; Step (2.2.2), Design of the predictive decoder; The predictive decoder processes the data in the latent space, using a repeating vector layer, an LSTM layer, and a time-distributed fully connected layer to predict battery voltage, temperature, and lithium plating risk index. And SEI membrane health.

4. The deep learning-based early warning and closed-loop control method according to claim 1, characterized in that: The In step (2.3), the expressions for the physical residual L1 and the prediction residual L2 in the loss function are as follows: in, This represents the reconstruction error; This represents the prediction error; This is the physical residual loss; Step 2.3.1, Reconstruction Error; Reconstruction error includes data reconstruction error and regularization terms to prevent model overfitting; The total reconstruction error is shown in the following formula: in, Mean square error, The encoder's raw input data, The data reconstructed from the reconstructed model, For regularization terms, The regularization coefficient is . For the parameters of the model, To reconstruct the weight coefficients corresponding to different parameters in the autoencoder; Step (2.3.2), prediction error; The prediction error is shown in the following formula: in, To predict the weight coefficients corresponding to different parameters in an autoencoder, The encoder's raw input data, To predict the data predicted by the autoencoder model; Step (2.3.3), physical residual; The physical residuals include lithium plating residuals and SEI film residuals, as shown in the following formula: in, These are lithium plating residual and SEI film residual, respectively. As a time factor, This is the negative overpotential. For SEI residual coefficients, The change rate of the SEI film impedance.

5. The deep learning-based early warning and closed-loop control method according to claim 1, characterized in that: Step (4) includes, Step (4.1), safety protection based on gas sensors; The gas sensor transmits its data directly to the server without going through a self-encoder; Based on the chemical mechanisms of lithium plating and SEI film failure, hydrogen, carbon monoxide, and carbon dioxide sensors are installed near the battery pack. Set concentration thresholds for hydrogen and carbon monoxide, and a warning threshold for carbon dioxide; if the concentration of any gas exceeds the limit, immediately cut off the main circuit and trigger emergency cooling. Step (4.2): Setting early warning thresholds based on loss function and predicted data; The loss function includes reconstruction residual, prediction residual, and physical residual. The prediction data output includes voltage, temperature, lithium plating risk index, and SEI film health. The original data includes negative electrode potential, SEI film growth index, and lithium plating sensitivity factor. Parameters related to lithium plating and SEI film failure are selected from these to set the warning threshold. Step (4.2.1), Lithium plating threshold setting; Prioritize reconstruction error and lithium plating sensitivity factors as early warning indicators, and then combine them with lithium plating residual and negative electrode overpotential for lithium plating verification; Predicting future lithium plating risk using a lithium plating risk index, with the following parameters selected: reconfiguration error L. r Lithium plating sensitive factor a, negative electrode overpotential Lithium plating residual L p1 Lithium plating risk index Z l ; Step (4.2.2), SEI membrane failure threshold setting; Reconstruction error and SEI film growth indicators are prioritized as early warning indicators, and then combined with SEI film residual and SEI film health to verify SEI film failure. The selected parameters are: reconstruction error L r SEI film growth parameters, SEI film impedance change rate, SEI film residual L p2 SEI membrane health level Z s .

6. The early warning and closed-loop control method based on deep learning according to claim 1, characterized in that: Step (5) includes, Step (5.1), Battery parameter control; Lithium plating can be mitigated by controlling battery parameters during charging, and the damage to the SEI film caused by battery load can be reduced by controlling battery parameters during discharging. Step (5.1.1): Control of battery parameters during charging; During charging, the charging speed and battery health are coordinated and adjusted to control the cutoff voltage at different temperatures. The formula for calculating the cutoff voltage is as follows: in, The reference cutoff voltage; This is the temperature compensation coefficient; This is the lithium plating risk compensation coefficient; This is the lithium plating sensitivity compensation coefficient; Battery temperature; The charging current is controlled based on the battery capacity, gradually decreasing as the battery capacity increases; the cutoff voltage and charging current are dynamically adjusted according to the lithium plating index. The charging current calculation formula is as follows: in, This represents the initial charging current; SOC represents the battery capacity. The descending flow index; This is the current attenuation coefficient; Characterized the current decay factor, when As the current increases, the charging current decreases further; Step (5.1.2): Control of battery parameters during discharge; The formula for calculating the discharge current is: in, This is the rated discharge current; This is a weighting coefficient for health status; This is the weighting coefficient for impedance change; It is a constant to prevent division by zero errors; Step (5.2), Pulse lithium plating repair control during charging; When lithium plating is detected in the battery, pulse repair is performed when the battery capacity is between 60% and 100%; when the negative electrode is overpotential... >0.1V, lithium plating residual L p1 >0.05, Lithium plating risk index Z l Repair should be performed when the value is >0.8; Step (5.2.1), Lithium plating repair pulse settings; The pulse period is set to 30 seconds, the pulse current to -0.5C, the discharge time to 200ms, and a 1-second rest period after each discharge. The chemical reaction formula is: ; Step (5.2.2): Pulse amplitude adaptive adjustment; The pulse period, pulse current magnitude, discharge time, and rest period are dynamically adjusted based on changes in gas concentration.

7. A deep learning-based early warning and closed-loop control system applying the deep learning-based early warning and closed-loop control method as described in any one of claims 1 to 6, characterized in that: Includes the following execution modules, The data acquisition and preprocessing module is used to acquire the individual cell voltage, total voltage, total current, temperature, SOH, SOC, and SEI film impedance of the battery, and estimate the battery internal resistance, calculate the negative electrode overpotential, lithium plating sensitivity factor, and SEI film growth index. At the same time, it acquires the concentrations of hydrogen, carbon monoxide, and carbon dioxide through a gas sensor. The collected multi-source heterogeneous data is processed by time alignment, standardization, and sliding window partitioning. A physically constrained dual-decoder autoencoder model, comprising: The encoder, consisting of a convolutional layer, an LSTM layer, and a physical constraint layer connected in sequence, is used to receive preprocessed data, extract features, and compress them to generate a latent space vector containing a lithium plating risk index and SEI film health. A reconstruction decoder is used to reconstruct the latent space vector back into battery state data with the same dimension as the encoder input. A predictive decoder is used to predict battery state parameters over a future period of time based on the latent space vector. The loss function calculation module is used to calculate the total loss of the model, which includes the reconstruction error of the reconstruction decoder, the prediction error of the prediction decoder, and the physical residual loss based on the electrochemical mechanism. The multi-level early warning module is used to determine the graded threshold based on gas concentration data, the loss function output value of the model, the predicted output value, and the calculated derived parameters, and to trigger early warning signals of different levels. The closed-loop control module is used to respond to the warning signal, adaptively adjust the battery's charging and discharging parameters, and / or trigger the lithium plating pulse repair program.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the deep learning-based early warning and closed-loop control method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: It stores a computer program that, when executed, implements the deep learning-based early warning and closed-loop control method as described in any one of claims 1 to 6.

Images