Battery full life cycle monitoring and control method

Abstract

The invention discloses a battery full life cycle monitoring and control method which is characterized by comprising the following steps: S1, a BMS (Battery Management System) obtains real-time state information data of a battery through an acquisition chip or a sensor; s2, de-noising processing is carried out on the obtained real-time state information data of the battery; s3, obtaining battery health state index data and generating fusion features; s4, selecting a lightweight DNN and CNN-LSTM hybrid model with a complementary structure as a generator and a verifier; s5, expanding a training sample set; s6, training the deep neural network to obtain a battery full life cycle monitoring model; s7, inputting the fused features into a battery full life cycle monitoring model, and outputting an SOH predicted value; s8, comparing the SOH predicted value with the SOH predicted values of the battery for nearly three times; the battery full life cycle monitoring and control method is suitable for actual scenes such as new energy automobiles and energy storage systems in which a large amount of labeled battery data is difficult to obtain and in complex temperature working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery control, and in particular to a battery full life cycle monitoring and control method. BACKGROUND

[0002] With the popularization of new energy vehicles and energy storage systems, battery state of health (SOH) estimation has become the core of safety and life guarantee. Traditional methods rely on supervised learning with a large amount of labeled data or electrochemical models based on empirical formulas, but the former faces the problems of high data collection cost and difficult labeling, and the latter is difficult to adapt to the nonlinear degradation characteristics under complex working conditions. Existing technologies also have limitations such as weak generalization ability and insufficient real-time performance, and there is an urgent need for efficient and low-data-dependent battery state of health (SOH) estimation solutions. SUMMARY

[0003] In view of the above defects or deficiencies in the prior art, a battery full life cycle monitoring and control method is provided.

[0004] The first aspect of the present application provides a battery full life cycle monitoring and control method, characterized by comprising the following steps:

[0005] S1. The BMS obtains real-time state information data of the battery through a collection chip or a sensor, wherein the real-time state information of the battery includes real-time voltage, current and temperature of the battery;

[0006] S2. The obtained real-time state information data of the battery is denoised;

[0007] S3. The battery health status index data is extracted from the denoised real-time state information data of the battery, and the battery health status index data is standardized, and the standardized battery health status index data is fused through an attention mechanism to generate a fusion feature;

[0008] S4. A lightweight DNN and a CNN-LSTM hybrid model with structural complementarity are selected as a generator and a verifier, and initially, only the lightweight DNN and the CNN-LSTM hybrid model are trained with labeled data;

[0009] S5. The trained lightweight DNN and CNN-LSTM hybrid model are used to screen unlabeled data to obtain pseudo-labeled data;

[0010] S6. The training sample set is used to train a deep neural network to obtain a battery full life cycle monitoring model;

[0011] S7. The fusion feature is input into the battery full life cycle monitoring model to output an SOH prediction value;

[0012] S8. Comparing the SOH prediction value with the battery's last three SOH prediction values, if the deviation of the SOH prediction value from any of the battery's last three SOH prediction values is greater than 5%, re-collect data and predict; if the deviation of the SOH prediction value re-collected and predicted is still greater than 5%, mark as "battery abnormal" and trigger an alarm.

[0013] Further, the battery health indicator data includes: the voltage platform slope in the battery charging stage, the peak intensity of the incremental capacity curve (ICA), and the internal resistance growth rate in the cycle process.

[0014] Further, the specific process of generating the fusion feature by the attention mechanism for the standardized battery health status indicator data is: splicing the standardized battery charging stage voltage platform slope feature F1, the incremental capacity curve (ICA) peak intensity feature F2, and the internal resistance growth rate feature F3 in the cycle process into a total feature F all , by introducing an attention matrix M, calculating the contribution weights W1, W2, and W3 of the features F1, F2, and F3 to SOH, wherein W1 = softmax(M*F1*F ^T all ), W2 = softmax(M*F2*F ^T all ), and W3 = softmax(M*F3*F ^T all ), the fusion feature F fuse = W1*F1 + W2*F2 + W3*F3.

[0015] Further, the screening of unlabeled data using the trained lightweight DNN and the CNN-LSTM hybrid model to expand the training sample set specifically includes: using the trained lightweight DNN to predict the fusion feature F fuse of the unlabeled data, obtaining the initial pseudo label SOH pseudoA of the unlabeled data; inputting the unlabeled data into the CNN-LSTM hybrid model to obtain the predicted label SOH pseudoB of the unlabeled data, calculating the deviation ΔSOH = |SOH pseudoA -SOH pseudoB |; if ΔSOH ≤ 0.02: considering that the initial pseudo label SOH pseudoA is reliable, retaining the initial pseudo label SOH pseudoA , and taking SOH pseudo = (SOH pseudoA + SOH pseudoB ) / 2 as the final pseudo label of the pseudo label data; if ΔSOH > 0.02, then discard the initial pseudo label SOH pseudoA, avoid noise introduction.

[0016] Further, the deep neural network comprises an input layer, a hidden layer, an output layer and a Dropout layer.

[0017] Further, the specific process of training the deep neural network to obtain the battery full life cycle monitoring model comprises: training the model only using the labeled data to obtain an initial model of the battery full life cycle monitoring model; mixing the labeled data and the pseudo label at a ratio of 1:1 to obtain a model training set; designing a loss function of the battery full life cycle monitoring model as:

[0018] Loss=α*Loss labeled +(1-α)*Loss pseudo ; wherein, a is a weight coefficient, Loss labeled is the MSE loss of the labeled data, and Loss pseudo is the MSE loss of the pseudo label data; training the initial model using the model training set, stopping when the model training set MSE is less than 0.3% and the validation set MSE is less than 0.8%, saving the model parameters to obtain the battery full life cycle monitoring model.

[0019] Further, the battery generates new labeled data in subsequent cycles, which is fed back to the battery full life cycle monitoring model, and the output layer weight of the battery full life cycle monitoring model is updated.

[0020] Further, for the long-term used battery, the battery health state index data is re-extracted every 100 cycles of the battery, and the input of the battery full life cycle monitoring model is updated.

[0021] The second aspect of the application provides a battery full life cycle monitoring and control device, comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the battery full life cycle monitoring and control as described above.

[0022] The third aspect of the application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to realize the battery full life cycle monitoring and control as described above.

[0023] Therefore, the application has at least the following beneficial effects:

[0024] The application fuses feature extraction, pseudo-label generation and optimization, and stage-by-stage model training, and in the application scenario where labeled data is scarce, the battery SOH estimation accuracy is significantly better than that of traditional supervised learning, and the real-time performance and generalization ability are also considered, and the application is especially suitable for new energy vehicles, energy storage systems and other practical scenarios that are difficult to obtain a large amount of labeled battery data and complex temperature conditions. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A flowchart of a battery full life cycle monitoring and control method. DETAILED DESCRIPTION

[0026] The embodiments of the application will be described in detail below. The following embodiments are implemented on the premise of the technical solutions of the application, and detailed implementation modes and specific operation processes are given, but the protection scope of the application is not limited to the following embodiments.

[0027] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0028] Embodiment 1

[0029] Figure 1 A flowchart of a battery full life cycle monitoring and control method. The method comprises the following steps:

[0030] S1. The BMS obtains real-time state information data of the battery through a collection chip or a sensor, and the real-time state information of the battery at least includes real-time voltage, current and temperature of the battery;

[0031] The BMS generally has high-precision sensing and real-time data transmission capabilities, and its core hardware includes a voltage collection unit, a current collection unit, a temperature collection unit and an edge computing unit. The BMS outputs structured data at a fixed period, and the structured data generally includes battery pack number, current cycle number, collection timestamp and other basic information, as well as real-time state information data of each battery cell voltage, total current, average temperature and cumulative charge capacity;

[0032] S2. The obtained real-time state information data of the battery is denoised; the denoising process can remove outliers and redundant information in the real-time state information data of the battery, and improve the accuracy of the data;

[0033] S3. Extract battery health status index data from the denoised real-time state information data of the battery, and standardize the battery health status index data. The standardized battery health status index data is generated into fusion features through an attention mechanism;

[0034] The voltage platform slope of the battery charging stage, the peak intensity of the incremental capacity curve (ICA), and the internal resistance growth rate during the cycle process are taken as the battery health index data.

[0035] The charging voltage platform slope is the ratio of the voltage change amount to the capacity change amount during the constant current stage of the battery. The charging voltage platform slope reflects the active material lithium extraction ability of the battery. The aging of the battery is because the positive electrode material structure degrades, the voltage platform segment voltage rises faster, and the absolute value of the slope increases.

[0036] The incremental capacity curve (ICA) reflects the internal electrochemical reaction of the battery through the derivative of capacity with respect to voltage (dQ / dV). The ICA peak of the aging battery will decrease and move to the right due to the thickening of the SEI film and the loss of active material. The peak intensity is directly related to the amount of active material. Since the BMS collects discrete voltage-capacity data, numerical differentiation is required to calculate dQ / dV. Take the voltage-capacity data with a window size of 11, use a 3rd order polynomial to fit Q=aV 3 +bV 2 +cV+d; the derivative is dQ / dV=3aV 2 +2bV+c; substitute each voltage point V i , calculate the corresponding dQ / dV i , and get the smooth ICA curve. Generally, the ICA peak is located by threshold method + extreme value judgment: set the peak threshold: take 50% of the maximum value of the ICA curve as the threshold to filter small fluctuations; extreme value judgment: if a point dQ / dV i is greater than the adjacent two points (dQ / dV i-1 , dQ / dV i+1 ) and greater than the threshold, it is determined as the peak, and the peak voltage V peak and the peak intensity (dQ / dV) peak are recorded. Taking a ternary lithium battery as an example, the main peak of the new ternary lithium battery is usually at about 3.8V, and the peak intensity is 5-6Ah / V. After 500 cycles, the peak intensity decreases to 3-4Ah / V, and the peak voltage moves to 3.85V. The BMS quantifies the loss of active material by the change in peak intensity. For example, if the peak intensity decreases by 20%, the SOH decreases by about 10%.

[0037] The internal resistance of the battery (including ohmic resistance and charge transfer resistance) increases with the increase of the cycle number. The ohmic resistance reflects the degradation of the electrode / electrolyte, and the charge transfer resistance reflects the interface reaction efficiency. The internal resistance growth rate is a direct indicator of SOH. The internal resistance of the battery at the first cycle (or factory calibration) is taken as the reference resistance R0, the BMS calculates the current internal resistance R t every 10 cycles, records the corresponding relationship between R t and the cycle number N, and the internal resistance growth rate γ=[(R t-R0) / R0]×100% / N;

[0038] Standardization can be achieved using Z-score normalization to avoid model bias caused by differences in the magnitude of features of different battery health indicators.

[0039] The voltage plateau slope characteristic F1 during the standardized battery charging phase, the peak intensity characteristic F2 of the incremental capacity curve (ICA), and the internal resistance growth rate characteristic F3 during cycling are concatenated to form the total characteristic F. all By introducing an attention matrix M, the contribution weights W1, W2, and W3 of features F1, F2, and F3 to SOH are calculated. 3， Where W1 = softmax(M*F1*F ^T all W2 = softmax(M*F2*F) ^T all W3 = softmax(M*F3*F) ^T all The fusion feature F fuse =W1*F1+W2*F2+W3*F3; By dynamically allocating the weights of each battery health indicator data feature, information strongly correlated with SOH is highlighted, key features of different aging stages are effectively captured, and the anti-interference capability is greatly improved compared with traditional splicing.

[0040] S4. Select two complementary lightweight DNN and CNN-LSTM hybrid models as generator and validator. Initially, only labeled data is used to train the lightweight DNN and the CNN-LSTM hybrid model.

[0041] S5. Use the trained lightweight DNN and the CNN-LSTM hybrid model to filter unlabeled data and obtain pseudo-labeled data;

[0042] The fusion feature F of unlabeled data is applied using a trained lightweight DNN. fuse Prediction is performed to obtain the initial pseudo-label SOH of the unlabeled data. pseudoA The unlabeled data is input into a CNN-LSTM hybrid model to obtain the predicted label SOH for the unlabeled data. pseudoB Calculate the deviation between the two, ΔSOH=|SOH pseudoA -SOH pseudoB |; If ΔSOH≤0.02: the initial pseudo-label SOH is considered to be pseudoA Reliable, retaining the initial pseudo-label SOH pseudoA , SOH pseudo =(SOH) pseudoA +SOH pseudoB) / 2 is used as the final pseudo-label of the pseudo-label data; if ΔSOH > 0.02, the initial pseudo-label is discarded.

[0043] SOH pseudoA To avoid introducing noise, the lightweight DNN focuses on quickly generating pseudo-labels, while the CNN-LSTM hybrid model, through a multi-structure fusion of one CNN layer, one LSTM layer, and a fully connected layer, focuses on verifying the reliability of the pseudo-labels. This invention, by employing a dual-model alternating verification strategy, ensures the reliability of the pseudo-labels.

[0044] S6. Train a deep neural network to obtain a battery lifecycle monitoring model; the deep neural network includes an input layer, hidden layers, an output layer, and a dropout layer; first, train the deep neural network model with a small amount of labeled data to enable the model to initially grasp the mapping relationship between fused features and battery SOH, providing reliable initial parameters for subsequent pseudo-label learning; after pre-training, the model's SOH estimation RMSE on the validation set should be <1.5% to ensure the reliability of the initial mapping relationship; mix the labeled data and the pseudo-label data in a 1:1 ratio to obtain the model training set, and expand the training set by adding labeled data and dynamically selected pseudo-label data to optimize model parameters and improve the adaptability to unlabeled batteries.

[0045] The loss function for the battery lifecycle monitoring model is: Loss = α * Loss labeled +(1-α)*Loss pseudo Where α is the weighting coefficient, and Loss labeled For labeled data, the MSE loss is Loss pseudo The MSE loss is calculated for pseudo-labeled data. The initial weight coefficient is 0.7, which is reduced by 0.1 every 5 epochs until it reaches 0.3, ensuring that the model relies on real labels in the early stages and gradually trusts pseudo labels in the later stages. Every 2 epochs of training, the pseudo-label generation and selection process is re-executed using the current model to update the pseudo-label sample set. The model is stopped when the MSE of the training set is <0.3% and the MSE of the validation set is <0.8%, and the model parameters are saved.

[0046] S7. Input the fused features into the battery life cycle monitoring model and output the SOH prediction value;

[0047] S8. Compare the predicted SOH value with the battery's last three predicted SOH values. If the current predicted value deviates by more than 5%, re-collect data and re-predict. If the deviation is still greater than 5%, mark it as "battery abnormal" and trigger an alarm. For example, if the battery's last three predicted SOH values ​​are 92.1%, 91.8%, and 91.5%, and the current output predicted SOH value is 89.0%, then re-collect data and re-predict.

[0048] Furthermore, in order to quickly correct prediction biases and identify changes in short-term battery aging characteristics, new labeled data generated by the battery in subsequent cycles can be fed back to the battery life cycle monitoring model, and the output layer weights of the battery life cycle monitoring model can be updated. The new labeled data generated by the battery in each cycle does not require retraining the entire model; only the output layer weights of the model are updated. By utilizing the direct mapping relationship between the output layer and SOH, the weights can be fine-tuned through gradient descent, and biases can be quickly corrected.

[0049] Because long-term cycling can lead to changes in the battery aging mechanism, new features can reflect the switching of core aging factors and avoid the decay of the correlation between old features and SOH. For batteries used for a long time, the battery health status index data is re-extracted every 100 cycles and the input of the battery life cycle monitoring model is updated.

[0050] Example 2

[0051] This embodiment provides a battery life cycle monitoring and control device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the steps of the battery life cycle monitoring and control method as described in Embodiment 1.

[0052] Example 3

[0053] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the steps of the battery life cycle monitoring and control method as described in Embodiment 1.

[0054] While specific embodiments of this application have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this application is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this application, but all such changes and modifications fall within the scope of protection of this application.

Claims

1. A battery full life cycle monitoring and control method, characterized by, The method comprises the following steps: S1. The BMS obtains real-time state information data of the battery through a collection chip or a sensor, wherein the real-time state information of the battery includes real-time voltage, current and temperature of the battery; S2. The real-time state information data of the battery is subjected to denoising processing; S3. Battery health state index data is extracted from the denoised battery real-time state information data, and the battery health state index data is subjected to standardization processing, and the battery health state index data subjected to the standardization processing is fused into features through an attention mechanism; S4. A lightweight DNN and a CNN-LSTM hybrid model with structural complementarity are selected as a generator and a verifier, and initially, only the labeled data is used to train the lightweight DNN and the CNN-LSTM hybrid model; S5. The trained lightweight DNN and the CNN-LSTM hybrid model are used to screen the unlabeled data to obtain pseudo-labeled data; S6. A deep neural network is trained to obtain a battery full-life cycle monitoring model; S7. The fused features are input into the battery full-life cycle monitoring model to output an SOH prediction value; S8. The SOH prediction value is compared with the SOH prediction values of the battery in the last three times, and if the deviation of the SOH prediction value from any one of the SOH prediction values of the battery in the last three times is greater than 5%, the data is re-collected and predicted; If the deviation of the SOH prediction value obtained by re-collecting and predicting the data is still greater than 5%, the battery is marked as "abnormal" and an alarm is triggered.

2. The battery full life cycle monitoring and control method of claim 1, wherein: The battery health index data includes a voltage platform slope in the battery charging stage, a peak intensity of an incremental capacity curve, and an internal resistance growth rate in the cycle process.

3. The battery full life cycle monitoring and control method of claim 1, wherein: The specific process of generating a fusion feature by the attention mechanism from the standardized battery health state index data includes: splicing the voltage platform slope feature F1 of the battery charging stage, the peak intensity feature F2 of the incremental capacity curve, and the internal resistance growth rate feature F3 in the cycle process into a total feature F all By introducing an attention matrix M, the contribution weights W1, W2, and W3 of the features F1, F2, and F3 to the SOH are calculated; wherein W1 = softmax(M*F1*F^ T all ), W2 = softmax(M*F2*F^ T all ), and W3 = softmax(M*F3*F^ T all ), and the fusion feature F fuse = W1*F1+W2*F2+W3*F3.

4. The battery full life cycle monitoring and control method of claim 1, wherein: The trained lightweight DNN and the CNN-LSTM hybrid model are used for screening of unlabeled data to obtain pseudo-label data, and the specific process includes: using the trained lightweight DNN to screen the fusion features F of the unlabeled data fuse Performing prediction to obtain an initial pseudo-label SOH of the unlabeled data pseudoA ; inputting the unlabeled data into the CNN-LSTM hybrid model to obtain a predicted label SOH of the unlabeled data pseudoB , calculating the deviation ΔSOH = |SOH pseudoA -SOH pseudoB |; if ΔSOH≤0.02, considering that the initial pseudo-label SOH pseudoA is reliable, retaining the initial pseudo-label SOH pseudoA , and taking SOH pseudo =(SOH pseud oA+SOH pseudoB ) / 2 as the final pseudo-label of the pseudo-label data; if ΔSOH>0.02, discarding the initial pseudo-label SOH pseudoA .

5. The battery full life cycle monitoring and control method of claim 4, wherein: The deep neural network comprises an input layer, a hidden layer, an output layer and a Dropout layer.

6. The battery full life cycle monitoring and control method of claim 5, wherein: The specific process of training the deep neural network to obtain the battery full-life cycle monitoring model comprises: first, only the labeled data is used to train the deep neural network to obtain an initial model of the battery full-life cycle monitoring model; the labeled data and the pseudo-labeled data are mixed in a 1:1 ratio to obtain a model training set; a loss function of the battery full-life cycle monitoring model is designed as: Loss = a * Loss labeled + (1 - a) * Loss pseudo ; wherein a is a weight coefficient, Loss labeled is the MSE loss of the labeled data, and Loss pseudo is the MSE loss of the pseudo-labeled data; the initial model is trained using the model training set, and the training is stopped when the model training set MSE < 0.3% and the validation set MSE < 0.8%, the model parameters are saved, and the battery full-life cycle monitoring model is obtained.

7. The battery full life cycle monitoring and control method of claim 6, wherein: The battery generates new labeled data in the subsequent cycles, which is fed back to the battery full-life cycle monitoring model, and the output layer weight of the battery full-life cycle monitoring model is updated.

8. The battery full life cycle monitoring and control method of claim 1, wherein: For the long-term used battery, the battery health state index data is re-extracted every 100 cycles, and the input of the battery full-life cycle monitoring model is updated.

9. A battery full life cycle monitoring and control device comprising: A memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the battery full-life cycle monitoring and control method according to any one of claims 1-8.

10. A computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the battery full-life cycle monitoring and control method according to any one of claims 1-8.

Images