Lithium ion battery health state prediction method and device based on multi-source data multi-task learning model

Abstract

The invention discloses a lithium ion battery health state prediction method and device based on a multi-source data multi-task learning model, and the method comprises the steps: screening out an experiment data set based on an NASA lithium ion battery data set; the method comprises the following steps: respectively extracting discharge capacity increment curve (IC curve) image data related to the state of health (SOH) of the lithium ion battery and time sequence data containing parameters such as temperature-voltage and the like, and respectively preprocessing to generate an SOH training set and a test set; a 2D-CNN module is adopted to train IC curve image data, a BiLSTM module is adopted to train time series data, outputs of the two modules are connected through feature fusion, and a multi-task SOH prediction neural network (MTSOHFNN) model is established; and training and testing by adopting the SOH training set and the test set, and outputting an evaluation result of the SOH of the lithium ion battery. According to the invention, the lithium ion battery SOH prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to a method and device for predicting the state of health of a lithium-ion battery based on a multi-source data multi-task learning model. Background Art

[0002] In the full life cycle management of lithium-ion batteries, accurate prediction of the state of health (SOH) of the battery is a core link to ensure the safety and reliability of the battery system. SOH reflects the remaining capacity and aging degree of the battery, and directly affects the performance degradation evaluation, life prediction and maintenance strategy formulation of the battery pack. Especially in key fields such as electric vehicles and energy storage systems, real-time monitoring of SOH can avoid safety hazards such as overcharging / overdischarging and thermal runaway of the battery, while optimizing the energy management strategy and extending the service life of the battery. However, existing SOH prediction technologies still face challenges such as insufficient data utilization and insufficient model generalization ability. Traditional methods mostly rely on a single type of data (such as voltage, current or temperature time series data), ignoring the complementarity of multi-source data (such as capacity increment curve, electrochemical characteristic image), resulting in incomplete information extraction and limited prediction accuracy.

[0003] In the prior art, although the incremental capacity analysis (ICA) can intuitively reflect the battery aging characteristics by analyzing the peak change of the IC curve, it has extremely high requirements for data sampling accuracy and relies on charge and discharge data under static conditions, making it difficult to apply in real time; the equivalent circuit model method extracts polarization parameters based on a first-order RC model, but the model complexity is high and it is sensitive to dynamic working conditions; data-driven methods (such as support vector regression) can handle non-linear relationships, but rely on artificial feature engineering and are difficult to effectively fuse high-dimensional multi-source data. In addition, although single-modal deep learning models (such as convolutional neural networks for processing IC curve images or bidirectional long short-term memory networks for processing time series data) can automatically extract features, they do not fully utilize the complementarity of multi-source data, are sensitive to noise and have insufficient generalization ability. In practical applications, data acquisition is easily interfered by factors such as environmental noise and uneven sampling intervals, and traditional filtering methods (such as moving average) are difficult to retain key features, further exacerbating the instability of the model.

[0004] To address the above problems, there is an urgent need for an SOH prediction method that can integrate multi-source data, adaptively denoise and achieve efficient feature fusion. The prior art lacks the ability to co-model multi-source data. Most methods are designed for a single modality only and cannot co-mine the spatial distribution characteristics of the IC curve and the dynamic evolution law of time series data, which limits the further improvement of prediction performance. At the same time, traditional methods are difficult to balance the model complexity and real-time requirements when dealing with complex working conditions, resulting in insufficient robustness and adaptability in practical applications. Therefore, developing an SOH prediction technology that takes into account the complementarity of multi-source data, noise suppression ability and dynamic feature extraction has become a key research direction to promote the intelligent upgrade of the battery management system. Summary of the Invention

[0005] The present invention aims to overcome the above-mentioned drawbacks of the prior art, and proposes a method and device for predicting the state of health of lithium-ion batteries based on a multi-source data multi-task learning model. The present invention proposes a complete data preprocessing method, combined with a high-precision multi-task learning model, to improve the estimation accuracy of the state of health (SOH) of lithium-ion batteries.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] The first aspect of the present invention relates to a method for predicting the state of health of lithium-ion batteries based on a multi-source data multi-task learning model, including the following steps:

[0008] S1: Based on the NASA lithium-ion battery dataset, screen the relevant datasets, and respectively extract the image data of the incremental discharge capacity curve (IC curve) related to the state of health (SOH) of lithium-ion batteries and the time series data including parameters such as temperature-voltage;

[0009] S2: Preprocess the data extracted related to the state of health of lithium-ion batteries to generate an SOH training set and a test set;

[0010] S3: Use the 2D-CNN module to train the IC curve image data, use the BiLSTM module to train the time series data, connect the outputs of the two modules through feature fusion, and establish a multi-task SOH prediction neural network MTSOHFNN model;

[0011] S4: Use the SOH training set and test set for training and testing, and output the evaluation result of the state of health (SOH) of lithium-ion batteries.

[0012] Further, in the step S1, the specific content of the constructed NASA dataset is as follows:

[0013] The dataset for predicting the SOH of lithium-ion batteries is divided into an image data part and a time series data part.

[0014] S1-1: The incremental capacity curve transforms the voltage plateau that is difficult to identify in the battery voltage and capacity change curve into an easily identifiable incremental capacity peak. The incremental capacity analysis method obtains by calculating the capacity change corresponding to the unit voltage:

[0015]

[0016] In the formula, Q k , Q k-1 respectively represent the capacity change amounts at times k and k-1, V B,k , V B,k-1respectively represent the battery terminal voltage values at time k and k-1. However, during the data acquisition process, the real-time capacity change is difficult to collect and obtain. Therefore, the current and time difference are used for replacement:

[0017]

[0018] where T represents the discharge time, and I B represents the current during the charge and discharge process of the battery.

[0019] S1-2: For the time-series data part, first select the first-order RC equivalent circuit model, identify it using the least squares method based on the NASA dataset, collect the identification results of each cycle, the ohmic resistance R0 of the lithium-ion battery, the polarization resistance R p of the lithium-ion battery, and the polarization capacitance C p of the lithium-ion battery.

[0020] In addition, based on the NASA dataset, perform screening, and respectively select the discharge voltage curve slope (in the interval from the maximum voltage to the cut-off voltage) Rate V , the temperature change rate calculation Rate T and the average discharge time T mean These three types of data including the three dimensions of time-voltage-temperature are used as part of the time-series input;

[0021] Furthermore, in step S2, the data related to the state of health of the lithium-ion battery is preprocessed as follows:

[0022] S2-1: For the extracted IC curve image, since the data sampling time in the NASA dataset is not fixed and the interval is large, in order to avoid ignoring important information on the IC curve, linear interpolation is performed on the collected voltage data, and the minimum sampling interval is set to prevent the noise step caused by dV B being too small.

[0023] In order to avoid the influence of noise on the SOH prediction accuracy during the training process, the Savitzky-Golay filtering method is introduced for noise reduction processing of the IC curve. Savitzky-Golay fits a low-order polynomial to the sliding window of adjacent data points through the linear least squares method, sets the local fitting order to third-order fitting, and the sliding window size to 60 sampling points.

[0024] S2-2: In the time-series part of the data, it is necessary to verify the correlation between the data and the state of health SOH of the lithium-ion battery to avoid the additional computing power burden on the prediction model caused by redundant data input. Therefore, the Pearson correlation coefficient method PCC is used to measure the linear relationship between the discrete data X and the target data Y (SOH), and its calculation process is as follows:

[0025]

[0026] Among them, and respectively represent the average values of the discrete data X and the target data Y, and n represents the sequence length. For the result of the PCC coefficient, the closer its absolute value is to 1, the higher the correlation with the target data. The minimum-maximum normalization is performed on the filtered data, and its formula is as follows:

[0027]

[0028] In the formula, X nrom is the normalization result, and X min and X max are the maximum and minimum values in the data respectively.

[0029] Furthermore, in step S3, the process of establishing the multi-task SOH prediction neural network (MTSOHFNN) model is as follows:

[0030] S3-1: The multi-task SOH prediction neural network trains the health state prediction model by designing a multi-task model. The 2D-CNN branch and the BiLSTM branch are respectively used to extract the lithium-ion battery aging characteristics in different types of data. Finally, the outputs of the two networks are connected together through a feature fusion layer and two fully connected layers to construct the MTSOHFNN model.

[0031] S3-2: The specific process of training the health state SOH prediction using the multi-task SOH prediction neural network (MTSOHFNN) model is as follows:

[0032] Construct a dataset corresponding to the IC curve image - SOH, and train it with the first 65% of the dataset as the training set and the last 35% as the test set;

[0033] Construct a dataset corresponding to the time series data - SOH, and construct the dataset in the same division method;

[0034] Combine the two datasets to construct a datastore with image - time series data;

[0035] Set the number of iterations and the learning rate of the training process;

[0036] Input the established dataset into the established MTSOHFNN model for training to obtain the trained MTSOHFNN model.

[0037] During the training process, both the 2D-CNN and the BiLSTM in the MTSOHFNN model are trained using the same mean squared error loss function. The expression of the mean squared error loss function is:

[0038]

[0039] Among them, MSE is the mean square error, n is the number of samples, and y i is the actual prediction result of the i-th data sample, and

[0040] In step S4, the evaluation method of the state of health SOH of the lithium-ion battery is output as follows:

[0041] To evaluate the prediction ability of the MTSOHFNN model for SOH, the coefficient of determination R 2 and the mean absolute percentage error MAE and the root mean square error RMSE are used to evaluate the SOH result, where y i and represent the true value and the estimated value respectively, represents the average value:

[0042]

[0043] The second aspect of the present invention relates to a lithium-ion battery state of health prediction device based on a multi-source data multi-task learning model, including a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the lithium-ion battery state of health prediction method based on the multi-source data multi-task learning model of the present invention.

[0044] The third aspect of the present invention relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the lithium-ion battery state of health prediction method based on the multi-source data multi-task learning model of the present invention.

[0045] The present invention includes screening out an experimental data set based on the NASA lithium-ion battery data set, respectively extracting image data of the incremental discharge capacity curve (IC curve) related to the state of health SOH of the lithium-ion battery and time series data including parameters such as temperature-voltage, and respectively performing preprocessing to generate an SOH training set and a test set; using a 2D-CNN module to train the IC curve image data, using a BiLSTM module to train the time series data, connecting the outputs of the two modules through feature fusion, and establishing a multi-task SOH prediction neural network (MTSOHFNN) model; using the SOH training set and the test set for training and testing, and outputting the evaluation result of the state of health SOH of the lithium-ion battery. The present invention improves the accuracy of the SOH prediction of the lithium-ion battery.

[0046] The beneficial effects of the present invention are:

[0047] 1. A method for predicting the state of health of a lithium-ion battery based on a multi-source data multi-task learning model is proposed, and the interaction between different neural networks is designed. While combining multi-source information, the prediction accuracy of the state of health (SOH) of the lithium-ion battery is improved.

[0048] 2. Based on the proposed prediction model, the IC curve is used as the image input, reducing the extraction and preprocessing process of various features in the IC curve. The IC curve is simply denoised by the Savitzky-Golay filtering method, avoiding the interference of redundant data on the prediction ability of the image prediction branch and reducing the computing power burden.

[0049] 3. In the time series data prediction part, the correlation between the parameters of the first-order RC equivalent circuit model and the state of health during the aging process of the lithium-ion battery is considered, and the correlation degree of the identification data of the first-order RC model is screened. And three-dimensional data including time-voltage-temperature are extracted from the data set and combined to ensure sufficient feature extraction in the time series data part. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a schematic flow chart of the method of the present invention.

[0051] Figure 2 is the IC curve diagram under different cycles in the present invention.

[0052] Figure 3 is the IC curve filtering comparison image in the present invention.

[0053] Figure 4 is a schematic diagram of 4 capacity decay curves in the present invention.

[0054] Figure 5 is the framework diagram of the multi-task SOH prediction neural network model in the present invention.

[0055] Figure 6(a) - Figure 6(d) is the prediction result of four batteries in the present invention. Among them, Fig. 6(a) is the SOH prediction result of battery B0005, Fig. 6(b) is the SOH prediction result of battery B0006, Fig. 6(c) is the SOH prediction result of battery B0007, and Fig. 6(d) is the SOH prediction result of battery B0018.

[0056] Figure 7 is the schematic diagram of the device of the present invention. SPECIFIC IMPLEMENTATION METHOD

[0058] The present invention will be further described below with reference to the accompanying drawings.

[0059] Example 1

[0060] Refer to Figure 1As shown in the figure, a method for predicting the state of health of a lithium-ion battery based on a multi-source data multi-task learning model includes the following steps:

[0061] S1: Based on the NASA lithium-ion battery dataset, screen the relevant datasets, and respectively extract the image data of the incremental discharge capacity curve (IC curve) related to the state of health SOH of the lithium-ion battery and the time-series data containing parameters such as temperature-voltage.

[0062] S2: Preprocess the data extracted related to the state of health of the lithium-ion battery to generate an SOH training set and a test set.

[0063] S3: Use the 2D-CNN module to train the IC curve image data, use the BiLSTM module to train the time-series data, connect the outputs of the two modules through feature fusion, and establish a multi-task SOH prediction neural network (MTSOHFNN) model.

[0064] S4: Use the SOH training set and test set for training and testing, and output the evaluation result of the state of health SOH of the lithium-ion battery.

[0065] Furthermore, in the step S1, the specific content of the constructed NASA dataset is as follows:

[0066] The dataset for predicting the SOH of lithium-ion batteries is divided into an image data part and a time-series data part.

[0067] S1-1: The incremental capacity curve transforms the voltage platform that is difficult to identify in the battery voltage and capacity change curve into an easily identifiable incremental capacity peak. The incremental capacity analysis method is obtained by calculating the capacity change corresponding to the unit voltage:

[0068]

[0069] In the formula, Q k , Q k-1 respectively represent the capacity change amounts at times k and k-1, and V B,k , V B,k-1 respectively represent the battery terminal voltage values at times k and k-1. However, during the data acquisition process, the real-time capacity change is difficult to collect and obtain, so the current and time difference are used for replacement:

[0070]

[0071] Among them, T represents the discharge time, and I B represents the battery charge and discharge process current. As Figure 2 shown, the peaks and valleys of the IC curve shift continuously under different cycles, and there is a certain correlation with the degradation curve of SOH.

[0072] S1-2: For the time-series data part, first select the first-order RC equivalent circuit model, identify it using the least squares method based on the NASA dataset, collect the identification results of each cycle, the ohmic internal resistance R0 of the lithium-ion battery, the polarization resistance R of the lithium-ion battery p and the polarization capacitance C of the lithium-ion battery p .

[0073] In addition, based on the NASA dataset screening, respectively select the discharge voltage curve slope (in the interval from the maximum voltage to the cut-off voltage) Rate V , the temperature change rate calculation Rate T and the average discharge time T mean Three kinds of data including the three dimensions of time-voltage-temperature are used as part of the time-series input;

[0074] Furthermore, in step S2, the data related to the state of health of the lithium-ion battery is preprocessed as follows:

[0075] S2-1: For the extracted IC curve image, since the NASA data sampling time is not fixed and the interval is large, in order to avoid ignoring important information on the IC curve, linear interpolation is performed on the collected voltage data, and the minimum sampling interval is set to prevent dV B from being too small resulting in noise steps.

[0076] In order to avoid the influence of noise on the SOH prediction accuracy during the training process, the Savitzky-Golay filtering method is introduced to perform noise reduction processing on the IC curve. Savitzky-Golay fits a low-order polynomial to the sliding window of adjacent data points through the linear least squares method, sets the local fitting order to third-order fitting, and the sliding window size to 60 sampling points. The changes in the IC curve before and after filtering are as Figure 3 shown. After Savitzky-Golay filtering, the IC curve retains most of the key features and the overall waveform is not affected, providing a more concise input for the subsequent image-based SOH prediction part.

[0077] S2-2: In the time-series part of the data, it is necessary to verify the correlation between the data and the state of health SOH of the lithium-ion battery to avoid the additional computing power burden on the prediction model caused by redundant data input. Therefore, the Pearson correlation coefficient method PCC is used to measure the linear relationship between the discrete data X and the target data Y (SOH), and its calculation process is as follows:

[0078]

[0079] Among them, and respectively represent the average values of the discrete data X and the target data Y, and n represents the sequence length. For the result of the PCC coefficient, the closer its absolute value is to 1, the higher the correlation with the target data. In the NASA dataset, the PCC results of the time-series data of four batteries are shown in Table 1.

[0080] Table 1 PCC Results of Time-Series Data in NASA Data

[0081]

[0082] As can be seen from Table 1, the lowest correlation coefficient between the features extracted from the NASA dataset and the battery SOH is 0.9229, and the lowest correlation coefficient with the identification result as the feature is 0.9092. Since both are greater than 0.9, it can be considered that they have a relatively high degree of correlation with the SOH change of the battery, so both are retained.

[0083] Perform min-max normalization on the filtered data, and its formula is as follows:

[0084]

[0085] In the formula, X nrom is the normalization result, X min and X max are the maximum and minimum values in the data respectively.

[0086] Furthermore, in step S3, the process of establishing the multi-task SOH prediction neural network MTSOHFNN model is as follows:

[0087] S3-1: The multi-task SOH prediction neural network trains the health state prediction model by designing a multi-task model, uses the 2D-CNN branch and the BiLSTM branch to extract the aging characteristics of lithium-ion batteries in different types of data respectively, and finally connects the outputs of the two networks through a feature fusion layer and two fully connected layers to construct the MTSOHFNN model.

[0088] S3-2: As Figure 4 shown, collect the capacity degradation trajectories of B0005, B0006, B0007, and B0018 in the NASA lithium-ion battery dataset. As the cycle period progresses, the capacities of the four batteries have all decreased to varying degrees. Convert the capacity data into SOH, and according to the definition of SOH, its formula is as follows:

[0089]

[0090] Among them, C i represents the available capacity of the battery at the i-th cycle, and C n represents the rated capacity of the battery.

[0091] S3-3: The specific process of training the health state of health (SOH) prediction using the multi-task SOH prediction neural network MTSOHFNN model is as follows:

[0092] Construct a dataset corresponding to the IC curve image - SOH, and train it with the first 65% of the dataset as the training set and the last 35% as the test set;

[0093] Construct a dataset corresponding to the time series data - SOH, and construct the dataset in the same division method;

[0094] Combine the two datasets to construct a datastore with image - time series data;

[0095] Set the number of iterations and learning rate of the training process;

[0096] Input the established dataset into the established MTSOHFNN model for training to obtain a trained MTSOHFNN model.

[0097] During the training process, both the 2D-CNN and BiLSTM in the MTSOHFNN model are trained using the same error loss function. The expression of the mean square error loss function used in this paper is:

[0098]

[0099] Among them, MSE is the mean square error, n is the number of samples, y i is the actual prediction result of the i-th data sample, is the theoretical predicted value of the i-th data sample.

[0100] S3-4: In the conventional LSTM, the hidden state is always passed from front to back, making it impossible for a simple LSTM network to fully extract the characteristics of time series data. In the actual working state of the battery, the capacity changes at the front and back moments are highly correlated, and it is necessary to take this characteristic into account.

[0101] Therefore, in order to effectively predict the SOH of the battery, it is necessary to extract the implicit features and complex changes in the sequence data. Compared with LSTM, BiLSTM consists of a forward and a backward LSTM unit, which can extract data features from both the forward and backward directions at the same time, and connect the input at the current position with the past information and future information, so as to better extract the characteristic relationship of the aging process. The hidden layer state h(t) of BiLSTM at time t can be obtained through the state of the forward hidden layer and the state of the backward hidden layer h(t). The BiLSTM formula is as follows:

[0102]

[0103] Among them, w i represents the corresponding weight of each layer. Therefore, h(t) in the BiLSTM is passed to the fully connected layer and the output is:

[0104] D v = tanh(w v h(t) + b v ) (8)

[0105] The model used consists of three BiLSTM layers with 128, 64, and 32 neurons respectively, and two fully connected layers with 64 and 32 neurons respectively.

[0106] S3-5 Traditional convolutional neural networks are difficult to fully exploit the feature coupling relationships between different waveform positions in the IC curve. The present invention introduces 2D-CNN to fully extract the feature relationships on different feature channels. The mathematical expression for converting the feature of the (i-1)-th layer of 2D-CNN to the feature map of the i-th layer through three-dimensional convolution operation is as follows:

[0107]

[0108] In the formula, where m represents the set of feature maps of the previous layer connected to the current feature map, P i and Q i are the height and width of the kernel respectively, w7 is the weight matrix of the kernel, and b7 is the corresponding bias. In addition, the size of the convolutional kernel has an important impact on the performance of the convolutional neural network. The kernel size of the 2D convolutional layer is selected as 32×32 according to the size of the IC curve image. The image feature vector is extracted by the 2D-CNN branch, and the output calculation is as follows:

[0109] D f = f(w f h(t) + b f ) (10)

[0110] where w f is the weight matrix and b f represents the bias. Based on the above 2D-CNN module, the IC curve picture can be recognized to obtain relevant health features for SOH prediction. Similarly, the 2D-CNN model used consists of three neuron convolutional layers and different 2D-CNN convolutional layers, two pooling layers, and two fully connected layers with 64 and 32 neurons respectively, where the convolutional layer kernels of the three 2D-CNN convolutional layers are all [3×3].

[0111] S3-6: As Figure 5As shown, it is difficult for the BiLSTM and 2D-CNN models to learn complete health features. Therefore, the outputs of the above two single-structure network models are connected together through a feature fusion layer and two fully connected layers. At the same time, Dropout layers are added to some Fullyconnected layers in each model branch to maintain the stability of the model. The output vectors of the two parallel tasks are connected through the feature fusion layer to obtain M i =[D v ,D f . M i Passes through the fully connected layer and outputs, as follows:

[0112] Output = f(w L M i + b L ) (11)

[0113] Among them, Output is the predicted value of the battery SOH, w L is the weight matrix, and b L is the corresponding bias.

[0114] On the one hand, the architecture of the multi-task prediction SOH model reflects the influence of different features on battery aging by separately inputting two different types of data, namely image and time-series data. On the other hand, the multi-task architecture makes different battery SOH prediction tasks independent of each other, and different branches extract different data features. This method avoids the information chaos caused by a large amount of data.

[0115] In step S4, the evaluation method for outputting the health state SOH of the lithium-ion battery is as follows:

[0116] To evaluate the prediction ability of the MTSOHFNN model for SOH, the coefficient of determination (R 2 ), the mean absolute percentage error MAE, and the root mean square error RMSE are used to evaluate the SOH results. Among them, y i and represent the true value and the estimated value respectively, represents the average value:

[0117]

[0118] The prediction evaluation results of four batteries in the NASA dataset are shown in Table 2. The worst RMSE, MAPE, and R 2 of the MTSOHFNN model for estimating the SOH of batteries B0005, B0006, B0007, and B0018 are 0.653%, 0.85%, and 0.94 respectively. Although there are certain differences in the maximum error values between different cycles, it may be due to the different complexities of SOH decay during different cycles. Therefore, R is introduced2 As an indicator predicted as a whole, it can be seen that the fitting capabilities of the four batteries are actually similar.

[0119] Table 2 Evaluation of SOH Prediction Results for NASA Dataset

[0120]

[0121] To compare the superiority of the proposed MTSOHFNN model of the present invention, it is compared with the commonly used GRU, 2D-CNN, and BiLSTM models in the current field of battery SOH estimation to verify the accuracy of the proposed model. Among them, the GRU and BiLSTM models use time-series data for prediction, and the 2D-CNN uses image data for prediction. The prediction results are as Figure 6(a) - Figure 6(d) shown. It can be seen that among the four batteries, the prediction accuracy of the proposed MTSOHFNN model is better than the other three comparison methods, which also indirectly shows that the training method of combining image and time-series data has a certain contribution to improving the SOH prediction accuracy.

[0122] In summary, this method proposes a method for predicting the health state of lithium-ion batteries based on a multi-source data multi-task learning model. The IC curve image and time-series data are jointly used as the input of the SOH prediction model. The advantages of 2D-CNN and BiLSTM in image and time-series prediction are respectively utilized to achieve the SOH prediction of multi-source data. The two prediction models are combined through a feature fusion method to realize a multi-task training model, achieve the mutual compensation and optimization of multi-source features, and effectively improve the prediction of the health state SOH of lithium-ion batteries by data-driven methods.

[0123] Example 2

[0124] Referring to Figure 7 , this embodiment relates to a device for predicting the health state of lithium-ion batteries based on a multi-source data multi-task learning model, including a memory and one or more processors. An executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the method for predicting the health state of lithium-ion batteries based on the multi-source data multi-task learning model of Example 1.

[0125] Example 3

[0126] This embodiment relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the method for predicting the health state of lithium-ion batteries based on the multi-source data multi-task learning model of Example 1.

[0127] In this specification, the illustrative descriptions of the present invention do not necessarily refer to the same embodiments or examples. Those skilled in the art can combine and assemble different embodiments or examples described in this specification. In addition, the content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the implementation cases. The protection scope of the present invention also includes equivalent technical means that those skilled in the art can think of based on the inventive concept of the present invention.

Claims

1. A lithium-ion battery health status prediction method based on a multi-source data multi-task learning model, characterized in that: The following steps are involved: S1: Based on NASA lithium-ion battery data set, relevant data sets are screened to extract the discharge capacity increment curve related to the health status SOH of lithium-ion batteries, namely the IC curve image data and the time series data containing parameters such as temperature-voltage; S2: preprocessing the extracted data related to the health status of the lithium-ion battery to generate a SOH training set and a test set; S3: The 2D-CNN module is used to train the IC curve image data, and the BiLSTM module is used to train the time series data. The outputs of the two modules are connected through feature fusion to establish a multi-task SOH prediction neural network MTSOHFNN model; S4: Use the SOH training set and test set for training and testing, and output the evaluation results of the health status SOH of the lithium-ion battery.

2. A lithium-ion battery health status prediction method based on a multi-source data multi-task learning model as claimed in claim 1, characterized in that: In step S1, the specific content of the constructed NASA data set is as follows: The dataset used for lithium-ion battery SOH prediction is divided into image data part and time series data part; S1-1: The capacity increment curve transforms the voltage platform that is difficult to identify in the battery voltage and capacity change curve into an easily identifiable capacity increment peak; the capacity increment analysis method obtains the capacity change corresponding to the unit voltage by calculating: In the formula, Q k , Q k-1 Respectively represent the capacity change at time k and k-1, V B,k ,V B,k-1 They represent the battery terminal voltage values ​​at time k and k-1 respectively; however, in the data acquisition process, the real-time capacity change is difficult to acquire, so the current and time difference are used for replacement: Where T represents the discharge time, I B Indicates the current during battery charging and discharging; S1-2: For the time series data, we first select the first-order RC equivalent circuit model and identify it using the least squares method based on the NASA data set. We collect the identification results of each cycle, the ohmic internal resistance R0 of the lithium-ion battery, and the polarization resistance R p And the polarization capacitance C of the lithium-ion battery p ; In addition, based on the NASA data set, the slope of the discharge voltage curve Rate is selected. V 、Temperature change rate calculation Rate T And the average discharge time T mean Three types of data including time-voltage-temperature dimensions are used as part of the timing input.

3. The method for predicting the health status of a lithium-ion battery based on a multi-source data multi-task learning model according to claim 2, characterized in that: In step S2, the data related to the health status of the lithium-ion battery is preprocessed as follows: S2-1: For the extracted IC curve image, since the NASA data sampling time is not fixed and the interval is large, in order to avoid ignoring important information on the IC curve, the collected voltage data is linearly interpolated and the minimum sampling interval is set to prevent dV B Noise steps caused by being too small; In order to avoid the influence of noise on the prediction accuracy of SOH during the training process, the Savitzky-Golay filtering method is introduced to reduce the noise of the IC curve. Savitzky-Golay fits a low-order polynomial to a sliding window of adjacent data points through the linear least squares method. The local fitting order is set to third-order fitting, and the sliding window size is 60 sampling points. S2-2: In the time series data, it is necessary to verify the correlation between the data and the health status SOH of the lithium-ion battery to avoid redundant data input causing additional computing power burden on the prediction model; therefore, the Pearson coefficient method PCC is used to measure the linear relationship between discrete data X and target data Y (SOH), and its calculation process is as follows: in, and Respectively represent the average values ​​of discrete data X and target data Y, and n represents the sequence length; the closer the absolute value of the PCC coefficient is to 1, the higher the correlation with the target data; the minimum-maximum normalization is performed on the filtered data, and the formula is as follows: Where, X nrom is the normalized result, X min and X max are the maximum and minimum values ​​in the data respectively.

4. A lithium-ion battery health status prediction method based on a multi-source data multi-task learning model as claimed in claim 3, characterized in that: In step S3, the process of establishing a multi-task SOH prediction neural network MTSOHFNN model is as follows: S3-1: Multi-task SOH prediction neural network trains the health status prediction model by designing a multi-task model. The 2D-CNN branch and the BiLSTM branch are used to extract the aging characteristics of lithium-ion batteries in different types of data. Finally, the outputs of the two networks are connected together through a feature fusion layer and two fully connected layers to construct the MTSOHFNN model. S3-2: The specific process of training the multi-task SOH prediction neural network MTSOHFNN model for health status SOH prediction is as follows: Construct a data set corresponding to IC curve image-SOH, and train it with the first 65% of the data set as the training set and the last 35% as the test set; Construct the data set corresponding to the time series data-SOH, and construct the data set according to the same division method; Combine the two datasets to build a datastore with image-time series data; Set the number of iterations and learning rate for the training process; Inputting the established data set into the established MTSOHFNN model for training to obtain a trained MTSOHFNN model; During the training process, the 2D-CNN and BiLSTM in the MTSOHFNN model are trained using the same mean square error loss function, and the mean square error loss function expression is: Among them, MSE is the mean square error, n is the number of samples, and y i is the actual prediction result of the i-th data sample, is the theoretical prediction value of the ith data sample.

5. The method for predicting the health status of a lithium-ion battery based on a multi-source data multi-task learning model according to claim 1, characterized in that: In step S4, the evaluation method of outputting the health status SOH of the lithium-ion battery is as follows: In order to evaluate the prediction ability of MTSOHFNN model for SOH, the determination coefficient R 2 The mean absolute percentage error (MAE) and the root mean square error (RMSE) are used to evaluate the SOH results, where y i and represent the true value and the estimated value respectively. Represents the average value:

6. A lithium-ion battery health status prediction device based on a multi-source data multi-task learning model, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the lithium-ion battery health status prediction method based on a multi-source data multi-task learning model as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the method for predicting the health status of a lithium-ion battery based on a multi-source data multi-task learning model as described in any one of claims 1 to 5 is implemented.

Images