Multi-task learning lithium battery state combined prediction system and method

Abstract

The invention belongs to the technical field of lithium battery health state evaluation, and discloses a multi-task learning lithium battery state combined prediction system and method, and the system comprises a sharing layer which is used for extracting battery features from lithium battery operation monitoring data; the sharing layer comprises a plurality of feature extraction modules; the multi-task layer is used for predicting the state of the lithium battery by taking the battery characteristics extracted by the sharing layer as input data after corresponding weights of the battery characteristics obtained by the gating layer are weighted; the multi-task layer comprises more than two task prediction modules and is used for predicting more than two battery state parameters; the gating layer comprises gating units with the same number as the task prediction modules, each gating unit takes lithium battery operation monitoring data as input, and the corresponding weight of each battery feature is obtained through a Softmax function. According to the invention, combined prediction of multiple battery states can be realized, and the accuracy is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery health status assessment, and relates to a multi-task learning joint prediction system and method for lithium battery status (capacity and remaining service life). Background Art

[0002] With the advancement of science and technology, global energy demand continues to rise, and the widespread use of fossil fuels has caused serious environmental problems and resource depletion. Therefore, new energy technologies, especially renewable energy and clean energy technologies, have received widespread attention since their advent. These technologies play a vital role in mitigating climate change and environmental pollution, as well as promoting sustainable development. As a new energy storage technology, lithium-ion batteries have been widely used in mobile devices, electric vehicles, and energy storage systems due to their advantages such as high energy density, long cycle life, low self-discharge rate, and lightweight. In the life cycle of lithium-ion batteries, their capacity and remaining useful life (RUL) are two key parameters to measure their performance. However, with the increase of usage time, the capacity and life of lithium batteries will gradually decrease, and there is a close correlation between the two. These parameters largely determine the safety and reliability of electric vehicles. Since these parameters cannot be directly measured by sensors, it has become an important research field to achieve real-time and accurate estimation of lithium-ion battery capacity and future prediction of RUL. Research in this area not only helps to improve consumers' driving experience and promote the sustainable development of the electric vehicle industry, but also has important practical significance for the batch recycling of batteries and the rational allocation of resources.

[0003] The estimation of state parameters of lithium-ion batteries, such as battery capacity degradation estimation and RUL prediction, is still a very challenging task due to the complex electrochemical reactions inside the battery. The existing estimation methods are mainly divided into electrochemical methods, equivalent circuit methods and data-driven methods. The electrochemical method can comprehensively estimate multiple parameters of the battery, but since it needs to deal with a large number of parameters and complex partial differential equations, this type of model is mainly suitable for theoretical analysis. On the other hand, the equivalent circuit method is more complicated in parameter identification and matrix calculation, and its accuracy may be affected as the battery ages, so it has certain limitations in practical applications. The data-driven method overcomes the complex modeling and parameter identification problems and has a good application prospect. In addition, most of the traditional data-driven method predictions only consider the future target point, without considering the auxiliary role of the information near the target point on the prediction target; and the current data-driven method can only realize the prediction of a single state parameter of lithium-ion batteries, that is, the single-task mode. The data-driven method of the single-task mode limits the performance of these models and may lead to insufficient prediction accuracy. Summary of the invention

[0004] The purpose of the present invention is to provide a multi-task learning lithium battery state joint prediction system to address the problems existing in the above-mentioned prior art, so as to realize the simultaneous prediction of multiple state parameters (including capacity and remaining service life) of the lithium battery.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions to achieve it.

[0006] The multi-task learning lithium battery state joint prediction system provided by the present invention includes:

[0007] A shared layer is used to extract battery features from lithium battery operation monitoring data; the shared layer includes a plurality of feature extraction modules; each type of lithium battery operation monitoring data is configured with a feature extraction module to extract battery features from the corresponding type of battery operation monitoring data;

[0008] A multi-task layer, which is used to predict the state of the lithium battery by taking the battery features extracted by the sharing layer and the corresponding weights of each battery feature obtained by the gating layer as input data; the multi-task layer includes more than two task prediction modules, which are used to predict more than two battery state parameters; each task prediction module has the same structure, including a convolutional neural network, an LSTM network and an attention mechanism layer;

[0009] The gating layer includes the same number of gating units as the task prediction modules. Each gating unit takes the lithium battery operation monitoring data as input and obtains the corresponding weight of each battery feature through the Softmax function.

[0010] In the above shared layer, the structure of each feature extraction module is the same, and all use LSTM network.

[0011] In the above-mentioned shared layer, the lithium battery operation monitoring data is mainly collected from the battery based on the online power battery test platform and the battery management system (Battery Management System, BMS), including the charging voltage, discharging voltage, charging current, discharging current, rated voltage, rated capacity, charging capacity (acceptable capacity under charging conditions), discharging capacity (capacity that can be released under discharging conditions), charging time, discharging time and ambient temperature corresponding to several charge and discharge cycles. In particular, for voltage data and current data, voltage / current characteristic values ​​can also be obtained through mathematical statistics, including maximum value, average value, skewness, and kurtosis. Therefore, for charging voltage, the characteristics can be selected from the maximum value of charging voltage, the average value of charging voltage, the skewness of charging voltage or / and the kurtosis of charging voltage; for discharging voltage, the characteristics can be selected from the maximum value of discharge voltage, the average value of discharge voltage, the skewness of discharge voltage or / and the kurtosis of discharge voltage; for charging current, the characteristics can be selected from the maximum value of charging current, the average value of charging current, the skewness of charging current or / and the kurtosis of charging current; for discharging current, the characteristics can be selected from the maximum value of discharge current, the average value of discharge current, the skewness of discharge current or / and the kurtosis of discharge current.

[0012] In the above-mentioned multi-task layer, the convolutional neural network is used to capture the features in the input data that are most relevant to the target, thereby reducing the noise in the input data; the convolutional neural network used in the present invention is a one-dimensional convolutional neural network (CNN).

[0013] In the above multi-task layer, the lithium battery status parameters include the lithium battery capacity and the remaining service life of the lithium battery.

[0014] The present invention also provides a training method for the multi-task learning lithium battery state joint prediction system, which comprises the following steps:

[0015] S1 builds a training data set and divides it into a training set and a test set;

[0016] S2 uses the training set to train the multi-task learning lithium battery state joint prediction system, and obtains the system's joint loss function, and optimizes the system parameters according to the joint loss function; the system's joint loss function is obtained by weighted addition of the corresponding loss functions obtained by each task prediction module;

[0017] S3 uses the test set data to test the trained lithium battery state prediction system.

[0018] The above step S1 includes the following sub-steps:

[0019] S11 obtains historical operation monitoring data of lithium batteries;

[0020] S12 pre-processes the historical operation monitoring data of the lithium battery;

[0021] S13 forms a feature set with the historical operation monitoring data of the lithium battery, and simultaneously creates a battery status label; the feature set and the corresponding label constitute a training data set;

[0022] S14 divides the training data set into a training set and a test set according to a set ratio;

[0023] S15 normalizes the data in the training set and the test set respectively.

[0024] In the above step S12, the preprocessing is mainly to remove invalid data, and at the same time fill in the data with the average value of the position data before and after the removed data position to improve the data quality.

[0025] In the above step S13, the lithium battery historical operation monitoring data includes charging voltage, discharging voltage, charging current, discharging current, rated voltage, rated capacity, charging capacity, discharging capacity charging time, discharging time, ambient temperature and other types of data of several charge and discharge cycles; for charging voltage, discharging voltage, charging current, and discharging current, relevant mathematical statistics need to be performed to obtain the maximum value, average value, skewness or / and kurtosis; therefore, the maximum value of charging voltage, the average value of charging voltage, the skewness of charging voltage, the kurtosis of charging voltage, the maximum value of discharging voltage, the average value of discharging voltage, the skewness of discharging voltage, the kurtosis of discharging voltage, the maximum value of charging current, the average value of charging current, the skewness of charging current, the kurtosis of discharging voltage, the maximum value of discharging current, the average value of discharging current, the skewness of discharging current, the kurtosis of discharging current, the rated voltage, rated capacity, charging capacity, discharging capacity, charging time, discharging time, ambient temperature and other types of data are selected to form a feature set; at the same time, the maximum discharge capacity and the remaining service life of each cycle (the number of life end times minus the number of current cycles) are used as labels. The feature set and the corresponding labels constitute a training data set. For each battery sample in the training data set, the lithium battery operation monitoring data corresponding to several charge and discharge cycles before the cycle to be predicted is used as the input data of the system.

[0026] In the above step S2, the joint loss function of the system is expressed as:

[0027] L(total)=∑W i L i

[0028] Among them, L(total) represents the joint loss function; W i represents the weight coefficient corresponding to the i-th task, L i represents the loss function of the i-th task;

[0029] The loss function used for each task in the present invention is mean square error (MSE).

[0030] During the training iterations, the current loss value of each task is compared with the result of the previous iteration and a ratio is calculated. If this ratio is higher, it means that the task is more challenging to optimize. In this way, the optimization difficulty of each task can be quantitatively evaluated. In the backpropagation phase, based on these evaluation results, the gradient values ​​corresponding to those tasks that are more difficult to optimize are multiplied by a larger specified weight factor to enhance their role in parameter updates. Correspondingly, for those tasks that are less difficult to optimize, their gradient values ​​are multiplied by a smaller specified weight factor to reduce their impact on parameter updates.

[0031] In one implementation, the battery status parameters include the lithium battery capacity and the remaining life of the lithium battery. In this case, the joint loss function of the system is:

[0032] L(total)=W Capacity L Capacity +W RUL L RUL

[0033] Among them, L(total) represents the joint loss function; W Capacity Represents the weight coefficient of the capacity prediction task; L Capacity represents the loss function of the capacity prediction task; W RUL Represents the weight coefficient of the remaining useful life prediction task; L RUL Represents the loss function for the remaining useful life prediction task.

[0034] During the training process, the upper limit of the number of training iterations is used as the training end condition.

[0035] In the above step S3, the prediction system is tested using the test set, and it is determined whether the joint loss function is less than the set loss threshold. If it is less than the set threshold, the prediction system meets the training requirements; if it is greater than the set threshold, the training set is expanded, and then return to step S2 to continue training.

[0036] The present invention also provides a multi-task learning lithium battery state joint prediction method, which comprises the following steps:

[0037] Step 1: Obtain the operation monitoring data of the lithium battery to be predicted and preprocess and normalize the data;

[0038] Step 2: Input the normalized lithium battery operation monitoring data to be predicted into the above lithium battery state prediction system to obtain the prediction results of each battery state operation parameter.

[0039] In the above step 1, the operation monitoring data of the lithium battery to be predicted refers to the operation monitoring data corresponding to several charge and discharge cycles before the charge and discharge cycle to be predicted. The operation monitoring data of the lithium battery to be predicted is preprocessed, mainly to eliminate invalid data, and at the same time fill it with the average value of the position data before and after the eliminated data position to improve the data quality; the normalization operation is: use the maximum and minimum values ​​in the historical operation monitoring data of the lithium battery to normalize the preprocessed operation monitoring data of the lithium battery to be predicted.

[0040] In the above step 2, the normalized lithium battery operation monitoring data to be predicted is input into the trained lithium battery state prediction system, and the output of each task prediction module is the battery state parameter of the charge and discharge cycle to be predicted.

[0041] So far, the present invention has realized a multi-task learning lithium battery state joint prediction system. The multi-task learning lithium battery state joint prediction system of the present invention is divided into a bottom parameter sharing layer and a top multi-task layer. The multiple feature extraction modules set in the bottom parameter sharing layer have powerful feature extraction capabilities and extract features from the input data respectively; the extracted features are weighted by the weight parameters output by the corresponding gated units of the gated layer and assembled as the input of the corresponding task prediction module; therefore, each task prediction module can perform specific learning tasks respectively; secondly, the prediction system also introduces a one-dimensional convolutional neural network and an attention mechanism to capture the most relevant features of the input data to the target, thereby reducing the noise in the input data. At the same time, multiple one-dimensional convolutional layers are used to learn the output sequence of the parameter sharing layer and convert it into feature sequences of different abstraction levels. The deep convolutional layer receives the features extracted by the previous convolutional layer as input, so the extracted features are more abstract and are used for prediction tasks that are far away from the historical capacity data. The attention mechanism layer allows the system to pay more attention to the time steps that have a greater impact on the results.

[0042] Compared with the prior art, the multi-task learning lithium battery state joint prediction system and method provided by the present invention has the following benefits:

[0043] (1) The present invention promotes learning by mining the relationship between multiple learning tasks (for example, between two different learning tasks of battery capacity and battery life), and sharing, complementing, and learning related information through a bottom parameter sharing layer, thereby achieving better results than traditional single-task learning. In addition, multi-task learning can use the correlation between task labels to supervise the training of prediction models, reduce the impact of noise, and improve the generalization of the battery status joint prediction system.

[0044] (2) The bottom parameter sharing layer takes all learning tasks into consideration at the same time, which avoids the overfitting of single tasks in traditional learning to a certain extent and improves the generalization of the system. In addition, LSTM can effectively capture the potential relationship between the previous and next information of a sequence, efficiently integrate the implicit features between different time series, and extract more universal high-dimensional spatial features.

[0045] (3) Learn multiple tasks at one time, which is highly efficient;

[0046] (4) The introduction of convolutional neural networks and attention mechanisms makes feature extraction more accurate, while allowing the system to pay more attention to the parts that have a greater impact on the output, thereby improving the accuracy, speed, and generalization of predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the structure of the lithium battery state joint prediction system for multi-task learning;

[0048] Figure 2 This is a schematic diagram of the LSTM network structure;

[0049] Figure 3 It is a schematic diagram of the principle of one-dimensional convolutional neural network;

[0050] Figure 4 Schematic diagram of the attention mechanism principle;

[0051] Figure 5 This is a flowchart of the training data set;

[0052] Figure 6 The capacity prediction results of No. 2 battery in different life cycles;

[0053] Figure 7 This is the prediction result for the full life cycle of No. 2 battery. DETAILED DESCRIPTION

[0054] The present invention is further described below in conjunction with the accompanying drawings.

[0055] Example 1

[0056] Most traditional machine learning algorithms are single-task learning. For complex problems, single-task learning can also decompose the problem into simple and independent sub-problems to solve separately, and then merge the results. However, this ignores the correlation between different problems. Many problems are also correlated with each other. Even if they can be decomposed, the sub-problems can be linked together. Treating real-life problems as independent single tasks ignores the rich correlation information between problems. Multi-task learning was born to solve this problem. Put multiple related tasks together to learn, and share some factors between multiple tasks. They can share the information they have learned during the learning process, which is not available in traditional single-task learning. In this embodiment, there is a strong correlation between the lithium battery capacity and the remaining service life prediction task, so it is very suitable to use multi-task learning to complete the prediction task.

[0057] Based on this, the multi-task learning lithium battery state joint prediction system provided in this embodiment is as follows: Figure 1 As shown, it includes a sharing layer, a gating layer and a multi-task layer. The sharing layer is used to extract battery features from lithium battery operation monitoring data; the multi-task layer is used to predict the lithium battery state by taking the battery features extracted by the sharing layer and the corresponding weights of each battery feature obtained by the gating layer as input data; the multi-task layer includes more than two task prediction modules, which are used to predict more than two battery state parameters.

[0058] In this embodiment, the lithium battery state includes capacity and remaining service life. Therefore, the lithium battery state prediction system provided in this embodiment mainly predicts capacity and remaining service life.

[0059] 1. Shared Layer

[0060] The shared layer includes several feature extraction modules. A feature extraction module is configured for each type of lithium battery operation monitoring data to extract battery features from the corresponding type of battery operation monitoring data.

[0061] The lithium battery operation monitoring data is mainly obtained by collecting the battery based on the online power battery test platform and the battery management system (Battery Management System, BMS), including the charging voltage, discharge voltage, charging current, discharge current, rated voltage, rated capacity, charging capacity (acceptable capacity under charging conditions), discharge capacity (capacity that can be released under discharge conditions), charging time, discharge time and ambient temperature corresponding to several charging and discharging cycles before the predicted charging and discharging cycle. In particular, for voltage data and current data, the voltage / current characteristic values, including maximum value, average value, skewness, and kurtosis, can also be obtained through mathematical statistics. Therefore, for charging voltage, the characteristics can be selected from the maximum value of charging voltage, the average value of charging voltage, the skewness of charging voltage, or / and the kurtosis of charging voltage; for discharging voltage, the characteristics can be selected from the maximum value of discharge voltage, the average value of discharge voltage, the skewness of discharge voltage, or / and the kurtosis of discharge voltage; for charging current, the characteristics can be selected from the maximum value of charging current, the average value of charging current, the skewness of charging current, or / and the kurtosis of charging current; for discharging current, the characteristics can be selected from the maximum value of discharge current, the average value of discharge current, the skewness of discharge current, or / and the kurtosis of discharge current.

[0062] In this embodiment, the lithium battery operation monitoring data serving as system input data includes the maximum charging voltage, the average charging voltage, the charging voltage skewness, the charging voltage peakness, the maximum discharge voltage, the average discharge voltage, the discharge voltage skewness, the discharge voltage peakness, the discharge capacity and the ambient temperature; for each type of data, a feature extraction module is set respectively, and the feature extraction module is used to extract battery features from the corresponding type of data.

[0063] The structure of each feature extraction module is the same, and all use LSTM networks. The structure of LSTM is as follows Figure 2 As shown. LSTM is a special recurrent neural network (RNN). Although traditional RNN can capture the long-term dependencies of sequence data in theory, it is often troubled by the problem of gradient disappearance or explosion in practice, which limits its ability to effectively model long-term patterns. As a special RNN variant, LSTM successfully overcomes the modeling difficulty of long-term dependencies by introducing a unique gating mechanism, greatly improving the learning and expression capabilities of time series data. The LSTM network used in the present invention can be found in Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput, 1997, 9: 1735-1780.

[0064] 2. Gating Layer

[0065] The gating layer includes the same number of gating units as the task prediction modules. Each gating unit takes the lithium battery operation monitoring data as input and obtains the corresponding weight of each battery feature through the Softmax function.

[0066] In this embodiment, the prediction system needs to complete two tasks: capacity and remaining service life. Therefore, two gating units (gating A and gating B) are required, and each gating unit corresponds to one task.

[0067] Each gating unit uses the original input data (here, the lithium battery operation monitoring data determined as the system input data previously) as input and outputs the weight parameter after Softmax. This parameter is used to weight the output features of the bottom parameter sharing layer and assemble them, thereby allowing different task models to utilize the bottom parameter sharing layer in different ways.

[0068] For example, after weighting the battery features extracted from the shared layer through gating A, data set A for predicting the capacity task is obtained, and after weighting the battery features extracted from the shared layer through gating B, data set B for predicting the capacity task is obtained.

[0069] 3. Multi-tasking layer

[0070] In this embodiment, since there are two prediction tasks, two task prediction modules are set in the multi-task layer. Each task prediction module has the same structure, including a convolutional neural network, an LSTM network and an attention mechanism layer.

[0071] The battery features extracted by the shared layer and the corresponding weights of the battery features obtained by the gated layer are weighted, and the reorganized features are used as input data (dataset A and data set B) and input into the convolutional neural network. The convolutional neural network is used to capture the most relevant features of the input data to the target, thereby reducing the noise in the input data and improving the accuracy of feature extraction; the convolutional neural network used in the present invention is a one-dimensional convolutional neural network (CNN), and the principle is as follows Figure 3 As shown in the figure, the input features are convolved with the convolution kernel and the extracted features are output.

[0072] The features extracted by the convolutional neural network are further input into the LSTM for data processing. The features extracted by the LSTM are further input into the attention mechanism layer.

[0073] According to different tasks, the attention mechanism is introduced before the output. The attention mechanism layer automatically identifies the influence of each input feature on the prediction result through self-learning and analysis of the input information, and generates corresponding attention weights at each time point accordingly. Figure 4As shown, during the training process, the model spontaneously learns the influence of each element in the input information on the result, forming the attention weight at each moment. As the input sequence changes, since the attention layer can calculate the attention weight according to the input value, the model can pay more attention to the changes in the input value more flexibly, thereby improving the accuracy of the prediction. Since the attention mechanism layer has the function of dynamically calculating the attention weight according to the input data, the system can adaptively adjust the focus to more sensitively capture the changes in the input value. For the attention mechanism layer used in this embodiment, see Vaswani A, Shazeer N, Parmar N, et al. Attention Is All You Need [J]. arXiv, 2017.

[0074] Finally, the attention mechanism layer outputs the prediction results of the corresponding battery status, which is the battery capacity and remaining service life.

[0075] This embodiment also provides a training method for the multi-task learning lithium battery state joint prediction system, which includes the following steps:

[0076] S1 constructs a training data set and divides it into a training set and a test set.

[0077] like Figure 5 As shown, this step includes the following sub-steps:

[0078] S11 obtains historical operation monitoring data of the lithium battery.

[0079] The lithium battery operation monitoring data in this step is mainly collected by the online power battery testing platform and the battery management system (Battery Management System, BMS).

[0080] According to the previous explanation, the lithium battery operation monitoring data here includes the charging voltage, discharge voltage, discharge capacity and ambient temperature corresponding to several charge and discharge cycles of the sample battery.

[0081] S12 pre-processes the historical operation monitoring data of the lithium battery.

[0082] The preprocessing mainly involves eliminating invalid data and filling in the data with the average value of the data before and after the eliminated data to improve data quality.

[0083] S13 forms a feature set with the historical operation monitoring data of the lithium battery and simultaneously produces a battery status label; the feature set and the corresponding label constitute a training data set.

[0084] For the charging voltage and discharging voltage, relevant mathematical statistics need to be performed to obtain the maximum value, average value, skewness and kurtosis, and obtain the maximum value of charging voltage, the average value of charging voltage, the skewness of charging voltage, the kurtosis of charging voltage, the maximum value of discharging voltage, the average value of discharging voltage, the skewness of discharging voltage, and the kurtosis of discharging voltage. Then, the maximum value of charging voltage, the average value of charging voltage, the skewness of charging voltage, the kurtosis of charging voltage, the maximum value of discharging voltage, the average value of discharging voltage, the skewness of discharging voltage, the kurtosis of discharging voltage, the discharge capacity and the ambient temperature constitute the feature set; at the same time, the maximum discharge capacity and the remaining service life of each cycle (the number of life end times minus the current number of cycles) are used as labels. The feature set and the corresponding labels constitute the training data set. For each battery sample in the training data set, the lithium battery operation monitoring data corresponding to several charge and discharge cycles before the cycle to be predicted are used as the input data of the system.

[0085] S14 divides the training data set into a training set and a test set according to a set ratio.

[0086] In this embodiment, the training data set is divided into a training set and a test set according to a ratio of 7:3.

[0087] S15 normalizes the data in the training set and the test set respectively.

[0088] The data in the training set and the test set are normalized to the interval [0, 1]; the normalization formula is:

[0089]

[0090] Among them, x0 represents the original data, x max Represents the maximum value in the data set, x min Represents the minimum value in the data set.

[0091] S2 uses the training set to train the multi-task learning lithium battery state joint prediction system, obtains the system's joint loss function, and optimizes the system parameters based on the joint loss function.

[0092] In this embodiment, the evaluation index of system performance adopts a multi-task joint loss function, which guides the optimization process of system parameters.

[0093] In this embodiment, gradient normalization is introduced to solve the problem of gradient imbalance and inconsistent convergence speed between different tasks. Therefore, the joint loss function of the system in this embodiment is obtained by weighted addition of the corresponding loss functions obtained by each task prediction module. Specifically, the joint loss function of multi-task learning is:

[0094] L(total)=∑W i L i

[0095] Among them, L(total) represents the joint loss function; W i represents the weight coefficient corresponding to the i-th task, L i represents the loss function of the i-th task.

[0096] The loss function for each task is the mean square error (MSE) calculated from the predicted value and the label value.

[0097] During the iteration of system training, the current loss value of each task is compared with the result of the previous iteration and the ratio is calculated. If this ratio is higher, it means that the task is more challenging to optimize. In this way, the optimization difficulty of each task can be quantitatively evaluated. In the back propagation stage, according to these evaluation results, the gradient values ​​corresponding to those tasks with greater optimization difficulty are multiplied by a larger weight factor to enhance their role in parameter updating. Correspondingly, for those tasks with less optimization difficulty, their gradient values ​​are multiplied by a smaller weight factor to reduce their influence on parameter updating.

[0098] Specifically, in this embodiment, the battery status parameters include the lithium battery capacity and the remaining life of the lithium battery. At this time, the joint loss function of the system is:

[0099] L(total)=W Capacity L Capacity +W RUL L RUL

[0100] Among them, L(total) represents the joint loss function; W Capacity Represents the weight coefficient of the capacity prediction task; L Capacity represents the loss function of the capacity prediction task; W RUL Represents the weight coefficient of the remaining useful life prediction task; L RUL Represents the loss function for the remaining useful life prediction task.

[0101] The sample data in the training set is input into the above system for training, and the joint loss function is obtained at the same time. Then, the system parameters are optimized using the Adam optimization algorithm based on the joint loss function. This step is repeated several times until the upper limit of the number of iterations is reached, and the system training is completed.

[0102] S3 uses the test set data to verify and test the trained lithium battery state prediction system.

[0103] In this step, the prediction system is tested using the test set, and it is determined whether the joint loss function is less than the set loss threshold. If it is less than the set threshold, the prediction system meets the training requirements; if it is greater than the set threshold, the training set is expanded, and then return to step S2 to continue training.

[0104] This embodiment also provides a multi-task learning lithium battery state joint prediction method, which includes the following steps:

[0105] Step 1: Obtain the operation monitoring data of the lithium battery to be predicted and preprocess and normalize the data.

[0106] In this step, the operation monitoring data of the lithium battery to be predicted refers to the operation monitoring data corresponding to several charge and discharge cycles before the charge and discharge cycle to be predicted. The operation monitoring data of the lithium battery to be predicted is preprocessed, mainly to eliminate invalid data, and at the same time fill it with the average value of the data before and after the eliminated data position to improve the data quality.

[0107] The normalization operation is: using the maximum and minimum values ​​in the historical operation monitoring data of the lithium battery to be predicted to normalize the preprocessed operation monitoring data of the lithium battery to be predicted.

[0108] Step 2: Input the normalized lithium battery operation monitoring data to be predicted into the above lithium battery state prediction system to obtain the prediction results of each battery state operation parameter.

[0109] In this step, the normalized lithium battery operation monitoring data to be predicted is input into the lithium battery state prediction system trained in the above steps S1-S3, and the output of each task prediction module is the battery state parameter of the charge and discharge cycle to be predicted.

[0110] In this embodiment, batteries No. 1 to No. 10 are selected for verification experiments, among which batteries No. 2, No. 6, and No. 10 are test samples, and the remaining batteries are training samples.

[0111] The training samples use LiNi 0.83 Co 0.11 Mn 0.06 O2 (NCM) commercial batteries, the specifications of the tested NCM batteries are shown in Table 1. The charging process includes a combination of constant current and constant voltage (CC and CV) modes at a charging rate of 0.5C, followed by a 30-min relaxation period and a discharge of 1C. Battery samples 1, 3-5, 7-9 were cycled at an ambient temperature of 25℃, 35℃, or 45℃ until the capacity dropped to 71% of the initial value.

[0112] Use the operation monitoring data of the training samples to train the multi-task learning lithium battery state joint prediction system according to the previous steps S1-S3.

[0113] Then, the trained multi-task learning lithium battery state joint prediction system is used to predict the LiNi with the same charge and discharge strategy at three different ambient temperatures. 0.83 Co 0.11 Mn 0.06 O2 (NCM) commercial battery, the specifications of the tested NCM battery are shown in Table 1. The charging process includes a combination of constant current constant voltage (CC and CV) modes at a charging rate of 0.5C, followed by a 30-minute relaxation period and a discharge of 1C. The batteries were cycled at different ambient temperatures, battery 2 was cycled at 25°C, battery 6 was cycled at 35°C, and battery 10 was cycled at 45°C until the capacity dropped to 71% of the initial value.

[0114] Table 1 Detailed specifications of training samples and test samples

[0115]

[0116] Table 2 shows the statistical errors of different models in predicting battery capacity, specifically for three batteries (battery 2, battery 6, and battery 10) operating under different temperature conditions. Taking battery 2 as an example, the mean absolute error (MAE) of the model is 1.36%, and the mean square error (MSE) is 6.12×10 -4 , which are lower than the results of other models. Therefore, the multi-task learning lithium battery state joint prediction system proposed in the present invention shows its robustness and excellent prediction accuracy in batteries with different operating temperatures, indicating that it is suitable for battery capacity degradation prediction.

[0117] Table 2 Statistical errors of battery capacity prediction under different models.

[0118]

[0119] For the No. 2 battery, the capacity prediction fitting effect of three different life cycles (10%, 30% and 50%) is shown as follows: Figure 6 (a)-(c) As shown in the figure, the predicted data is basically consistent with the test data, and the fitting effect is good, indicating that the overall fitting of the battery capacity by the system of the present invention is good, and its prediction accuracy and robustness are very good. It can be seen that the future capacity trajectory in the three cases can be accurately predicted. The results show that even if the amount of prior data is reduced, the prediction accuracy does not decrease significantly, which verifies that the proposed model has strong robustness in predicting battery degradation.

[0120] The full life cycle of No. 2 battery is predicted, and the prediction curve is as follows: Figure 7 shown.

[0121] Table 3 Comparison of RUL estimation results of different models.

[0122]

[0123] Taking the full life cycle prediction results of No. 2 battery as an example, it can be seen from Table 3 that the estimated values ​​of the four single-task learning models are quite close, and the MAE errors are all greater than 3; while the MAE and RMSE errors of the system of the present invention are significantly smaller than those of other models, indicating that the multi-task joint estimation method proposed in the present invention has higher accuracy and stronger adaptability. Figure 7 The prediction results of RUL are displayed more intuitively.

[0124] Therefore, the prediction effect of the system of the present invention is good, and a relatively small error is always maintained overall, indicating that the model that uses the multi-task learning method to establish the relationship between different tasks can effectively improve the accuracy and reliability of the model.

[0125] During the training process of the multi-task learning lithium battery state joint prediction system of the present invention, based on the correlation between the lithium battery capacity and the remaining service life prediction evaluation, the data of the lithium battery charging and discharging process is obtained from the battery management system (BMS) and a lithium battery capacity and remaining service life prediction data set is prepared. The multiple feature extraction modules in the bottom parameter sharing layer of the multi-task learning system extract the health indicator features related to the capacity degradation trend of the battery, and the two task prediction modules set in the top multi-task layer are used to respectively perform the learning tasks of lithium battery capacity prediction and lithium battery remaining service life prediction. At the same time, the attention mechanism is used to make the system prediction pay more attention to the part that has a greater impact on the output, so as to achieve accurate and reliable prediction of the lithium battery capacity and remaining service life.

[0126] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical enlightenment disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A multi-task learning lithium battery state joint prediction system, characterized in that: include: The shared layer is used to extract battery features from lithium battery operation monitoring data; The shared layer includes several feature extraction modules; A feature extraction module is configured for each type of lithium battery operation monitoring data to extract battery features from the corresponding type of battery operation monitoring data; A multi-task layer, which is used to predict the state of the lithium battery by taking the battery features extracted by the sharing layer and the corresponding weights of each battery feature obtained by the gating layer as input data; the multi-task layer includes more than two task prediction modules, which are used to predict more than two battery state parameters; each task prediction module has the same structure, including a convolutional neural network, an LSTM network and an attention mechanism layer; The gating layer includes the same number of gating units as the task prediction modules. Each gating unit takes the lithium battery operation monitoring data as input and obtains the corresponding weight of each battery feature through the Softmax function.

2. The multi-task learning lithium battery state joint prediction system according to claim 1 is characterized in that: The structure of each feature extraction module in the shared layer is the same, and all use LSTM networks.

3. The multi-task learning lithium battery state joint prediction system according to claim 1 is characterized in that: The lithium battery operation monitoring data in the shared layer includes the charging voltage, discharging voltage, charging current, discharging current, rated voltage, rated capacity, charging capacity, discharging capacity, charging time, discharging time and ambient temperature corresponding to several charge and discharge cycles.

4. The multi-task learning lithium battery state joint prediction system according to claim 3 is characterized in that: For the charging voltage, the characteristics are selected from the maximum charging voltage, the average charging voltage, the charging voltage skewness and / or the charging voltage kurtosis; for the discharging voltage, the characteristics are selected from the maximum discharge voltage, the average discharge voltage, the discharge voltage skewness and / or the discharge voltage kurtosis; for the charging current, the characteristics are selected from the maximum charging current, the average charging current, the charging current skewness and / or the charging current kurtosis; for the discharging current, the characteristics are selected from the maximum discharge current, the average discharge current, the discharge current skewness and / or the discharge current kurtosis.

5. The multi-task learning lithium battery state joint prediction system according to claim 1, characterized in that: The convolutional neural network uses a one-dimensional convolutional neural network.

6. The multi-task learning lithium battery state joint prediction system according to any one of claims 1 to 5, characterized in that: The lithium battery status parameters in the multi-task layer include the lithium battery capacity and the remaining service life of the lithium battery.

7. The training method of the multi-task learning lithium battery state joint prediction system according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1 builds a training data set and divides it into a training set and a test set; S2 uses the training set to train the multi-task learning lithium battery state joint prediction system, and obtains the system's joint loss function, and optimizes the system parameters according to the joint loss function; the system's joint loss function is obtained by weighted addition of the corresponding loss functions obtained by each task prediction module; S3 uses the test set data to test the trained lithium battery state prediction system.

8. The training method of the multi-task learning lithium battery state joint prediction system according to claim 7 is characterized in that: Step S1 includes the following sub-steps: S11 obtains historical operation monitoring data of lithium batteries; S12 pre-processes the historical operation monitoring data of the lithium battery; S13 forms a feature set with the historical operation monitoring data of the lithium battery, and simultaneously creates a battery status label; the feature set and the corresponding label constitute a training data set; S14 divides the training data set into a training set and a test set according to a set ratio; S15 normalizes the data in the training set and the test set respectively.

9. The training method of the multi-task learning lithium battery state joint prediction system according to claim 8, characterized in that: In step S13, some of the maximum charging voltage, average charging voltage, charging voltage skewness, charging voltage kurtosis, maximum discharge voltage, average discharge voltage, discharge voltage skewness, discharge voltage kurtosis, maximum charging current, average charging current, charging current skewness, charging current kurtosis, maximum discharge current, average discharge current, discharge current skewness, discharge current kurtosis, rated voltage, rated capacity, charging capacity, discharging capacity, charging time, discharging time, and ambient temperature are selected to constitute a feature set; at the same time, the maximum discharge capacity and remaining service life of each cycle are used as labels, and the remaining service life is obtained by subtracting the current number of cycles from the end of life number; the feature set and the corresponding labels constitute a training data set.

10. A multi-task learning lithium battery state joint prediction method, characterized in that: The following steps are involved: Step 1: Obtain the operation monitoring data of the lithium battery to be predicted and preprocess and normalize the data; Step 2: Input the normalized lithium battery operation monitoring data to be predicted into the lithium battery state prediction system according to any one of claims 1 to 6 to obtain the prediction results of each battery state operation parameter.

Images