A method for predicting the remaining service life of lithium-ion batteries

Abstract

The present invention relates to the field of lithium-ion battery capacity detection and discloses a method for predicting the remaining useful life of lithium-ion batteries. The method comprises the following steps: extracting the V, I, and T data of each battery during the charging process and processing them into the matrix shapes required by two subnets; inputting the data into the AFSC subnet and adaptively assigning a weight to each element in each charging characteristic spectrum; thirdly, inputting the data into the ConvLSTM subnet to obtain hidden information that fuses the characteristic spectrum into 20 cycle states; fusing the contributions of the two subnets through two multi-layer perceptrons to provide a highly accurate early life prediction value and guide the predictor to perform RUL prediction; the life predictor remains silent until the percentage remaining useful life reaches a threshold, at which point the predictor is activated to predict the remaining useful life. The present invention improves the model's ability to identify battery aging characteristics and provides a highly accurate prediction of the remaining useful life of the battery.

Description

Technical Field

[0001] The present invention relates to the field of lithium-ion battery capacity detection, and in particular to a method for predicting the remaining service life of a lithium-ion battery. Background Art

[0002] In recent years, lithium-ion batteries have demonstrated significant advantages in terms of high energy density, reduced memory effect, low self-discharge, and long lifespan. They have played a vital role in a wide range of applications, including electric vehicles, portable electronics, aerospace, and intelligent power systems. However, over time, lithium-ion batteries inevitably age and experience performance degradation, manifested by decreased capacity and increased internal resistance. Battery aging can lead to leakage, insulation damage, and partial short circuits. If not promptly detected, these conditions can lead to more serious consequences. Battery life prediction and health management are methods and technologies for assessing system reliability, detecting early failures, and predicting their progression under realistic lifecycle conditions. This allows users to make proactive maintenance decisions and prevent losses caused by unexpected failures. Therefore, accurate remaining useful life prediction is crucial for monitoring battery health, enabling timely replacement of failed batteries, and ensuring safe operation.

[0003] However, as a typical long-term series data, battery capacity decay poses a severe challenge in inferring the capacity decay trend in the later period using limited early cycle data. In current prediction models, most models extract HIs from the charge and discharge process, such as internal resistance, discharge power, AC impedance, etc. These HIs are usually difficult to obtain online, and HIs such as discharge power vary due to the random load of the equipment, which brings more data noise to the model and leads to poor robustness. Moreover, in the early cycles of the battery, the changes in various HIs are very small, which makes the modeling work of early life prediction more challenging. In addition, the size of the battery dataset for training the model greatly affects the generalization performance of the model.

[0004] In summary, the current remaining useful life prediction model for lithium-ion batteries is not capable of solving the long-term dependency and gradient explosion problems brought about by long-term series prediction. It is subject to key challenges such as random workload interference, inability to identify and predict small differences in early cycle data, and small datasets. As a result, the remaining useful life prediction accuracy is low, the robustness is weak, and the generalization performance of the model is poor. To this end, we propose a method for predicting the remaining useful life of lithium-ion batteries. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In view of the shortcomings of the prior art, the present invention provides a method for predicting the remaining service life of a lithium-ion battery to solve the above problems.

[0007] (2) Technical solution

[0008] To achieve the above-mentioned purpose, the present invention provides the following technical solutions:

[0009] A method for predicting the remaining service life of a lithium-ion battery comprises the following steps:

[0010] Step 1: Extract the V, I, and T data of each battery during the charging process as HI to measure the battery aging trend, and process the data into the matrix shape required by the two subnets;

[0011] Step 2: Input the data into the AFSC subnet and go through depthwise convolution, global attention, pointwise convolution, and local attention respectively. The contribution of V, I, and T to the model is adaptively adjusted and assigned corresponding weights. The local attention mechanism traverses the sampled data of different cycles under the same feature and adaptively assigns a weight to each element in each charging feature spectrum.

[0012] Step 3: Input the data into the ConvLSTM subnet. First, perform preliminary data processing through deep convolution. Global attention adaptively adjusts the input weights of feature data in 20 cyclic states. Before local attention embedding, ConvLSTM uses a "gate" structure to determine the retention, forgetting, and output of features in the input time series. The resulting feature spectrum integrates the hidden information of the 20 cyclic states.

[0013] Step 4: The contributions of the two subnetworks are fused through two multi-layer perceptrons to provide a highly accurate early lifespan prediction value N EOL , and guide the predictor to make RUL prediction;

[0014] Step 5: The life predictor remains silent until the percentage of remaining useful life N RUL,％ When the threshold reaches 10%, 7.5%, 5% or 2.5%, the predictor is activated to make a prediction of the remaining useful life.

[0015] Preferably, the first step is specifically implemented according to the following steps:

[0016] S1: The setup consists of an experimental dataset consisting of 124 commercial lithium iron phosphate / graphite batteries with a rated capacity of 1.1 Ah and a rated voltage of 3.3 V;

[0017] S2: Cycling to failure under 72 charging strategies and a fixed discharge rate in a 48-channel Abin charge-discharge cabinet and a 30°C constant temperature chamber;

[0018] S3: randomly divided into training set and test set in a ratio of 8:2;

[0019] S4: The first five cycles of each battery are regarded as the measure of the battery's healthiest state and recorded as the initial state. The data of the 15 cycles before the prediction starting point are recorded as the real-time state.

[0020] Preferably, the second step comprises the following steps:

[0021] S1: Perform depth convolution on the original 3D matrix. One convolution kernel is responsible for one channel. We take one of the channels for illustration and record it as X. (i,V) , for X (i,V) Fill with zero. (i,V) Take out the convolved sub-matrix, recorded as X (i,V) (n), the convolution process can be expressed as:

[0022]

[0023] in, and are the weights and biases from the k1th 2-D convolution kernel, ⊙ is the Hadamard product, For X (i,V) (n) is the result of the operation with the k1th convolution kernel. In order to perform a separate convolution on each curve, the width and step of the convolution kernel are set to 1, the number of convolution kernels for each convolution is 1, and X (i,V) (n) has the same shape as the convolution kernel, and the k1-th 2-D feature map output can be expressed as:

[0024]

[0025] The same operation is performed on the I and T channels. The output results will be batch normalized, max pooled and Leaky Rectified Linear Unit activated. The four layers are usually treated as a convolutional unit operation. The Leaky ReLU activation function (α = 0.05) is as follows:

[0026]

[0027] S2: The output volume is defined as V, and the global average pooling layer is used to obtain the attention scores s of the three channels n , to adaptively adjust their contribution to the model, take the feature map of the nth channel in V as V n , s n The calculation is as follows:

[0028]

[0029] Normalize the obtained attention scores to get the final global weight factor The global weight factor is used to adaptively assign weights to different feature maps, allowing the model to focus on feature extraction of important variables. The calculation is as follows:

[0030]

[0031] By connecting g n (n=1,2,···,K) will get the output volume with weights

[0032] S3: Yes Fill in zero and take the sub-matrix of the mth convolution kernel size as The convolution result of point convolution can be expressed as:

[0033]

[0034] yes The result of the k2th convolution kernel has the same shape before and after convolution. and are the weights and biases from the k2th 3-D convolution kernel respectively, and the k2th 2-D feature map output can be expressed as:

[0035]

[0036] S4: Denote the output volume of the previous convolutional layer as F, define a matrix A with the same shape as F, and the elements A in A i,j is the element in the corresponding feature map The attention weights are generated by two fully connected layers to generate the attention weight matrix A, and the corresponding elements A i,j It can be expressed as:

[0037]

[0038] Among them, δ and ω are weights, b and c are biases, and n Fc is the number of neurons, the subscript is the index of the element in the matrix, g(·) and f(·) represent the hyperbolic tangent function and the sigmoid function, respectively, which are expressed as follows:

[0039]

[0040]

[0041] The Sigmoid function controls the elements of the weight matrix within the range of 0-1, and controls the progress through the Hadamard product of the two matrices.

[0042] The size of the information flow into the next layer of the network, the output volume L lAs shown below:

[0043] L l =A⊙F l .

[0044] Preferably, the local attention in the third step is to further extract useful information from the feature spectrum after convolutional LSTM processing, so that the model focuses on the common features shared by all data frames.

[0045] Preferably, the specific steps of the third step are as follows:

[0046] The input matrix is ​​first operated by two depthwise convolution units and then embedded in the global attention mechanism. The obtained input volume is increased in dimension to become a 5-dimensional tensor to meet the input of ConvLSTM. The key equations for the gate function and data flow transmission within the ConvLSTM cell are as follows:

[0047] f t =σ(W Xf *X t +W hf *h t-1 +b f );

[0048] i t =σ(W Xi *X t +W hg *h t-1 +b i );

[0049] g t =tanh(W Xg *X t +W hg *h t-1 +b g );

[0050] o t =σ(W Xo *X t +W ho *h t-1 +b o );

[0051] C t =f t ⊙C t-1 +i t ⊙g t ;

[0052] h t =o t ⊙tanh(C t );

[0053] Among them, f represents the output of the forget gate, i represents the output of the input gate, o represents the output of the output gate, g represents the candidate memory, C represents the cell state of ConvLSTM, h represents the hidden layer output, X represents the input, W X～ and W h～ It is a 2-D convolution kernel. The subscript t indicates the corresponding time. The symbol “*” represents the convolution operation and “⊙” represents the Hadamard product. The output of ConvLSTM becomes the output of subnet 2 after adding local attention.

[0054] Preferably, the fourth step specifically includes:

[0055] The outputs of the two subnetworks are connected and defined as The remaining service life of the battery is predicted by fusing the contribution values ​​of the two subnetworks through two multi-layer perceptrons. The calculation is as follows:

[0056]

[0057] (3) Beneficial effects

[0058] Compared with the prior art, the method for predicting the remaining service life of a lithium-ion battery provided by the present invention has the following beneficial effects:

[0059] 1. This lithium-ion battery remaining service life prediction method is based on battery charging process data modeling, eliminating the interference of random workload on the model, making the model more robust, practical and generalizable.

[0060] 2. This lithium-ion battery remaining service life prediction method uses directly measurable V, I, and T data as HI. The data is simple and easy to obtain and is strongly correlated with battery life, which is conducive to real-time monitoring of battery health and life prediction.

[0061] 3. This lithium-ion battery remaining service life prediction method proposes for the first time a predictor that can be used for both early prediction of battery life and RUL.

[0062] 4. This lithium-ion battery remaining useful life prediction method uses the AFSC subnetwork to adaptively adjust the input weights of early cycle data and fuse different features. The ConvLSTM subnetwork excels at capturing spatiotemporal features from late cycle data, reducing the redundancy of memory cells, effectively capturing long-term dependencies in long-sequence data, and mitigating gradient explosion. The fusion of these two methods contributes to accurate prediction of both early and late RUL.

[0063] 5. In this lithium-ion battery remaining service life prediction method, the addition of the attention mechanism improves the model's ability to extract important information and discard useless information, further improving the model's prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of the process of the remaining useful life prediction method according to an embodiment of the present invention;

[0065] Figure 2 Capacity decay curves of 124 battery data sets used in the embodiments of the present invention;

[0066] Figure 3 The regularity differences of Examples V, I and T under different cycles of the present invention are shown;

[0067] Figure 4 This is the data input structure of the two sub-networks in the embodiment of the present invention;

[0068] Figure 5 This is a schematic diagram of the remaining battery life prediction model structure according to an embodiment of the present invention;

[0069] Figure 6 This is a schematic diagram of the structure of the ConvLSTM in an embodiment of the present invention;

[0070] Figure 7 The early prediction performance of the model for battery life under different prediction starting points in the embodiment of the present invention;

[0071] Figure 8 This is the RUL prediction performance of the model at different warning points in the embodiment of the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0073] Example

[0074] See also Figure 1-8 The method for predicting the remaining service life of a lithium-ion battery provided by an embodiment of the present invention includes the following steps:

[0075] In step 1, the V, I, and T data of each battery during the charging process are extracted as HI to measure the battery aging trend, and the data are processed into the matrix shape required by the two subnets.

[0076] Step 1 is implemented as follows:

[0077] In Step 1.1, the experimental dataset consists of 124 commercial lithium iron phosphate / graphite batteries with a rated capacity of 1.1 Ah and a rated voltage of 3.3 V. These batteries were cycled to failure in a 48-channel Abin charging and discharging cabinet and a 30°C constant temperature chamber under 72 charging strategies and a fixed discharge rate. The dataset was randomly divided into a training set and a test set in an 8:2 ratio. This means that there will be 99 battery data with 52 charging strategies and 25 battery data with 20 charging strategies, which will be used for RUL estimation and model generalization performance verification, respectively.

[0078] Figure 2 The capacity decay curves of 124 battery data sets are given.

[0079] In step 1.2, the battery's V, I, and T data exhibit varying degrees of drift as cycling progresses. Furthermore, batteries of varying lifespans exhibit significant differences in their initial state during cycling, and these regular differences are strongly correlated with battery lifespan. Therefore, the raw input data should reflect these differences across cycling states. The first five cycles of each battery are considered the benchmark for the battery's healthiest state and are recorded as the initial state. The data for the 15 cycles prior to the prediction starting point is recorded as the real-time state and compared with the battery's initial state. This data input structure can be used for both early-stage battery life prediction and later-stage real-time remaining useful life prediction, depending on the chosen starting point. This model offers significant flexibility. For example, when using V, I, and T data from the first 20 cycles for early-stage life prediction, with the prediction starting point = 20, the data structure is [1,5] + [6,20] (initial state + real-time state), with the numbers in the interval representing the cycle numbers used. For RUL prediction, for a battery with a cycle life of 1000, if the prediction starting point = 800, the data structure is [1,5]+[786,800].

[0080] Figure 3 The regularity differences of V, I and T under different cycles.

[0081] Figure 4 Enter the data structure for both subnets.

[0082] In step 2, the data is input into the AFSC subnet, which undergoes depthwise convolution, global attention, pointwise convolution, and local attention steps to adaptively adjust the contribution of V, I, and T to the model and assign corresponding weights. The local attention mechanism traverses the sampling data of different cycles under the same feature, and adaptively assigns a weight to each element in each charging feature spectrum, making the AFSC subnet highly sensitive to the changing rules at each time sampling point.

[0083] Step 2 is implemented as follows:

[0084] Step 2.1, perform depthwise convolution on the original 3D matrix, with one convolution kernel responsible for one channel. Take one of the channels (voltage) for illustration, denoted as X (i,V) In order to avoid the loss of edge data, (i,V) Fill with zero. (i,V) Take out the convolved sub-matrix, recorded as X (i,V) (n), the convolution process can be expressed as:

[0085]

[0086] in, and are the weights and biases from the k1th 2-D convolution kernel, and ⊙ is the Hadamard product. For X (i,V) (n) and the result of the k1-th convolution kernel. In order to perform a separate convolution on each curve, the width and stride of the convolution kernel are set to 1, and the number of convolution kernels for each convolution is 1. (i,V) (n) has the same shape as the convolution kernel. The k1th 2-D feature map output can be expressed as:

[0087]

[0088] The same operations are performed on the I and T channels. The output is batch normalized, max-pooled, and activated with a Leaky Rectified Linear Unit (Leaky ReLU) to mitigate gradient explosion during training. The above four layers are usually treated as a single convolutional unit. The Leaky ReLU activation function (α = 0.05) is as follows:

[0089]

[0090] Step 2.2, define the output volume of the previous step as V, and the global average pooling layer is used to obtain the attention scores s of the three channels n , to adaptively adjust their contribution to the model. Take the feature map of the nth channel in V as V n , s n The calculation is as follows:

[0091]

[0092] Normalize the obtained attention scores to get the final global weight factor The global weight factor is used to adaptively assign weights to different feature maps, allowing the model to focus on feature extraction of important variables. The calculation is as follows:

[0093]

[0094] By connecting g n (n=1,2,···,K) will get the output volume with weights Used as input to the next layer.

[0095] Step 2.3, Fill in zero and take the sub-matrix of the mth convolution kernel size as The convolution result of point convolution can be expressed as:

[0096]

[0097] Similarly, yes The result of the k2-th convolution kernel has the same shape before and after convolution. and are the weights and biases from the k2th 3-D convolution kernel respectively. The k2th 2-D feature map output can be expressed as:

[0098]

[0099] In step 2.4, the output volume of the previous convolutional layer is represented as F. In order to explore the contribution of the data of each time sampling point in the variable curve to the model, a matrix A with the same shape as F is defined, and the elements A in A are i,j is the element in the corresponding feature map The attention weights of . In this paper, the attention weight matrix A is generated by two fully connected layers, and the corresponding elements A i,j It can be expressed as:

[0100]

[0101] Among them, δ and ω are weights, b and c are biases, and n Fc is the number of neurons. The subscript is the index of the element in the matrix. g(·) and f(·) represent the hyperbolic tangent function and the sigmoid function, respectively, and are expressed as follows:

[0102]

[0103]

[0104] The Sigmoid function controls the elements of the weight matrix within 0-1, and controls the size of the information flow into the next layer of the network through the Hadamard product of the two matrices. The output volume L l As shown below:

[0105] L l =A⊙Fl (11)

[0106] Figure 5 A schematic diagram of the structure of a new lithium-ion battery remaining service life prediction model based on AFSC-ConvLSTM disclosed in the present invention is given.

[0107] In step 3, the data is input into the ConvLSTM subnetwork. This data is initially processed through deep convolution. Global attention adaptively adjusts the input weights of feature data in 20 recurrent states. Before local attention embedding, the ConvLSTM uses a "gate" structure to determine the retention, forgetting, and output of features in the input time series. The resulting feature spectrum incorporates the hidden information of the 20 recurrent states. Local attention further extracts useful information from the feature spectrum processed by the convolutional LSTM, allowing the model to "focus attention" on common features across all data frames.

[0108] Step 3 is implemented as follows:

[0109] In step 3.1, similar to subnet 1, the input matrix is ​​first processed through two depthwise convolutional units before being embedded in the global attention mechanism. This will not be repeated here. The resulting input volume is then dimensionalized into a 5-dimensional tensor to feed the ConvLSTM. The key equations for the gate function and data flow within the ConvLSTM cell are as follows:

[0110] f t =σ(W Xf *X t +W hf *h t-1 +b f ) (12)

[0111] i t =σ(W Xi* X t +W hg *h t-1 +b i ) (13)

[0112] g t =tanh(W Xg *X t +W hg *h t-1 +b g ) (14)

[0113] o t =σ(W Xo *X t +W ho *h t-1 +b o ) (15)

[0114] C t =f t ⊙C t-1 +i t ⊙g t (16)

[0115] h t =o t ⊙tanh(C t ) (17)

[0116] Among them, f represents the output of the forget gate; i represents the output of the input gate; o represents the output of the output gate; g represents the candidate memory; C represents the cell state of ConvLSTM; h is the hidden layer output; X represents the input; W X～ and W h～ is a 2-D convolution kernel; the subscript t indicates the corresponding time. The symbol "*" represents the convolution operation, and "⊙" represents the Hadamard product. The output of the ConvLSTM is added with local attention and becomes the output of subnet 2. This is not detailed here.

[0117] Figure 6 The structural diagram of ConvLSTM is given.

[0118] Step 4: The contributions of the two subnetworks are fused through two multilayer perceptrons to provide a highly accurate early lifespan prediction value N. EOL , and guide the predictor to make RUL predictions.

[0119] Step 4 specifically includes:

[0120] The outputs of the two subnetworks are connected and defined as The remaining service life of the battery is predicted by fusing the contribution values ​​of the two subnetworks through two multi-layer perceptrons. The calculation is as follows:

[0121]

[0122] The symbolic interpretation is similar to that of Equation (8), except that f(·) in Equation (18) is represented by the activation function shown in Equation (3).

[0123] Figure 7 The early prediction performance of the model for battery life under different prediction starting points of the method disclosed in the present invention is given.

[0124] Step 5: The lifespan predictor remains silent until the percentage of remaining useful life N RUL,％ (Relative N EOL When the threshold value reaches 10%, 7.5%, 5% or 2.5%, the predictor is activated to predict the remaining useful life.

[0125] Figure 8 The RUL prediction performance of the model under different warning points of the method disclosed in the present invention is given.

[0126] The focus of the above embodiments of the present invention is: extracting the V, I and T data of each battery charging process as HI; inputting the HI in parallel into the AFSC and ConvLSTM sub-networks, AFSC adaptively adjusts the input weights of the early cycle data and fuses different features, and ConvLSTM is good at capturing spatiotemporal features from the later cycle data and capturing the long-term dependencies in the HI sequence; the two sub-networks are respectively embedded with the global attention mechanism and the local attention mechanism, and the key data frames in the V, I and T matrices are identified through the global attention mechanism. The local attention mechanism helps to select important features accurately to each data sampling point under the key data frame, thereby improving the model's ability to identify battery aging characteristics; the outputs of the last two sub-networks are constructed through two multi-layer perceptrons to provide a high-precision prediction of the remaining battery service life.

[0127] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining service life of a lithium-ion battery, characterized in that: The following steps are involved: Step 1: Extract the V, I, and T data of each battery during the charging process as HI to measure the battery aging trend, and process the data into the matrix shape required by the two subnets; Step 2: Input the data into the AFSC subnet and go through depthwise convolution, global attention, pointwise convolution, and local attention respectively. The contribution of V, I, and T to the model is adaptively adjusted and assigned corresponding weights. The local attention mechanism traverses the sampled data of different cycles under the same feature and adaptively assigns a weight to each element in each charging feature spectrum. Step 3: Input the data into the ConvLSTM subnet. First, perform preliminary data processing through deep convolution. Global attention adaptively adjusts the input weights of feature data in 20 cyclic states. Before local attention embedding, ConvLSTM uses a "gate" structure to determine the retention, forgetting, and output of features in the input time series. The resulting feature spectrum integrates the hidden information of the 20 cyclic states. Step 4: The contributions of the two subnetworks are fused through two multi-layer perceptrons to provide a highly accurate early lifespan prediction value N EOL , and guide the predictor to make RUL prediction; Step 5: The life predictor remains silent until the percentage of remaining useful life X RUL,％ When the threshold reaches 10%, 7.5%, 5% or 2.5%, the predictor is activated to make a prediction of the remaining useful life.

2. The method for predicting the remaining useful life of a lithium-ion battery according to claim 1, wherein: The first step is specifically implemented according to the following steps: S1: The setup consists of an experimental dataset consisting of 124 commercial lithium iron phosphate / graphite batteries with a rated capacity of 1.1 Ah and a rated voltage of 3.3 V; S2: Cycling to failure under 72 charging strategies and a fixed discharge rate in a 48-channel Abin charge-discharge cabinet and a 30°C constant temperature box; S3: randomly divided into training set and test set in a ratio of 8:2; S4: The first five cycles of each battery are regarded as the measure of the battery's healthiest state and recorded as the initial state. The data of the 15 cycles before the prediction starting point are recorded as the real-time state.

3. The method for predicting the remaining useful life of a lithium-ion battery according to claim 1, wherein: The second step has the following steps: S1: Perform depth convolution on the original 3D matrix. One convolution kernel is responsible for one channel. We take one of the channels for illustration and record it as X. (i,V) , for X (i,V) Fill with zeros, starting from X (i,V) Take out the convolved sub-matrix, recorded as X (i,V) (n), the convolution process can be expressed as: in, and are the weights and biases from the k1th 2-D convolution kernel, ⊙ is the Hadamard product, For X (i,V) (n) is the result of the operation with the k1th convolution kernel. In order to perform a separate convolution on each curve, the width and step of the convolution kernel are set to 1, the number of convolution kernels for each convolution is 1, and X (i,V) (n) has the same shape as the convolution kernel, and the k1-th 2-D feature map output can be expressed as: The same operation is performed on the I and T channels. The output results will be batch normalized, max pooled and Leaky ReLU activated. The four layers are usually treated as a convolution unit operation. The Leaky ReLU activation function is as follows: S2: The output volume is defined as V, and the global average pooling layer is used to obtain the attention scores s of the three channels n , to adaptively adjust their contribution to the model, take the feature map of the nth channel in V as V n , s n The calculation is as follows: The obtained attention scores are normalized to obtain the final global weight factor. The global weight factor is used to adaptively assign weights to different feature maps, allowing the model to focus on feature extraction of important variables. The calculation is as follows: By connecting g n (n=1,2,…,K) will get the output volume with weights S3: Yes Fill in zero and take the sub-matrix of the mth convolution kernel size as The convolution result of point convolution is expressed as: yes The result of the k2th convolution kernel has the same shape before and after convolution. and are the weights and biases from the k2th 3-D convolution kernel, and the k2th 2-D feature map output is expressed as S4: Denote the output volume of the previous convolutional layer as F, define a matrix A with the same shape as F, and the elements A in A i,j is the element in the corresponding feature map The attention weights are generated by two fully connected layers to generate the attention weight matrix A, and the corresponding elements A i，j It can be expressed as: in, and is the weight, and c i，j All are biased, n Fc is the number of neurons, the subscript is the index of the element in the matrix, g(·) and f(·) represent the hyperbolic tangent function and the sigmoid function, respectively, which are expressed as follows: The Sigmoid function controls the elements of the weight matrix within 0-1, and controls the size of the information flow into the next layer of the network through the Hadamard product of the two matrices. The output volume L l As shown below: L l =A⊙F l 。 4. The method for predicting the remaining useful life of a lithium-ion battery according to claim 1, wherein: The local attention in the third step is to further extract useful information from the feature spectrum after convolutional LSTM processing, so that the model focuses on the common features of all data frames.

5. The method for predicting the remaining useful life of a lithium-ion battery according to claim 1, wherein: The specific steps of the third step are as follows: The input matrix is ​​first operated by two depthwise convolution units and then embedded in the global attention mechanism. The obtained input volume is increased in dimension to become a 5-dimensional tensor to meet the input of ConvLSTM. The key equations for the gate function and data flow transmission within the ConvLSTM cell are as follows: f t =σ(W Xf *X t +W hf *h t-1 +b f ) i t =σ(W Xi *X t +W hg *h t-1 +b i ) g t =tanh(W Xg *X t +W hg *h t-1 +b g ) o t =σ(W Xo *X t +W ho *h t-1 +b o ) C t =f t ⊙C t-1 +i t ⊙g t h t =o t ⊙tanh(C t ) Among them, f t Represents the output of the forget gate, i t Represents the input gate output, o t Represents the output of the output gate, g t represents candidate memory, C t and C t-1 Represents the cell state of ConvLSTM, h t-1 is the hidden layer output, X t Represents input, W Xf 、W Xi 、W Xg 、W Xo and W hf 、W ho 、W hg It is a 2-D convolution kernel. The subscript t indicates the corresponding time. The symbol "*" represents the convolution operation and "⊙" represents the Hadamard product. The output of ConvLSTM becomes the output of subnet 2 after adding local attention.

6. The method for predicting the remaining useful life of a lithium-ion battery according to claim 1, wherein: The fourth step specifically includes: The outputs of the two subnetworks are connected and defined as Two multilayer perceptrons are used to generate an early prediction of the remaining useful life of a lithium-ion battery, as shown below:

Images