Method and device for predicting remaining service life of lithium battery of electric excavator

Abstract

The invention relates to the technical field of lithium battery service life, in particular to an electric excavator lithium battery remaining service life prediction method and device, and the method comprises the steps: obtaining the capacity time series data of a lithium battery based on an electric excavator coulombmeter sensor; constructing full step-by-step real-time decomposition sampling to perform information leakage-free real-time decomposition on the lithium battery capacity time sequence so as to relieve the capacity regeneration phenomenon of the lithium battery and constructing a training set and a test set; building a deep learning model of a tight coupling convolution Informer, and learning local fluctuation and global degeneration of a capacity time sequence; evaluating the prediction precision of the deep learning model by using an evaluation index; and estimating the test set by using the trained network, and verifying the performance of the model. The end-to-end residual service life prediction of the lithium battery of the electric excavator can be realized, so that the problems that the dynamic change of the capacity regeneration phenomenon cannot be accurately captured and the prediction precision is low due to the lag of the existing prediction technology are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery usage prediction for electric engineering equipment, and in particular to a method and device for predicting the remaining service life of a lithium battery for an electric excavator. Background Art

[0002] Lithium-ion batteries, due to their high energy density, long cycle life, and light weight, have become the primary battery type for energy storage systems in construction machinery such as electric excavators. However, lithium-ion battery performance gradually degrades over time due to factors such as operating temperature, charge and discharge rates, and battery chemistry, posing a serious threat to battery system safety. Remaining useful life prediction, a core task of battery management systems, can proactively identify potential battery issues, reduce the risk of failure, and enhance system safety. Therefore, accurate prediction of the remaining useful life of lithium-ion batteries plays a vital role in ensuring the safe and reliable operation of electric excavators.

[0003] Traditional model-based approaches for predicting the remaining useful life of lithium-ion batteries use physical models to describe the battery degradation process. However, the complexity of battery electrochemical reactions makes it difficult to establish accurate physical and chemical models. Currently, mainstream data-driven approaches rely on large amounts of training data to improve the accuracy of remaining useful life predictions. While data-driven approaches have significantly improved the accuracy and reliability of lithium-ion battery remaining useful life predictions, the capacity regeneration phenomenon that occurs during the degradation process of lithium-ion batteries has hindered further development of remaining useful life predictions. Summary of the Invention

[0004] The present invention aims to provide a method and device for predicting the remaining useful life of lithium batteries for electric excavators, aiming to address the problem of low remaining useful life prediction accuracy caused by the regeneration of lithium-ion battery capacity. To achieve this objective, the present invention provides the following technical solutions.

[0005] The present invention provides a method for predicting the remaining service life of a lithium battery of an electric excavator, comprising the following steps:

[0006] Step 1: Obtain the full-life charge and discharge cycle capacity time series data of the electric excavator lithium battery based on the coulomb counter sensor, the number of cycles at the end of the battery life, and the remaining service life;

[0007] Step 2: Perform full-step real-time decomposition sampling on the capacity time series data and construct a training set and a test set;

[0008] Based on the coulomb meter sensor, the full life cycle charge and discharge cycle capacity time series data of the electric excavator lithium battery, the number of cycles at the end of the battery life and the remaining service life are obtained;

[0009] The capacity time series data of the lithium battery is represented by S={S1, S2, S n ,…,S N}, S n The battery capacity data for the nth charge and discharge cycle;

[0010] The remaining service life RUL is expressed as: RUL=n EOL -n(1);

[0011] n EOL is the number of cycles at the end of the battery life, and n represents the nth charge and discharge cycle;

[0012] Step 2: Perform full-step real-time decomposition sampling on the capacity time series data, and construct a training set and a test set;

[0013] Step 3: Add cross-stage local attention, extended causal convolution, and transfer mechanism to the deep learning prediction model Informer to build a deep learning model with tightly coupled convolutional informer;

[0014] Step 4: Build a hybrid deep learning model of full step-by-step real-time decomposition and tightly coupled convolutional informer to learn the local volatility and global degradation of capacity time series;

[0015] Step 5: Evaluate the prediction accuracy of the hybrid deep learning model of the full step-by-step real-time decomposition-tightly coupled convolutional informer using evaluation indicators;

[0016] Step 6: Use the trained hybrid deep learning model of full step-by-step real-time decomposition-tightly coupled convolutional informer to estimate the test set, obtain the remaining service life prediction results of the lithium battery, and verify the performance of the model.

[0017] Optionally, in step 2, performing full step-by-step real-time decomposition sampling on the capacity time series data and constructing a training set and a test set includes the following steps:

[0018] Step 2.1: Take the pth lithium battery charge and discharge cycle as the prediction starting point, where p < n EOL , the capacity time series data {S1,S2,S n ,…,S N} is divided into training cycles {S1, S2, ..., S p} and test cycle {S p+1 ,S p+2 ,…,S N};

[0019] Step 2.2: Use a sliding window with a window size of m and a step size of 1 to segment the capacity time series data. The m past observations that lag behind the data to be predicted by one window are used as candidate explanatory variables, and the value to be predicted after m is used as the response variable.

[0020] Step 2.3: Segment the first sequence segment sample {S1, S2, ..., S m}∈R 1×m ,m<N is decomposed into k subsequences through variational mode decomposition,

[0021] The k subsequences are resampled in series into a time series. Then, the explanatory variable of the first sequence segment sample is The response variable is S m+1 ;

[0022] Step 2.4: Use a real-time sliding window with a window size of m and a step size of 1, and continue sliding the window along the capacity time series to obtain the second sequence segment sample {S2, S3, ..., S m+1}∈R 1×m ,m<N;

[0023] Decompose the second sequence fragment sample into k subsequences, then the explanatory variable of the second sample is The response variable is S m+2 ;

[0024] Step 2.5: Slide and decompose step by step until the last sequence fragment {S N-m ,S N-m+1 ,…,S N-1} is decomposed and resampled in series, and finally a total of (Nm)∈R (N-m)×(m*k) explanatory variable samples, ending decomposition and resampling;

[0025] Step 2.6: Use the first pm explanatory variable samples as the training set and the remaining Np explanatory variable samples as the test set.

[0026] Optionally, in step 3, the Informer deep learning model includes an encoder and a decoder. The process of adding cross-stage local attention, extended causal convolution, and transfer mechanisms to the deep learning prediction model Informer to establish a tightly coupled convolutional Informer deep learning model includes the following steps:

[0027] Step 3.1: In the encoder, a certain historical data X∈R of the pm explanatory variable samples is converted into L×d As the input of multi-head sparse cross-stage local attention, where L < (pm) is the input length and d < (m*k) is the input feature dimension;

[0028] Step 3.2: In the multi-head sparse cross-stage local attention, input X∈R L×d Divided into two parts along the feature dimension X1 serves as the input of the sparse self-attention network, and X2 is connected to the end of the multi-head sparse cross-stage local attention network after a 1×1 convolution. The outputs of the two branches X1 and X2 are concatenated through the feature dimension as the output of the entire multi-head sparse cross-stage local attention block, which can be expressed as:

[0029]

[0030] where Y∈R L×d is the output of multi-head sparse cross-stage local attention, A(X 1h ) is the scaled dot product of the h-th self-attention block, W h It is d h ×d h The linear projection matrix of , H is the number of multi-head sparse cross-stage local attention heads, d h is the dimension of each head, W c is the weight matrix of the 1×1 convolutional layer;

[0031] Step 3.3: The output of the multi-head sparse cross-stage local attention Y∈R L×d As the input of the dilated causal convolutional layer, for the fth dilated causal convolutional layer after the i-th multi-head sparse cross-stage local self-attention block, in the sequence Y∈R L×d The element y in g ∈R d ,g∈L uses the extended causal convolution operation Z with kernel size u f (y g ) can be defined as:

[0032]

[0033] where Y∈R L×d is the input sequence, y g ∈R d , g∈L is an element in Y, W f ∈R d×d′ is the weight matrix of the f-th extended causal layer, d is the input dimension, d′ is the output dimension, v is the dilation factor, and the filter of the f-th extended causal convolutional layer skips between two adjacent filter taps (2 v-1 -1) elements to effectively improve the receptive field of traditional convolution;

[0034] Step 3.4: Dilate the output of the causal convolution Z∈R L×dAs the input of the maximum pooling layer to extract the most significant local pattern in the time series, let e be the pooling time window size, s be the pooling step size, z l ′∈R d is the pooling result of the l∈Lth time period, then the maximum pooling layer can be defined as:

[0035]

[0036] The maximum pooling layer selects the maximum value of the jth feature from the continuous e time windows of the output Z of the extended causal convolution for each feature dimension j to obtain the output sequence Z′∈R L′×d ,in

[0037] Step 3.5: Use three multi-head sparse cross-stage local attentions in the encoder, and use an extended causal convolution layer and a maximum pooling layer to connect every two multi-head sparse cross-stage local attentions. There are three layers of output, namely And usually L′3<L′2<L′1. The transfer mechanism is used to splice the features of the three scales Z′1, Z′2, and Z′3 to fuse multi-scale information and enhance the capacity and time granularity perception ability, which can be expressed as:

[0038]

[0039] Step 3.6: Use the transfer mechanism to splice the features of the three scales Z′1, Z′2, and Z′3 to fuse multi-scale information and enhance the capacity time granularity perception capability. The multi-scale fusion feature is expressed as; the multi-scale fusion feature output by the transfer mechanism Input transition layer to compress feature dimension to unify feature dimension, ensure the consistency of subsequent decoder input structure, multi-scale fusion feature is expressed as:

[0040]

[0041] in is the output of the transition layer, ReLU is the activation function, BN is batch normalization, Conv1D 1×1 It is a one-dimensional convolution;

[0042] Step 3.7: Fusion of multi-scale features output by the transfer mechanism The input transition layer is used to compress the feature dimension to unify the feature dimension and ensure the consistency of the input structure of the subsequent decoder. The output of the transition layer is expressed as:

[0043]

[0044] in is the output of the transition layer, ReLU is the activation function, BN is batch normalization, Conv1D1×1 It is a one-dimensional convolution;

[0045] Step 3.8: Transition layer output A certain period of historical data X of the explanatory variable sample de ∈R T×d The decoder is fed together with the masked multi-head sparse cross-stage local attention to obtain the sequence feature learning results of the decoder;

[0046] In step 3.9, the sequence feature learning results of the decoder are input into the fully connected layer to build the final tightly coupled convolutional Informer deep learning model.

[0047] Optionally, the process of the full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model in step 4 includes the following steps:

[0048] Step 4.1: Based on the electric excavator battery management system, obtain the capacity degradation curve under different discharge conditions by using different charge and discharge conditions and a coulomb counter sensor;

[0049] Step 4.2: Processing the capacity degradation curve based on the full stepwise real-time decomposition sampling to obtain a series resampled explanatory variable sample;

[0050] Step 4.3: Train the weight parameters in the tightly coupled convolutional informer deep learning model based on the training set. When the tightly coupled convolutional informer deep learning model converges to the optimal state, stop training and save the optimal model weight parameters. During training, the mean square error (MSE) is used as the loss function of the model training. The loss function is expressed as:

[0051]

[0052] Where N is the number of samples, C i is the actual capacity value of the i-th sample, C′ i is the predicted capacity value of the i-th sample;

[0053] Step 4.4: combining the full step-by-step real-time decomposition sampling and the tightly coupled convolutional informer deep learning model, and training an optimal full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model based on the optimal tightly coupled convolutional informer deep learning model parameters;

[0054] Step 4.5: Predict the test set based on the optimal full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model to obtain a capacity degradation prediction value of the test set.

[0055] Optionally, in step 5, the evaluation indicators include absolute error, mean absolute percentage error, root mean square error and mean absolute error, and the absolute error, mean absolute percentage error, root mean square error and mean absolute error are used to evaluate the prediction accuracy of the full step-by-step real-time decomposition-tightly coupled convolutional Informer model respectively.

[0056] Optionally, in step 6, the trained full stepwise real-time decomposition-tightly coupled convolutional informer model is used to estimate the test set to obtain the remaining service life prediction result of the lithium battery, which specifically includes the following steps:

[0057] Step 6.1: Based on the capacity degradation prediction value of the test set, determine the number of cycles when the full step-by-step real-time decomposition-tightly coupled convolutional informer model predicts 80% of the rated capacity;

[0058] Step 6.2: Based on the predicted number of cycles when the rated capacity reaches 80%, the number of cycles at the predicted starting point is subtracted from the predicted number of cycles when the rated capacity reaches 80%, and the remaining useful life prediction result of the test set using the full step-by-step real-time decomposition-tightly coupled convolutional informer model is obtained;

[0059] Step 6.3: Calculate the evaluation index values of the prediction results of the full step-by-step real-time decomposition-tightly coupled convolutional informer model based on the evaluation index;

[0060] Step 6.4: Predicting the remaining service life of the test set based on the comparison method, and obtaining the remaining service life prediction value and corresponding evaluation index value of the lithium battery for each comparison method;

[0061] Step 6.5: Compare the evaluation index value of the full step-by-step real-time decomposition-tightly coupled convolutional Informer model with the evaluation index values of each comparison method. The smaller the evaluation index value, the higher the accuracy of the model. The models are compared and verified based on the size of the evaluation index value.

[0062] According to another aspect of the present invention, the present invention also provides a device for predicting the remaining service life of a lithium battery for an electric excavator. The device includes a data acquisition module, a data processing module, a data resampling module, a model training module, and a remaining service life prediction module. The acquisition output end of the data acquisition module is electrically connected to the remaining service life prediction module through the data processing module, the data resampling module, and the model training module in sequence.

[0063] The data acquisition module uses a coulomb counter sensor to collect real-time capacity data of the lithium battery and stores the capacity time series in a computer device through serial communication;

[0064] The data processing module is used to decompose the capacity time series into multiple smooth capacity subsequences step by step and in real time;

[0065] The data resampling module is used to resample multiple smooth capacity subsequences to obtain a complete smooth lithium battery capacity time series;

[0066] The model training module is used to train the model by taking the complete flat lithium battery capacity time series as the input of the full step-by-step real-time decomposition-tightly coupled convolutional informer model;

[0067] The remaining service life prediction module is used to evaluate the remaining service life of future electric excavator lithium batteries using the trained full step-by-step real-time decomposition-tightly coupled convolutional Informer model, thereby realizing real-time prediction of the remaining service life of electric excavator lithium batteries.

[0068] The present invention provides a method and device for predicting the remaining useful life of a lithium battery in an electric excavator. The method first obtains the capacity time series data of the lithium battery through the coulomb counter sensor of the electric excavator. A full stepwise real-time decomposition sampling is constructed to decompose the lithium battery capacity time series in real time without information leakage to alleviate the capacity regeneration phenomenon of the lithium battery. The resampled data is divided into a training set and a test set based on the prediction starting point. A deep learning model based on a tightly coupled convolutional informer is constructed to learn the local volatility and global degradation of the capacity time series.

[0069] The present invention trains and optimizes the parameters of a full-step real-time decomposition-tightly coupled convolutional informer model using a training set; inputs a test set into the trained full-step real-time decomposition-tightly coupled convolutional informer model to predict remaining useful life; evaluates the prediction accuracy of the deep learning model using evaluation metrics; and uses the trained network to estimate the test set and verify the model's performance. The method of the present invention can effectively alleviate the capacity regeneration phenomenon of lithium batteries and improve the model's remaining useful life prediction performance. It can achieve end-to-end remaining useful life prediction for lithium batteries in electric excavators, thereby resolving the problem of lag in existing prediction technologies, which leads to the inability to accurately capture the dynamic changes of capacity regeneration and low prediction accuracy.

[0070] In summary, the present invention has the following beneficial technical effects:

[0071] (1) The method of the present invention can effectively alleviate the capacity regeneration phenomenon of lithium batteries and improve the remaining service life prediction performance of the model in a non-information leakage manner, and can realize the end-to-end remaining service life prediction of lithium batteries for electric excavators, thereby solving the problem that the existing prediction technology is lagging and cannot accurately capture the dynamic changes of the capacity regeneration phenomenon and has low prediction accuracy.

[0072] (2) The RTVMD constructed by the present invention cleverly avoids the problems of capacity regeneration and information leakage through the full step-by-step real-time decomposition sampling technology. The established RTVMD-TCCI can effectively capture the local fluctuations and global degradation of the capacity sequence and has a strong robust RUL prediction result.

[0073] (3) The tightly coupled convolutional informer prediction model proposed in this invention tightly couples the various modules of the informer, maximizes the prediction performance, and brings the best prediction performance to the hybrid model.

[0074] (4) The lithium battery remaining service life prediction device designed by the present invention has the characteristics of fast prediction speed, high accuracy of remaining service life prediction, light and efficient model, and strong anti-noise interference ability. It can meet the practical application needs of lithium battery remaining service life prediction and also has great application potential in other application fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0076] Figure 1 The present invention is a flowchart of the steps of a method for predicting the remaining service life of a lithium battery for an electric excavator.

[0077] Figure 2 This is a schematic diagram of the full step-by-step real-time decomposition sampling calculation process proposed by the present invention.

[0078] Figure 3 This is a schematic diagram of the calculation process of the tightly coupled convolution Informer of the present invention.

[0079] Figure 4 A schematic structural diagram of a device for predicting the remaining useful life of a lithium battery for an electric excavator according to the present invention.

[0080] Figure 5 It is the capacity decay time series of four batteries according to a specific embodiment of the present invention.

[0081] Figure 6 It is a remaining service life prediction diagram of different models on different capacity degradation curves according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0082] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0083] The following is an explanation of the English abbreviations used in this invention:

[0084] RTVMD-TCCI: Fully Step-by-Step Real-Time Decomposition-Tightly Coupled Convolutional Informer

[0085] ODVMD-TCCI: Holistic Decomposition Sampling-Tightly Coupled Convolution Informer

[0086] RTDWT-TCCI: Real-time Discrete Wavelet Transform-Tightly Coupled Convolution Informer

[0087] TCCI: Tightly Coupled Convolutional Informer

[0088] SP: Prediction starting point

[0089] EOL: End of Life

[0090] RUL: Remaining useful life

[0091] See also Figure 1 The present invention provides a method for predicting the remaining service life of a lithium battery of an electric excavator, comprising the following steps:

[0092] S1: Using a coulomb counter sensor, we obtain the full-life charge and discharge cycle capacity time series data of the electric excavator lithium battery, as well as the number of cycles at the end of the battery life and the remaining service life.

[0093] S2: performing full step-by-step real-time decomposition sampling on the capacity time series data to alleviate the capacity regeneration phenomenon of the lithium battery and construct a training set and a test set;

[0094] S3: We incorporate cross-stage local attention, extended causal convolution, and a transfer mechanism into the deep learning prediction model Informer to build a deep learning model with tightly coupled convolutional informers.

[0095] S4: Build a hybrid deep learning model of full step-by-step real-time decomposition and tightly coupled convolutional informer to learn local volatility and global degradation of capacity time series;

[0096] S5: Evaluate the prediction accuracy of the full step-by-step real-time decomposition-tightly coupled convolutional informer deep learning model using evaluation metrics;

[0097] S6: Use the trained full step-by-step real-time decomposition-tightly coupled convolutional informer model to estimate the test set, obtain the remaining service life prediction results of the lithium battery, and verify the performance of the model.

[0098] The following is further explained with reference to specific embodiments and execution steps:

[0099] In step S1, the capacity time series data of the electric excavator lithium battery obtained based on the coulomb meter sensor is expressed as S = {S1, S2, S n ,…,S N}, S n is the battery capacity data of the nth charge and discharge cycle. The number of cycles at the end of battery life is defined as the number of charge and discharge cycles when the battery capacity degrades to 80% of the rated capacity. The remaining useful life (RUL) is defined as the remaining cycles available from now to the scheduled end of life (EOL), which can be expressed as:

[0100] RUL=n EOL -n (1);

[0101] where n EOL is the number of cycles at the end of the battery life, and n is the number of cycles at the current battery capacity.

[0102] like Figure 2 As shown, step S2 of performing full step-by-step real-time decomposition sampling on the capacity time series data to alleviate the capacity regeneration phenomenon of the lithium battery and constructing a training set and a test set includes the following sub-steps:

[0103] Step 2.1: Take the pth lithium battery charge and discharge cycle as the prediction starting point, where p < n EOL , the capacity time series data {S1,S2,S n ,…,S N} is divided into training cycles {S1, S2, ..., S p} and test cycle {S p+1 ,S p+2 ,…,S N As a capacity time series, the battery will go through N charge and discharge cycles from the beginning of use to the final failure. Each cycle corresponds to a capacity data measured by the coulomb meter sensor. Then S1 is the capacity data of the first cycle of the battery, S N It is the capacity data of the last cycle of the training period. Assuming there are N cycles in total, p is the number of cycles of the training period. Therefore, the data from 1 to N constitute a battery capacity time series of the entire life cycle.

[0104] Step 2.2: Use a sliding window with a window size of m and a step size of 1 to segment the capacity time series data. The m past observations that lag behind the data to be predicted by one window are used as candidate explanatory variables, and the value to be predicted after m is used as the response variable.

[0105] In this invention, the remaining useful life prediction utilizes the sliding window method used in time series prediction. This method primarily uses the previous m data points to predict the next data point (i.e., the sliding window method). This "sliding window" approach allows multiple training samples to be constructed from a long sequence, facilitating model learning. Essentially, a fixed-length segment of continuous data is taken from the time series as input to predict the next data point. The window gradually slides, generating multiple training samples, allowing the model to learn the time series' changing patterns. Assuming the original time series is: S = [1.0, 1.2, 1.4, 1.3, 1.5, 1.8, 2.0], and the window size m = 3, the input samples can be constructed as follows: input [1.0, 1.2, 1.4] → output prediction result 1.3, input [1.2, 1.4, 1.3] → output prediction result 1.5, and input [1.4, 1.3, 1.5] → output prediction result 1.8. This allows for multiple training samples to be drawn from a single time series, enhancing data utilization and robustness. Therefore, the m past observations here are the m past time series data with a window size of m, and the value to be predicted after m is the value to be predicted in the next time step after m;

[0106] Step 2.3: Segment the first sequence segment sample {S1, S2, ..., S m}∈R 1×m ,m<N is decomposed into k subsequences through variational mode decomposition, and m is the window size;

[0107] The first sequence segment is from 1 to m (m sequence segment data), and the k subsequences are resampled in series into a time series. Then, the explanatory variable of the first sequence segment sample is The response variable is S m+1 ;

[0108] Step 2.4: Use the real-time sliding window method with a step size of 1 to obtain the second sequence segment sample {S2, S3, ..., S m+1}∈R 1×m , m<N, decompose the second sequence fragment sample into k subsequences, and similarly resample the k subsequences in series into a time series. Then the explanatory variable of the second sample is The response variable is S m+2 ;

[0109] Step 2.5: Slide and decompose step by step until the last sequence fragment {SN-m ,S N-m+1 ,…,S N-1} is decomposed and resampled in series, and finally a total of (Nm)∈R (N-m)×(m*k) explanatory variable samples, ending decomposition and resampling;

[0110] Step 2.6: Use the first pm explanatory variable samples as the training set, and the remaining Np explanatory variable samples as the test set; the data obtained by real-time decomposition sampling is the data formed by concatenating several subsequences obtained after the capacity data is decomposed; the m window data is decomposed into k subsequences with a dimension of m×k by the variational mode, and then the concatenation dimension is 1×(m*k). Then, the remaining Nm data are decomposed by real-time sliding window decomposition with a step size of 1, and the final data dimension is (Nm)×(m*k);

[0111] In the use of lithium batteries, the remaining service life prediction of lithium batteries generally selects a part of the data set before 80% of the rated capacity as the training set, and the remaining part as the test set. However, because the number of cycles that each battery can be used is different, the predicted data is divided into training cycle and test cycle. Here, p is used instead of the division node, that is, the capacity data of the first p cycles is used as the training set. Then, after real-time decomposition and sampling, data with a dimension of (Nm)×(m*k) is obtained. In essence, it is a time series prediction, and m is the window. Therefore, the first pm samples are used as the training set, and the remaining Np samples are used as the test set. The dimension of each sample is 1×(m*k).

[0112] like Figure 3 As shown in step S3, the Informer deep learning model includes an encoder and a decoder. Adding cross-stage local attention, extended causal convolution, and transfer mechanisms to the Informer to build a tightly coupled convolutional Informer deep learning model includes the following sub-steps:

[0113] Step 3.1: In the encoder, a certain historical data X∈R of the pm explanatory variable samples is converted into L×d As the input of multi-head sparse cross-stage local attention, where L < (pm) is the input length and d < (m*k) is the input feature dimension;

[0114] Step 3.2: In the multi-head sparse cross-stage local attention, input X∈R L×d Divided into two parts along the feature dimension X1 is used as the input of sparse self-attention, and X2 is connected to the end of the multi-head sparse cross-stage local attention after a 1×1 convolution. The outputs of the two branches X1 and X2 are connected through the feature dimension as the output of the entire multi-head sparse cross-stage local attention block. In deep learning, data is connected through the feature dimension, which can be expressed as:

[0115]

[0116] Where Y∈R L×d is the output of multi-head sparse cross-stage local attention, A(X 1h ) is the scaled dot product of the h-th self-attention block, W h It is d h ×d h The linear projection matrix of , H is the number of multi-head sparse cross-stage local attention heads, d h is the dimension of each head, W c is the weight matrix of the 1×1 convolutional layer;

[0117] Step 3.3: The output of the multi-head sparse cross-stage local attention Y∈R L×d As the input of the dilated causal convolutional layer, for the fth dilated causal convolutional layer after the i-th multi-head sparse cross-stage local self-attention block, in the sequence Y∈R L×d The element y in g ∈R d , g∈L uses the dilated causal convolution operation Z with kernel size u f (y g ) can be defined as:

[0118]

[0119] where Y∈R L×d is the input sequence, y g ∈R d , g∈L is an element in Y, W f ∈R d×d′ is the weight matrix of the f-th extended causal layer, d is the input dimension, d′ is the output dimension, v is the dilation factor, and the filter of the f-th extended causal convolutional layer skips between two adjacent filter taps (2 v-1 -1) elements to effectively improve the receptive field of traditional convolution;

[0120] Step 3.4: Dilate the output of the causal convolution Z∈R L×d As the input of the maximum pooling layer to extract the most significant local pattern in the time series, let p be the pooling window size, s be the pooling step size, z l ′∈R dis the pooling result of the l∈Lth time period, then the maximum pooling layer can be defined as:

[0121]

[0122] The maximum pooling layer selects the maximum value of the jth feature from the consecutive p time steps of the input Z for each feature dimension j to obtain the output sequence Z′∈R L′×d ,in

[0123] Step 3.5: Use three multi-head sparse cross-stage local attentions in the encoder, and use an extended causal convolution layer and a maximum pooling layer to connect every two multi-head sparse cross-stage local attentions. There are three layers of output, namely And usually L′3<L′2<L′1, where Z′1, Z′2, Z′3 are the outputs of the first, second and third multi-head sparse cross-stage local attention respectively, and L′1, L′2, L′3 are the time steps of the first, second and third multi-head sparse cross-stage local attention output respectively;

[0124] In step 3.6, the transfer mechanism is used to concatenate the features of the three scales Z′1, Z′2, and Z′3 to fuse multi-scale information and enhance the capacity and time granularity perception capability, which can be expressed as:

[0125]

[0126] Step 3.7: Fusion of multi-scale features output by the transfer mechanism Input transition layer to compress feature dimensions to unify feature dimensions and ensure the consistency of subsequent decoder input structure, which can be expressed as:

[0127]

[0128] in is the output of the transition layer, ReLU is the activation function, BN is batch normalization, Conv1D 1×1 It is a one-dimensional convolution;

[0129] Step 3.8: Transition layer output A certain period of historical data X of the explanatory variable sample de ∈R T×d The decoder is fed together with the masked multi-head sparse cross-stage local attention to obtain the sequence feature learning results of the decoder;

[0130] Step 3.9: Input the sequence feature learning results of the decoder into the fully connected layer to build the final tightly coupled convolutional Informer deep learning model.

[0131] In step S4, the hybrid deep learning model of full step-by-step real-time decomposition-tightly coupled convolutional informer is established to learn the local volatility and global degradation of the capacity time series, which includes the following sub-steps:

[0132] Step 4.1: Based on the electric excavator battery management system, use a coulomb counter sensor to collect capacity degradation curves under different discharge conditions;

[0133] Step 4.2: Processing the capacity degradation curve based on the full stepwise real-time decomposition sampling to obtain a serially resampled explanatory variable sample; resampling the subsequences ensures that the prediction model is trained only once;

[0134] Step 4.3: Train the weight parameters in the tightly coupled convolutional informer deep learning model based on the training set. When the tightly coupled convolutional informer deep learning model converges to the optimal state, stop training and save the optimal model weight parameters. During training, the mean square error (MSE) is used as the loss function of the model training. The loss function can be expressed as:

[0135]

[0136] Where N is the number of samples, C i is the actual capacity value of the i-th sample, C′ i is the predicted capacity value of the i-th sample.

[0137] Step 4.4: Combining the full step-by-step real-time decomposition sampling and the tightly coupled convolutional informer deep learning model, and obtaining an optimal full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model based on the optimal tightly coupled convolutional informer deep learning model parameters; the core of the full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model is a neural network, which is composed of multiple neurons. Each neuron has some weight parameters. These weight parameters determine the connection strength between neurons, thereby affecting the behavior of the entire network. During the training process, the weight parameters are optimized by backpropagation through the loss function;

[0138] Step 4.5: Predict the test set based on the optimal full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model to obtain a capacity degradation prediction value of the test set.

[0139] In step S5, the prediction accuracy of the deep learning model is evaluated using an evaluation index;

[0140] Specifically, four indicators, namely, absolute error (AE), mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE), are used to evaluate the prediction accuracy of the full step-by-step real-time decomposition-tightly coupled convolution (RTVMD-TCCI model):

[0141] AE=|RUL pr -RUL re | (8);

[0142]

[0143] Where N is the number of samples, C i is the actual capacity value of the i-th battery, C′ i is the predicted capacity value of the i-th battery. pr is the predicted value of RUL, RUL re is the actual value of RUL.

[0144] In step S6, the method of estimating the test set using the trained full step-by-step real-time decomposition-tightly coupled convolutional informer model to obtain the remaining useful life prediction result of the lithium battery and verifying the performance of the full step-by-step real-time decomposition-tightly coupled convolutional informer model includes the following sub-steps:

[0145] S0601: Based on the capacity degradation prediction value of the test set, determine the number of cycles when the full step-by-step real-time decomposition-tightly coupled convolutional informer model predicts 80% of the rated capacity;

[0146] S0602: Based on the predicted number of cycles when the rated capacity reaches 80%, the number of cycles at the prediction starting point is subtracted from the predicted number of cycles when the rated capacity reaches 80%, thereby obtaining a remaining useful life prediction result of the test set using the full step-by-step real-time decomposition-tightly coupled convolutional informer model;

[0147] S0603: Calculating various evaluation index values of the prediction results of the full step-by-step real-time decomposition-tightly coupled convolutional informer model based on the evaluation index;

[0148] S0604: Predicting the remaining service life of the test set based on the comparison method, and obtaining the remaining service life prediction value of the lithium battery and the corresponding evaluation index value for each comparison method;

[0149] S0605: Compare the evaluation index value of the full step-by-step real-time decomposition-tightly coupled convolutional Informer model with the evaluation index values of each comparison method. The smaller the evaluation index value, the higher the accuracy of the model. The models are compared and verified based on the size of the evaluation index value.

[0150] See also Figure 4 , Figure 4 This is a schematic structural diagram of a device for predicting the remaining useful life of a lithium battery for an electric excavator according to the present invention, which specifically includes a data acquisition module, a data processing module, a data resampling module, a model training module, and a remaining useful life prediction module;

[0151] The data acquisition module uses a coulomb counter sensor to collect real-time capacity data of the lithium battery and stores the capacity time series in a computer device through serial communication;

[0152] The data processing module is used to decompose the capacity time series into multiple smooth capacity subsequences step by step and in real time;

[0153] The data resampling module is used to resample multiple flat capacity subsequences to obtain a complete flat lithium battery capacity time series;

[0154] The model training module is used to train the model by taking the complete flat lithium battery capacity time series as the input of the full step-by-step real-time decomposition-tightly coupled convolutional informer model;

[0155] The remaining service life prediction module is used to evaluate the remaining service life of future electric excavator lithium batteries using the trained full step-by-step real-time decomposition-tightly coupled convolutional informer model, thereby realizing real-time prediction of the remaining service life of electric excavator lithium batteries.

[0156] Furthermore, in order to verify the effectiveness of the remaining useful life prediction of the present invention, specifically, comparative methods such as ODD-TCCI, RTVMD-Informer, RTDWT-TCCI, TCCI, and Informer were established to verify the performance of the full step-by-step real-time decomposition-tightly coupled convolution Informer model. In the embodiment, the method of the present invention and the comparative method were used to predict the remaining useful life of the lithium battery of an electric excavator.

[0157] Specifically, in order to ensure the reliability and robustness of the RTVMD-TCCI remaining service life prediction, the operating capacity degradation curves of four lithium batteries of the electric excavator were selected, such as Figure 5 As shown in Figure 2, 80% of the rated capacity is defined as the EOL threshold. Figure 5 It can be seen that each battery undergoes capacity regeneration during the discharge process, which makes the battery capacity time series locally fluctuating and nonlinearly degraded.

[0158] The prediction starting point SP is set to 300, and prediction starts from the 300th cycle, that is, the first 300 cycles are selected for training and the remaining cycles are used for testing. The experimental results of the four batteries are shown in Table 1. It can be seen from Table 1 that the proposed RTVMD-TCCI achieves the best prediction performance on all four batteries, and it obtains the lowest four performance indicators.

[0159] Battery 1 reached the EOL threshold at its 370th cycle, with an error of 1 cycle, thus predicting failure in advance. Battery 2 reached the EOL threshold at its 511th cycle, with an error of 1 cycle. Battery 3 reached the EOL threshold at its 574th cycle, with an error of 1 cycle. Battery 4 reached the EOL threshold at its 487th cycle, with an error of 0 cycle. The RUL predictions for all four battery datasets achieved minimal error.

[0160] On Battery 1, compared with the best comparison model ODD-TCCI, RTVMD-TCCI achieves a MAPE reduction of 18.92%, a RMSE reduction of 24.21%, and a MAE reduction of 20%, all of which demonstrate the effectiveness of the proposed RTVMD-TCCI.

[0161] Table 1 Prediction performance of all methods on the self-test dataset

[0162]

[0163] In order to better analyze the prediction effect of each prediction model, the prediction results of the RTVMD-TCCI model on four batteries are visualized, as shown in the figure. Figure 6 As shown. Figure 6 It can be found that under the same SP starting prediction point, RTVMD-TCCI can well predict the capacity degradation trajectory of the four lithium batteries. In addition, the predicted degradation curve of RTVMD-TCCI almost reaches the EOL threshold at the same time as the original degradation curve, which shows that RTVMD-TCCI has a high-precision RUL prediction performance. Thanks to good data preprocessing technology and the powerful prediction ability of TCCI, RTVMD-TCCI can well capture this local feature and achieve a good fitting effect. Overall, the proposed RTVMD-TCCI method can still achieve high prediction accuracy and prediction stability when processing battery capacity data with unknown frequent disturbances.

[0164] The above disclosure is only a preferred embodiment of the present invention, and certainly cannot be used to limit the scope of the rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for predicting the remaining service life of a lithium battery for an electric excavator, characterized in that: The following steps are involved: Step 1: Obtain the full life cycle charge and discharge cycle capacity time series data of the electric excavator lithium battery based on the coulomb counter sensor, the number of cycles at the end of the battery life, and the remaining service life; The capacity time series data of the lithium battery is represented by S={S1, S2, S n ,…,S N }, S n The battery capacity data for the nth charge and discharge cycle; The remaining service life RUL is expressed as: RUL=n EOL -n(1); n EOL is the number of cycles at the end of the battery life, and n represents the nth charge and discharge cycle; Step 2: Perform full-step real-time decomposition sampling on the capacity time series data, and construct a training set and a test set; Step 3: Add cross-stage local attention, extended causal convolution, and transfer mechanism to the deep learning prediction model Informer to build a deep learning model with tightly coupled convolutional informer; Step 4: Build a hybrid deep learning model of full step-by-step real-time decomposition and tightly coupled convolutional informer to learn the local volatility and global degradation of capacity time series; Step 5: Evaluate the prediction accuracy of the hybrid deep learning model of the full step-by-step real-time decomposition-tightly coupled convolutional informer using evaluation indicators; Step 6: Use the trained hybrid deep learning model of full step-by-step real-time decomposition-tightly coupled convolutional informer to estimate the test set, obtain the remaining service life prediction results of the lithium battery, and verify the performance of the model.

2. The method for predicting the remaining service life of a lithium battery for an electric excavator according to claim 1, wherein: In step 2, the capacity time series data is subjected to full step-by-step real-time decomposition sampling and a training set and a test set are constructed, including the following steps: Step 2.1: Take the pth lithium battery charge and discharge cycle as the prediction starting point, where p < n EOL , the capacity time series data {S1,S2,S n ,…,S N } is divided into training cycles {S1, S2, ..., S p } and test cycle {S p+1 ,S p+2 ,…,S N }; Step 2.2: Use a sliding window with a window size of m and a step size of 1 to segment the capacity time series data. The m past observations that lag behind the data to be predicted by one window are used as candidate explanatory variables, and the value to be predicted after m is used as the response variable. Step 2.3: Decompose the first sequence fragment sample obtained by sliding window segmentation into k subsequences through variational mode, and resample the k subsequences in series into a time series. Then, the explanatory variable of the first sequence fragment sample is The response variable is S m+1 ; Step 2.4: Use a real-time sliding window with a window size of m and a step size of 1, and continue sliding the window along the capacity time series to obtain the second sequence fragment sample. Decompose the second sequence fragment sample into k subsequences, and similarly resample the k subsequences in series into a time series. The explanatory variable of the second sequence fragment sample is The response variable is S m+2 ; Step 2.5: Slide and decompose step by step until the last sequence fragment sample is decomposed and resampled in series, and finally a total of (Nm)∈R is obtained (N-m)×(m*k) explanatory variable samples, ending decomposition and resampling; Step 2.6: Use the first pm explanatory variable samples as the training set and the remaining Np explanatory variable samples as the test set.

3. The method for predicting the remaining service life of a lithium battery for an electric excavator according to claim 1, wherein: In step 3, the Informer deep learning model includes an encoder and a decoder. The process of adding cross-stage local attention, extended causal convolution, and transfer mechanisms to the deep learning prediction model Informer to establish a tightly coupled convolutional Informer deep learning model includes the following steps: Step 3.1: In the encoder, a certain historical data X∈R of pm explanatory variable samples is converted into L×d As the input of multi-head sparse cross-stage local attention, where L < (pm) is the input length and d < (m*k) is the number of input feature dimensions; Step 3.2: In the multi-head sparse cross-stage local attention, input X∈R L×d Divided into two parts along the feature dimension X1 is used as the input of the sparse self-attention, and X2 is connected to the end of the multi-head sparse cross-stage local attention after a 1×1 convolution. The outputs of the two branches X1 and X2 are connected through the feature dimension as the output of the entire multi-head sparse cross-stage local attention block. The output is expressed as: where Y∈R L×d is the output of multi-head sparse cross-stage local attention, A(X 1h ) is the scaled dot product of the h-th self-attention block, W h It is d h ×d h The linear projection matrix of , H is the number of multi-head sparse cross-stage local attention heads, d h is the dimension of each head, W c is the weight matrix of the 1×1 convolutional layer; Step 3.3: The output of the multi-head sparse cross-stage local attention Y∈R L×d As the input of the dilated causal convolutional layer, for the fth dilated causal convolutional layer after the i-th multi-head sparse cross-stage local self-attention block, in the sequence Y∈R L ×d The element y in g ∈R d , g∈L uses the dilated causal convolution operation Z with kernel size u f (y g ) is defined as: where Y∈R L×d is the input sequence, y g ∈R d , g∈L is an element in Y, W f ∈R d×d′ is the weight matrix of the f-th extended causal layer, d is the input dimension, d′ is the output dimension, and v is the expansion factor; Step 3.4: Dilate the output of the causal convolution Z∈R L×d As the input of the maximum pooling layer to extract the most significant local pattern in the time series, let e be the pooling time window size, s be the pooling step size, z l ′∈R d is the pooling result of the l∈Lth time period, then the maximum pooling layer can be defined as: The maximum pooling layer selects the maximum value of the jth feature from the continuous e time windows of the output Z of the extended causal convolution for each feature dimension j to obtain the output sequence Z′∈R L′×d ,in Step 3.5: Use three multi-head sparse cross-stage local attentions in the encoder, and use an extended causal convolution layer and a maximum pooling layer to connect every two multi-head sparse cross-stage local attentions. There are three layers of output, namely And usually L′3<L′2<L′1, where Z′1, Z′2, Z′3 are the outputs of the first, second and third multi-head sparse cross-stage local attention respectively, and L′1, L′2, L′3 are the time steps of the first, second and third multi-head sparse cross-stage local attention output respectively; In step 3.6, a transfer mechanism is used to concatenate the features of the three scales Z′1, Z′2, and Z′3 to fuse multi-scale information and enhance the capacity and time granularity perception capability. The multi-scale fusion feature is expressed as: Step 3.7: Fusion of multi-scale features output by the transfer mechanism The input transition layer is used to compress the feature dimension to unify the feature dimension and ensure the consistency of the input structure of the subsequent decoder. The output of the transition layer is expressed as: in is the output of the transition layer, ReLU is the activation function, BN is batch normalization, Conv1D 1×1 It is a one-dimensional convolution; Step 3.8, the output of the transition layer A certain period of historical data X of the explanatory variable sample de ∈R T×d The decoder is fed together with the masked multi-head sparse cross-stage local attention to obtain the sequence feature learning results of the decoder; In step 3.9, the sequence feature learning results of the decoder are input into the fully connected layer to build the final tightly coupled convolutional Informer deep learning model.

4. The method for predicting the remaining service life of a lithium battery for an electric excavator according to claim 1, wherein: The process of the full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model in step 4 includes the following steps: Step 4.1: Based on the charge and discharge conditions of the electric excavator battery, use a coulomb counter sensor to collect capacity degradation curves under different discharge conditions; Step 4.2: Processing the capacity degradation curve based on the full stepwise real-time decomposition sampling to obtain a series resampled explanatory variable sample; Step 4.3: Train the weight parameters in the tightly coupled convolutional informer deep learning model based on the training set. When the tightly coupled convolutional informer deep learning model converges to the optimal value, stop training and save the optimal model weight parameters. During training, the mean square error is used as the loss function of the model training. The loss function Loss is expressed as: Where N is the number of samples, C i is the actual capacity value of the i-th sample, C i ′ is the predicted capacity value of the i-th sample; Step 4.4: Combine the full step-by-step real-time decomposition sampling and the tightly coupled convolutional informer deep learning model, and add training based on the weight parameters of the best tightly coupled convolutional informer deep learning model to obtain the best full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model; Step 4.5: Predict the test set based on the optimal full step-by-step real-time decomposition-tightly coupled convolutional informer hybrid deep learning model to obtain a capacity degradation prediction value of the test set.

5. The method for predicting the remaining service life of a lithium battery for an electric excavator according to claim 1, wherein: In step 5, the evaluation indicators include absolute error, mean absolute percentage error, root mean square error and mean absolute error. The absolute error, mean absolute percentage error, root mean square error and mean absolute error are used to evaluate the prediction accuracy of the full step-by-step real-time decomposition-tightly coupled convolutional Informer model respectively.

6. The method for predicting the remaining service life of a lithium battery for an electric excavator according to claim 1, wherein: In step 6, the trained full step-by-step real-time decomposition-tightly coupled convolutional informer model is used to estimate the test set and obtain the remaining service life prediction result of the lithium battery. The specific steps include the following: Step 6.1: Based on the capacity degradation prediction value of the test set, determine the number of cycles when the full step-by-step real-time decomposition-tightly coupled convolutional informer model predicts 80% of the rated capacity; Step 6.2: Based on the predicted number of cycles when the rated capacity reaches 80%, the number of cycles at the predicted starting point is subtracted from the predicted number of cycles when the rated capacity reaches 80%, and the remaining useful life prediction result of the test set using the full step-by-step real-time decomposition-tightly coupled convolutional informer model is obtained; Step 6.3: Calculate the evaluation index values of the prediction results of the full step-by-step real-time decomposition-tightly coupled convolutional informer model based on the evaluation index; Step 6.4: Predicting the remaining service life of the test set based on the comparison method, and obtaining the remaining service life prediction value and corresponding evaluation index value of the lithium battery for each comparison method; Step 6.5: Compare the evaluation index value of the full step-by-step real-time decomposition-tightly coupled convolutional Informer model with the evaluation index values of each comparison method. The smaller the evaluation index value, the higher the accuracy of the model. The models are compared and verified based on the size of the evaluation index value.

7. A device for predicting the remaining service life of a lithium battery for an electric excavator, characterized in that: The life prediction device is used to execute the remaining service life prediction method of the electric excavator lithium battery according to claim 1, and the life prediction device includes a data acquisition module, a data processing module, a data resampling module, a model training module and a remaining service life prediction module. The acquisition output end of the data acquisition module is electrically connected to the remaining service life prediction module through the data processing module, the data resampling module, and the model training module in sequence; The data acquisition module uses a coulomb counter sensor to collect real-time capacity data of the lithium battery and stores the capacity time series in a computer device through serial communication; The data processing module is used to decompose the capacity time series into multiple smooth capacity subsequences step by step and in real time; The data resampling module is used to resample multiple smooth capacity subsequences to obtain a complete smooth lithium battery capacity time series; The model training module is used to train the model by taking the complete flat lithium battery capacity time series as the input of the full step-by-step real-time decomposition-tightly coupled convolutional informer model; The remaining service life prediction module is used to evaluate the remaining service life of future electric excavator lithium batteries using the trained full step-by-step real-time decomposition-tightly coupled convolutional Informer model, thereby realizing real-time prediction of the remaining service life of electric excavator lithium batteries.

Images