A method for evaluating the remaining life state of a lithium battery based on an improved TCN

Abstract

The application discloses a lithium battery residual life state evaluation method based on an improved TCN, obtains measured voltage and current data of a lithium battery under a cycle working condition, and obtains a cycle period and an actual capacity sequence after data slicing, ampere-hour integration and averaging treatment, so as to serve as input of an A-TCN-DAE model, battery capacity data Cap cur is predicted, then the subsequent capacity of the battery is recursively predicted until Cap cur ≤Cap EOL , the recursive prediction is ended, the number of residual cycle periods reaching an EOL capacity is counted to obtain the current residual life of the battery. The lithium battery residual life state evaluation method has strong local feature capturing capability and self-adaptive capability of input data noise reduction and reconstruction, and the precision of battery residual service life RUL estimation is obviously improved compared with existing deep learning network technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium batteries, and particularly relates to a lithium battery remaining life state evaluation method based on an improved TCN. BACKGROUND

[0002] The energy supply process of a lithium battery is actually a chemical reaction process of the positive and negative electrodes and internal components in the battery, and inevitable wear occurs with the increase of the use time. The single battery has differences when leaving the factory, and the wear degrees of the single batteries are inconsistent during the charging and discharging process, so that the service life of the single batteries is also different. After the single battery reaches the end of life (EOL) state, according to the short board effect after the series and parallel connection, the normal single battery performance is affected, and then the reliability and safety of the vehicle operation are affected, which is very serious. Therefore, based on the data-driven method, the remaining useful life (RUL) of the battery is predicted, and the battery reaching the EOL state in the battery pack is maintained in time, so that the safe use of the vehicle system can be ensured.

[0003] Research shows that there are many RUL prediction methods. Based on the battery model, the particle filter is used to construct the battery performance degradation data at different time intervals, the obtained battery impedance data at different aging states are fitted, the relationship between the impedance and the aging state is obtained, and then the RUL is predicted by using the impedance characteristics. The accuracy and prediction ability of the model are very reliable under stable conditions, but the model is easily affected by external factors. For the power battery system, the characteristics of the battery itself change due to different working conditions and different environmental temperatures. At this time, the model is difficult to obtain accurate battery mechanism characterization ability under the current actual operation. The SVM model is used to establish the mapping relationship between the capacity and the cycle period of the battery close to the EOL state, which can predict the RUL of the battery close to the EOL state, but the disadvantage of the SVM is that it cannot process the prediction of large data. Considering the change of the battery impedance, the existing battery RUL prediction method does not reasonably and effectively capture the capacity recovery characteristics in the battery attenuation process, only considers the general trend of the attenuation process, and cannot guarantee the accuracy of the local prediction. On the other hand, most of the models have poor robustness, and the environmental noise will greatly affect the prediction results.

[0004] Therefore, it is necessary to improve the prior art. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a lithium battery remaining life state evaluation method based on an improved TCN, which can quickly and accurately estimate the remaining useful life of the lithium battery.

[0006] To address the aforementioned technical problems, this invention provides a method for assessing the remaining lifespan of lithium batteries based on an improved TCN, comprising the following specific steps:

[0007] Step S01: Charge the lithium battery under test using constant current and constant voltage to fully charge it, and then discharge it with constant current of 0.5C and 1C respectively until the battery voltage drops to the threshold of 2.7V; repeat the operation 5 times to obtain the measured voltage and current data of the lithium battery under 5 cycles, and obtain the voltage and current curves under the cycle conditions.

[0008] Step S02: For the voltage data under the cyclic operating condition, the cutoff voltage of 2.7V is used as the cutoff point for data slicing, which is the end of the current cycle and the beginning of the next cycle. Each data slice is numbered in sequence, and one data slice is one cycle.

[0009] Step S03: Apply the ampere-hour integration method to the discharge process in each cycle and perform averaging to obtain the actual capacity; summarize the data according to the mapping relationship between the segment number and the actual capacity to obtain the cycle and actual capacity sequence.

[0010] Step S04: Use the cycle period and actual capacity sequence as input to the A-TCN-DAE model. The A-TCN-DAE model includes a denoising autoencoder DAE network and an A-TCN model. The reconstructed and optimized cycle period and actual capacity sequence output by the denoising autoencoder DAE network are used by the A-TCN model to predict the battery capacity data Cap at the current moment. cur Then, the subsequent battery capacity is recursively predicted. Once the predicted capacity value Cap is obtained... cur Preset battery EOL corresponding capacity value Cap EOL Eligible: Cap cur ≤Cap EOL The recursive prediction process ends, and then the number of remaining cycle cycles before reaching the end-of-life (EOL) capacity is counted to obtain the current remaining battery life.

[0011] As an improvement to the lithium battery remaining life state assessment method based on improved TCN of the present invention:

[0012] The A-TCN model is based on the TCN model, and the following improvements are made to each residual module RB: 1) causal convolution is replaced with hybrid dilated convolution, 2) the activation function ReLU is replaced with A-ReLU, and 3) a 1×1 convolution structure is added between the input and output of the residual module RB; each residual module RB calculates its own residual and then passes it layer by layer backward.

[0013] As a further improvement to the lithium battery remaining life state assessment method based on improved TCN of the present invention:

[0014] The noise-reducing autoencoder DAE includes an input sample after mixing noise v in the input x Input sample The data features are extracted by the encoding network E(·), and the decoding network D(·) is operated and the optimized output is reconstructed, and the cycle period and the actual capacity sequence y are optimized and reconstructed:

[0015]

[0016] Wherein, is the reconstructed optimized output, is the bias term of the decoding network;

[0017]

[0018] Wherein, the output f(i) is the sample data high-dimensional features after the encoding network, W is the weight matrix of the encoding network, b i is the bias term corresponding to the encoding network.

[0019] As a further improvement of the improved TCN-based lithium battery remaining life state evaluation method of the application:

[0020] The process of the recursive prediction in step S04 is: define the feature data set of the input and output as:

[0021] CR={(X1,Cap2),(x2,Cap3),…,(X L ,Cap L+1 )} (Formula 4)

[0022] Wherein, X L represents the model input at the Lth moment, Cap L+1 represents the capacity output at the L+1th moment;

[0023] The predicted capacity Cap L+1 at the Lth moment is added to the input capacity sequence X L+1 at the next moment, and the subsequent capacity of the prediction model is continued; the new feature data set obtained by recursion is:

[0024] CR rec ={(X L+1 ,Cap L+1 ),…,(X cur-1 ,Cap cur )} (Formula 5)

[0025] Once the predicted capacity value Cap cur ≤Cap EOL, stop the prediction process; count the number of remaining cycles that reach the EOL capacity during the entire prediction process to obtain the remaining service life of the battery;

[0026] Cap EOL = Cap rated ·70% (Formula 3)

[0027] Cap rated is the rated capacity of the battery.

[0028] As a further improvement of the improved TCN-based lithium battery remaining life state evaluation method of the application:

[0029] The expansion coefficient of the mixed expansion convolution is designed as [1, 2, 4, 8, 16, 32, 64];

[0030] The activation function A-ReLU is:

[0031] f(x) = max{ax, x}, (0 < a < 1) (Formula 8)

[0032] Wherein, a is an input correction coefficient, and x is the input layer feature value of the model.

[0033] As a further improvement of the improved TCN-based lithium battery remaining life state evaluation method of the application:

[0034] The training and testing process of the A-TCN-DAE model is:

[0035] Randomly initialize the weight parameters and bias parameters of the network model, and the training parameters include learning rate, dropout rate, convolution kernel size and expansion factor; input the training set into the A-TCN-DAE model; input a mini-batch of the input sequence, forward propagate according to the input data, calculate the parameters of the hidden layer and the output of the output layer; calculate the loss function, and perform back propagation with the goal of minimizing the loss function, calculate the hidden layer error and find the partial derivative, update the weight parameters and bias parameters in the network in the form of gradient descent, continuously iterate to reduce the error between the predicted value and the true value, until the preset iteration number is reached, save the model parameters to obtain the trained A-TCN-DAE model;

[0036] Then recursively input the test set into the trained A-TCN-DAE model, predict the subsequent capacity according to the recursion, and once the predicted capacity value Cap cur corresponding to the preset battery EOL is obtained, the recursion process is ended; the average absolute error MAE, the root mean square error RMSE, and the R EOL satisfy the condition: Cap cur ≤ Cap EOL , end the recursion process; the average absolute error MAE, the root mean square error RMSE, and the R2 To measure the index on the test set statistics and verify the accuracy of the model prediction, so as to obtain the A-TCN-DAE model available online.

[0037] As a further improvement of the improved TCN-based lithium battery remaining life state evaluation method of the application:

[0038] The acquisition process of the training set and the test set is:

[0039] 1) Four groups of the same lithium batteries are charged to the full charge state by using the constant current constant voltage CCCV method, then the No. 1 lithium battery and the No. 2 lithium battery are discharged at a constant current of 0.5C, the No. 3 and No. 4 lithium batteries are discharged at a constant current of 1C, until the battery voltage respectively drops to the minimum threshold voltage 2.7V as a cycle period;

[0040] 2) Repeat the above operation to obtain the measured voltage and current data of the lithium battery under the cycle condition until the current capacity of the lithium battery decays to the EOL state, and obtain the voltage and current curve under the cycle condition;

[0041] 3) The measured voltage and current data of the lithium battery under the cycle condition are supplemented with missing points or abnormal points by using the linear interpolation method:

[0042] V t = (V t-1 + V t+1 ) / 2 (Formula 1)

[0043] Wherein V t represents the current missing point or abnormal point, V t-1 represents the data of the previous time point, and V t+1 represents the data of the next time point;

[0044] 4) All voltage data is divided into pieces with the cutoff voltage 2.7V as the piece cutoff point, and each cycle is labeled in sequence:

[0045] 5) The ampere-hour integral of the discharge process of each battery from full to empty in the cycle period is obtained to obtain the capacity of the battery in the current cycle period;

[0046] Cap t = Cap t-1 + ∫I(t)dt (Formula 2)

[0047] Wherein Cap t-1 represents the capacity of the battery at t-1, and Cap t represents the capacity of the battery at t;

[0048] 6), the capacity of the battery in a cycle period is averaged according to a sampling time interval to obtain the actual capacity in each cycle period;

[0049] 7), the cycle period and the actual capacity sequence are obtained by summarizing the mapping relationship between the slice label and the actual capacity of each cycle period, and then the cycle period and the actual capacity sequence data of the No. 2 and No. 4 lithium batteries are taken as a training set, and the cycle period and the actual capacity sequence data of the No. 1 and No. 3 lithium batteries are taken as a test set.

[0050] The beneficial effects of the present application mainly include:

[0051] 1, the present application adopts mixed dilated convolution on the basis of time convolution network (TCN) to effectively expand the receptive field of the convolution process, and enhances the model's ability to capture local recovery features of the capacity sequence;

[0052] 2, the present application adopts A-ReLU activation function on the basis of time convolution network (TCN) to enhance certain feature capture ability, improve prediction accuracy and model fitting degree;

[0053] 3, the A-TCN model of the present application has the best real value following ability and the error and fluctuation are effectively reduced after the A-TCN-DAE model combined with the denoising autoencoder DAE, the reconstruction process of the original capacity sequence by DAE is equivalent to further extracting data features and removing noise effects, so that the reconstructed data has more essential features of data and robustness and generalization ability, and the accuracy and robustness of the battery remaining useful life RUL prediction are improved;

[0054] 4, compared with the prior art, the model proposed by the present application has strong local feature capture ability and adaptive ability of input data denoising reconstruction, and the accuracy of the battery remaining useful life RUL estimation is obviously improved compared with the existing deep learning network technology. BRIEF DESCRIPTION OF DRAWINGS

[0055] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0056] Figure 1 is a flowchart of a lithium battery remaining life state evaluation method based on an improved TCN of the present application;

[0057] Figure 2 is a schematic diagram of a data slicing process;

[0058] Figure 3 is a schematic diagram of the overall architecture of the basic TCN model;

[0059] Figure 4A schematic diagram of a denoising autoencoder network structure and a denoising process;

[0060] Figure 5 A schematic diagram of a residual module after introducing a mixed dilated convolution, A-ReLU and residual connection;

[0061] Figure 6 A schematic diagram of the A-TCN-DAE model of the application. DETAILED DESCRIPTION

[0062] The application will be further described below in conjunction with specific embodiments, but the scope of protection of the application is not limited thereto:

[0063] Embodiment 1, a lithium battery remaining life state evaluation method based on an improved TCN, as shown in the figure, including the following specific processes: Figures 1-6

[0064] Step 1, obtain the cycle period and actual capacity sequence by cyclic constant current discharge

[0065] Step 1.1, use the lithium battery charging and discharging platform to perform offline charging and discharging on the lithium battery, and use the battery management system device to collect the measured voltage and current of the lithium battery under the cycle condition.

[0066] 1), select four groups of the same lithium batteries as test batteries, and use the lithium battery charging and discharging platform to charge the four groups of test batteries in a constant current and constant voltage (CCCV) manner to reach the full charge state. The No. 1 lithium battery and the No. 2 lithium battery are discharged at a constant current of 0.5C, the No. 3 and No. 4 lithium batteries are discharged at a constant current of 1C, until the battery voltage respectively drops to the minimum threshold voltage 2.7V, which is regarded as a cycle period;

[0067] 2), repeat the experiment according to step 1) to obtain the measured voltage and current data of the lithium battery under the cycle condition, and obtain the voltage and current curve under the cycle condition; according to the IOS stipulated standard for lithium battery cycle life test, the lithium battery whose current capacity decays to 70% of the rated capacity is regarded as having reached the EOL state, and the cycle condition experiment of the single lithium battery in the charging and discharging experiment platform is also judged according to the set EOL standard.

[0068] Step 1.2, raw data preprocessing

[0069] 1), during the sampling and storage process, the fluctuation of data interaction between the collection hardware device and the communication module will inevitably bring dirty data. The dirty data in the test data mainly includes data missing problem and abnormal jump problem at some sampling points, and the processing method adopts the following steps: first, remove the abnormal points, and then use linear interpolation method to supplement the missing points or abnormal points caused by the sampling process: ​

[0070] V t =(V t-1 +V t+1 ) / 2 (Equation 1)

[0071] Where V t V represents the current missing point or outlier. t-1 V represents the data at the previous time point. t+1 This represents the data at the next point in time.

[0072] 2) Data sharding based on the number of cycle operations

[0073] Based on the characteristics of the cyclic operating condition cutoff voltage, the cutoff voltage point of each cyclic operating condition can be used as the end of the current cycle and the beginning of the next cycle. For example... Figure 2 As shown, all voltage data are segmented with a cutoff voltage of 2.7V as the segmentation cutoff point, and each data segment is numbered sequentially: each data segment is a cycle, and each cycle has a segmentation number.

[0074] 3) Perform ampere-hour integration on the discharge process of the battery in each cycle from full charge to full discharge to obtain the battery capacity in the current cycle.

[0075] Cap t =Cap t-1 +∫I(t)dt (Equation 2)

[0076] Cap t-1 Cap represents the capacity of a specific battery cell at time t-1. t This represents the capacity of a specific battery cell at time t;

[0077] 4) The battery capacity obtained from Equation 2 for each cycle may contain some dirty data. Therefore, the battery capacity for each cycle is averaged according to the sampling time interval to obtain the actual capacity for each cycle.

[0078] 5) By summarizing the mapping relationship between the segment number and the actual capacity of each cycle, a sequence of cycle periods and actual capacity can be obtained, which can be used as input to the A-TCN-DAE model.

[0079] Step 2: Recursive Prediction Process and Stopping Conditions

[0080] When the actual capacity of a lithium battery decreases to 70% of its rated capacity, the battery reaches the End-of-Life (EOL) state. The EOL threshold is set to ensure that when the battery capacity degrades to a certain level, it is considered to have entered an unreliable state, at which point maintenance of the battery system is required. The battery capacity at EOL is Cap. EOL It can be represented as:

[0081] Cap EOL =Cap rated 70% (Formula 3)

[0082] Cap rated This refers to the battery's rated capacity.

[0083] To design a reasonable prediction model that is more suitable for predicting remaining battery life (RUL), the input and output feature datasets are defined as follows:

[0084] CR={(X1,Cap2),(X2,Cap3),…,(X L Cap L+1 (Equation 4)

[0085] Among them, X L Cap represents the model input at time L. L+1 This represents the capacity output at time L+1. According to the causal theorem of time series data, the output of the model at the current time is only related to the data at the current time and before, and is independent of future data.

[0086] The capacity Cap predicted at time L. L+1 Add to the input capacity sequence X at the next time step L+1 Continue predicting the model's capacity for the future. The new feature dataset obtained through recursion is:

[0087] CR rec ={(X L+1 Cap L+1 ),…,(X cur-1 Cap cur (Equation 5)

[0088] Once the predicted capacity value Cap is obtained cur ≤Cap EOL The prediction process can then be stopped. By counting the number of remaining cycles before reaching the end-of-life (EOL) capacity, the remaining battery life (RUL) can be obtained.

[0089] Step 3: Construct the A-TCN-DAE model

[0090] The A-TCN-DAE model comprises a denoising auto-encoding (DAE) network and an A-TCN model (an improved TCN model), and the reconstruction optimized cycle period output by the denoising auto-encoding (DAE) network and the actual capacity sequence are subjected to the A-TCN model to obtain the prediction result of the remaining useful life RUL of the battery. In view of the problem that the commonly used sequence prediction model has poor capturing ability for local capacity recovery characteristics, the TCN model is taken as a basis, and the following improvements are made to each residual block RB (Residual Block) inside the TCN model to construct the A-TCN model: 1) the causal convolution is replaced with a mixed dilated convolution, 2) the activation function ReLU is replaced with A-ReLU, and 3) a 1x1 convolution structure is added between the input and output of the residual block RB; the A-TCN model is combined with the extended feature extraction capability of the denoising auto-encoding network (DAE) to improve the performance and robustness of the model.

[0091] Step 3.1: Construction of a basic TCN model

[0092] The overall architecture of the TCN model is shown in Figure 3 : a full connection layer with a sigmoid activation function is circumscribed by 7 residual blocks RB (Residual Block). The dilation factor of the residual block is subject to a 2 n increasing rule, which is used to expand the receptive field of the model. For example, under the dilated factor of 2 6 , if the convolution kernel kernel is set to 2, the model can perceive a visual field range of 128, and for the capacity and cycle period data, it indicates that the influence factor of the current output contains the capacity information of the last 128 inputs. The residual block contains two layers of causal convolution, and ReLU activation function and dropout layer with a dropout rate of 0.2 are used after each layer of causal convolution. Finally, the full connection layer is subjected to gradient descent learning, and the final output of the next time capacity prediction result is output, that is, the output of the last full connection layer can be expressed as:

[0093] Cap k = σ(w k h k + b k ) (Formula 6)

[0094] wherein σ represents a sigmoid activation function, w k and b k are weight matrix and bias, respectively, and h k is the hidden layer tensor result of the last residual layer.

[0095] Step 3.2: Convolution replacement

[0096] The causal convolution of the residual module RB in the TCN model established in step 3.1 is replaced using mixed dilated convolution. In the convolution operation in the feature extraction process, the receptive field is used to describe the number of data units in the bottom layer network associated with the output decision of a single unit in the top layer network. The larger the receptive field, the fewer the number of network layers required to extract all features of the same size of data features, which means that the network parameters are reduced, thereby improving the processing speed of the network and being able to contain more historical data features. Dilated convolution is used to increase the receptive field, and its characteristics are as follows:

[0097] 1) All convolution dilation coefficients cannot have a common divisor other than 1, for example, [2, 4, 6] is an unreasonable dilation coefficient, which will cause grid effect;

[0098] 2) The dilation coefficient is designed to be jagged, such as: d = [1, 2, 4], which is used to extract feature information at different intervals;

[0099] 3) The maximum interval between non-zero feature values in the convolution kernel needs to meet:

[0100] M i ≤ max{M i+1 - 2r i , M i+1 - 2(M i+1 - r i ), r i} (Equation 7)

[0101] Where r i represents the dilation coefficient of the i-th layer, M i represents the maximum available dilation coefficient at the i-th layer, and if there are n layers of structure, by default r n = M n , combined with the capacity sequence characteristics, in order to capture the local capacity recovery characteristics, all historical capacity information needs to be associated, if the selected convolution kernel kernel size is k x 1, the target is M k ≤ k, so that at least one layer of standard convolution with a dilation coefficient of 1 can be used to ensure that no historical capacity features are missed, and the dilation coefficient is designed as [1, 2, 4, 8, 16, 32, 64];

[0102] Step 3.3, A-ReLU replaces original ReLU

[0103] ReLU activation function characteristics make negative output values are filtered, only consider the output of the neuron features are positive. This feature reduces the amount of calculation, so as to accelerate the training and convergence of the model. However, the filtered neurons may cause some data features to be lost. For the capacity decay feature, although the overall trend is decreasing, due to the existence of capacity recovery characteristics, negative values may appear at local feature points during training, and the fitting ability near 0 needs to be enhanced, so the parameterized activation function A-ReLU is introduced to replace the activation function ReLU of the residual module RB in the TCN model established in step 3.1, and its characteristics are defined as:

[0104] f(x)=max{ax,x},(0<a<1) (Formula 8)

[0105] Where a is the input correction coefficient, and x is the input layer feature value of the model.

[0106] The parameter a is added to the training process as a learnable variable in the network, which increases the model calculation a little, but in return, the sensitivity of the model to the value near 0 is improved, solving the problem of feature loss when the input is less than 0.

[0107] Step 3.4, introduce residual connection:

[0108] By introducing residual connection, the upper layer convolution can contain detailed local feature information while not losing the original input data feature information.

[0109] In order to make an L+n layer network model at least have the ability of an L layer network without causing the network to be deeper and the effect to be worse due to network degradation, etc., a 1x1 convolution structure is added between the input and output of each residual module RB in each residual module RB. After each residual module RB calculates its own residual, it is transmitted back layer by layer, so that the residual modules RB form a residual connection, and all residual information is preserved during the training of the network, as shown in Figure 5 . Assuming x is the input of the model, F(x) is the process of linear conversion and activation function processing, then before the activation of the second layer linear processing, F(x) is combined with the bottom layer input x, and the final output is represented as:

[0110] o=F2(x+F1(x)) (Formula 9)

[0111] After introducing the residual connection, if the features extracted by the hierarchical discrete correlation HDC can improve the generalization ability, i.e. more detailed local features can be obtained based on the input data features, the parameters will be updated normally, and if no effective feature information is learned, the network will not degrade.

[0112] Step 3.5, introduce denoising autoencoder DAE

[0113] The noise reduction process of the denoising autoencoder DAE includes mixing noise v in the input x. The way of introducing noise v refers to the Dropout operation in deep learning, which randomly sets some data in a certain input sequence to 0. Then the input sample containing noise is used The data features are extracted through the encoding network E(·), and finally the original data structure is recovered and reconstructed through the decoding network D(·) decoding operation. The output y is the reconstructed and optimized cycle period and actual capacity sequence, as shown in Figure 4 .

[0114] For the model input X L at the L-th moment defined in step 2, the noise v is introduced in the denoising autoencoder DAE process to obtain The processing flow of the encoding network can be represented as:

[0115]

[0116] Wherein, the output f(i) is the high-dimensional features of the sample data after the encoding network, W is the weight matrix of the encoding network, b i is the bias term corresponding to the encoding network.

[0117] The decoding process can be represented as:

[0118]

[0119] Wherein, y and is the reconstructed and optimized output, is the bias term of the decoding network. The reconstructed and optimized y is used as the input of the subsequent A-TCN.

[0120] Step 4, train the A-TCN-DAE model to predict the remaining service life of the battery

[0121] Step 4.1, construct the training set and test set of the A-TCN-DAE model

[0122] According to the method of step 1, four groups of lithium batteries are tested respectively to obtain the relationship data of the cycle period and the actual capacity sequence of the four groups. Then, the four groups of data are grouped according to the discharge conditions. The cycle period and actual capacity sequence data of the No. 2 and No. 4 lithium batteries are used as the training set, and the cycle period and actual capacity sequence data of the No. 1 and No. 3 lithium batteries are used as the test set. The input of the A-TCN-DAE model is the cycle period and actual capacity sequence, and the output is the predicted capacity.

[0123] Step 4.2, construct the loss function

[0124] Define the predicted value y i and the true value The error between the predicted value and the actual value is the loss function:

[0125]

[0126] Step 4.3, training and testing the A-TCN-DAE model

[0127] Step 4.3.1, training process of the A-TCN-DAE model

[0128] (1) Training parameter setting: the expansion factor d of the dilated convolution is set to [1, 2, 4, 8, 16, 32, 64], the convolution kernel size is 3x1, the dropout rate of the deep learning network is 0.2, and the Adam optimizer is used for optimization during the training process, with dynamic adjustment of the learning rate, the initial learning rate is set to 0.005, and the first-order momentum parameter and the second-order momentum parameter are set to 0.5 and 0.9 respectively. The number of training iterations is 80 times.

[0129] (2) Randomly initialize the weight and bias parameters of the network model: the initial weight is a random number following a Gaussian distribution with a mean of 0 and a standard deviation of 0.01, and the initial bias is all set to 0;

[0130] (3) Input the cycle period and actual capacity sequence in the training set into the A-TCN-DAE model; input a mini-batch of the input sequence, forward propagate according to the input data, calculate the parameters of the hidden layer and the output of the output layer; calculate the loss function, and perform back propagation with the goal of minimizing the loss function (equation 12), calculate the hidden layer error and find the partial derivative, update the weight parameters and bias parameters in the network in the form of gradient descent, continuously iterate to reduce the error between the predicted value and the actual value, until the preset number of iterations is reached, save the model parameters to obtain the trained A-TCN-DAE model;

[0131] Step 4.3.2, testing process of the A-TCN-DAE model:

[0132] Recursively input the current cycle period and actual capacity sequence in the test set into the trained A-TCN-DAE model, recursively predict the subsequent capacity according to equation 5 in step 2, and once the predicted capacity value Cap cur corresponding to the preset battery EOL meets the condition: Cap EOL ≤ Cap cur , the prediction is successful, otherwise, the prediction fails. EOLIf the result is reached, the recursive process ends. Based on the recursive result, the capacity decay curve is obtained, and the number of remaining cycles before reaching the EOL capacity is counted as the remaining battery life (RUL). The accuracy of the model prediction is statistically verified on the test set using MAE, RMSE, and R2 as metrics, thus obtaining the A-TCN-DAE model that can be used online.

[0133] Step 5: Online application process of the A-TCN-DAE model

[0134] 1) The lithium battery under test is charged using a lithium battery charge and discharge platform using a constant current and constant voltage charging method until it is fully charged. Then it is discharged using a constant current of 0.5C and 1C respectively until the battery voltage drops to the threshold of 2.7V. Repeat the operation 5 times to obtain the measured voltage and current data of the lithium battery under 5 cycles, and obtain the voltage and current curves under the cycle conditions.

[0135] 2) The measured voltage and current data of lithium battery under cyclic conditions are segmented. The method is to use the cutoff voltage of 2.7V as the segment cutoff point for all voltage data as the end of the current cycle and the beginning of the next cycle, and to segment and label each cycle in sequence.

[0136] 3) The discharge process in each cycle is processed by ampere-hour integration and averaged to obtain the actual capacity; the cycle and actual capacity sequence is obtained by summarizing the mapping relationship between the segment number and the actual capacity.

[0137] 4) Use the cycle period and actual capacity sequence obtained in step 3) as input to the A-TCN-DAE model to predict the battery capacity data Cap at the current moment. cur Then, predict the subsequent battery capacity using Equation 5 in step 2. Once the predicted capacity value Cap is obtained... cur Cap corresponding to the preset battery EOL EOL Eligible: Cap cur ≤Cap EOL If the recursive process ends, then the number of remaining cycle cycles before reaching the EOL capacity is counted to obtain the current remaining battery life.

[0138] Experiment 1:

[0139] The experimental hardware configuration was as follows: processor 12*E5-2609v3@1.9GHz; memory 16GB; graphics card model Intel Tesla P4 8GB (384.81); operating system CentOS 7.3.1611; model building and training used Python 3.7 + Tensorflow 1.4.

[0140] The experimental data set uses the training set and test set constructed in step 4.1 in embodiment 1, and four comparative models are used for comparison experiments with the A-TCN-DAE model of the application to verify the effectiveness of the A-TCN-DAE model of the application for battery remaining useful life RUL prediction. The four comparative networks are respectively:

[0141] 1) Before replacing the dilated convolution: the basic TCN network constructed in step 3.1 in embodiment 1.

[0142] 2) Before improving the activation function: replace the causal convolution of the TCN network with a hybrid dilated convolution according to step 3.2 in embodiment 1.

[0143] 3) Without using residual connection: replace the causal convolution of the TCN network with a hybrid dilated convolution according to steps 3.2 and 3.3 in embodiment 1, and replace the activation function ReLU with A-ReLU.

[0144] 4) A-TCN: replace the causal convolution of the TCN network with a hybrid dilated convolution according to steps 3.2, 3.3 and 3.4 in embodiment 1, replace the activation function ReLU with A-ReLU, and then introduce a residual connection.

[0145] The experiments use the mean absolute error MAE, the mean square error MSE, the root mean square error RMSE and the R 2 As a measurement index, the specific definitions are as follows:

[0146] Mean absolute error MAE: This index is the expected value of the absolute error loss, which is the average value of the absolute value of the difference between the predicted value and the true value of the sample after summation, which can effectively avoid the positive and negative offset of the error. The average absolute error of N samples can be represented by the following formula:

[0147]

[0148] is the true value of the sample, y i is the predicted value of the model.

[0149] Mean square error MSE: This index is the expectation of the square error, which is inconsistent with the dimension of the target variable, and is the average value of the square of the difference between the predicted value and the true value of the sample after summation. The mean square error of N samples can be represented by the following formula:

[0150]

[0151] is the true value of the sample, y i is the predicted value of the model.

[0152] Root Mean Square Error (RMSE): is the square root of the mean squared error of the model prediction value and the true value. It is very sensitive to the display of very large or very small errors in a group of error data, and can well describe the precision of the model evaluation result. The root mean square error of N samples can be expressed as follows:

[0153]

[0154] R-square: also known as the coefficient of determination or goodness of fit, which reflects the fitting degree between the predicted value and the true value. The closer to 1, the better the model fitting. The R-square of N samples can be expressed as follows:

[0155]

[0156] is the true value of the sample, y i is the predicted value of the model, y0 is the mean value of the true sample.

[0157] Using the training set constructed in step 4.1, five network models are trained to obtain trained network models, and then the test set constructed in step 4.1 is used for testing and the mean absolute error MAE, RMSE, and R 2 As shown in Table 1

[0158] Table 1, comparison results of various measurement indexes of experiment 1

[0159] Model MSE MAE RMSE [R 2 ]]> Before dilated convolution replacement 0.015443 0.092317 0.124270 0.674913 Before activation function improvement 0.009657 0.064217 0.098272 0.796704 Without residual connection 0.008588 0.057751 0.092670 0.819220 A-TCN 0.002470 0.033291 0.049695 0.9480133 A-TCN-DAE 0.000265 0.010763 0.016292 0.9944123

[0160] According to the results of experiment 1, it is shown that the introduction of dilated convolution effectively expands the receptive field of the convolution process, and enhances the model's ability to capture local recovery features of the capacity sequence; the improved activation function enhances the feature capture ability to a certain extent, and improves the prediction accuracy and the goodness of fit of the model; the introduction of residual connection can further improve the network prediction accuracy, but the optimized A-TCN still has a large fluctuation between the true value and the model, and the robustness problem still exists. The A-TCN-DAE model combined with the denoising autoencoder DAE has the best true value following ability and the error and fluctuation are effectively reduced. The reason is that the DAE reconstruction process of the original capacity sequence is equivalent to further extracting the data features and removing the noise effect, so that the reconstructed data has more essential features of the data and robustness and generalization ability.

[0161] Experiment 2:

[0162] Select representative network models for battery remaining life prediction, which specifically include recurrent neural network RNN [1] , long short-term memory network LSTM [2]Gated recurrent unit network, GRU [3] In order to evaluate the accuracy and robustness of the model, the same training set constructed in step 4.1 is used to train the four network models to obtain the trained network models, and then the test set constructed in step 4.1 is used for testing and the mean absolute error MSE, MAE, RMSE, and R 2 As shown in the following table 2.

[0163] Table 2, comparison results of various evaluation indexes of experiment 2

[0164] Model MSE MAE RMSE [R 2 ]]> RNN 0.063097 0.161502 0.179966 0.326071 LSTM 0.032388 0.114287 0.118471 0.654070 GRU 0.005333 0.044645 0.073028 0.850089 A-TCN-DAE 0.000388 0.010185 0.019706 0.995852

[0165] Through the quantitative analysis of the evaluation indexes, it can be seen that the A-TCN-DAE model has advantages in various evaluation indexes.

[0166] References:

[0167] [1] Yu W, Kim I Y, Mechefske C. An improved similarity-based prognosticalgorithm for RUL estimation using an RNN autoencoder scheme[J]. ReliabilityEngineering System Safety, 2020, 199: 106926.

[0168] [2] Sayah M, Guebli D, Noureddine Z, et al. Deep LSTM Enhancement for RULPrediction Using Gaussian Mixture Models[J]. Automatic Control and ComputerSciences, 2021, 55(1): 15-25.

[0169] [3] Lu Y W, Hsu C Y, Huang K C. An Autoencoder Gated Recurrent Unit forRemaining Useful Life Prediction. 2020.

[0170] Finally, it should be noted that the above only lists several specific embodiments of the present application. Obviously, the present application is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or inferred from the disclosed content by those skilled in the art should be considered as falling within the scope of the present application.

Claims

1. A method for assessing the remaining life status of lithium batteries based on improved TCN, characterized in that: The specific process includes: Step S01: Charge the lithium battery under test using constant current and constant voltage to fully charge it, and then discharge it with constant current of 0.5C and 1C respectively until the battery voltage drops to the threshold of 2.7V; repeat the operation 5 times to obtain the measured voltage and current data of the lithium battery under 5 cycles, and obtain the voltage and current curves under the cycle conditions. Step S02: For the voltage data under the cyclic operating condition, the cutoff voltage of 2.7V is used as the cutoff point for data slicing, which is the end of the current cycle and the beginning of the next cycle. Each data slice is numbered in sequence, and one data slice is one cycle. Step S03: Apply the ampere-hour integral method to the discharge process in each cycle and perform averaging to obtain the actual capacity; The cycle period and actual capacity sequence are obtained by summarizing the mapping relationship between the segment number and the actual capacity. Step S04: Use the cycle period and actual capacity sequence as input to the A-TCN-DAE model. The A-TCN-DAE model includes a denoising autoencoder DAE network and an A-TCN model. The reconstructed and optimized cycle period and actual capacity sequence output by the denoising autoencoder DAE network are used by the A-TCN model to predict the battery capacity data at the current moment. Then, the subsequent capacity of the battery is predicted recursively. Once the predicted capacity value is obtained... Capacity value corresponding to the preset battery EOL Eligible conditions: The recursive prediction process ends, and then the number of remaining cycle cycles to reach the EOL capacity is counted to obtain the current remaining battery life. The A-TCN model is based on the TCN model, and the following improvements are made to each residual module RB: 1) causal convolution is replaced with hybrid dilated convolution; 2) the activation function is... Replace with 3) Add a 1×1 convolution structure between the input and output of the residual module RB; each residual module RB calculates its own residual and then passes it backward layer by layer.

2. The method for assessing the remaining life status of a lithium battery based on an improved TCN according to claim 1, characterized in that: Noise reduction autoencoder (DAE) includes input Mixed noise The following is the input sample. Input Sample After coding network Extract data features, and then process them through a decoding network. Decoding operation and reconstruction optimization output: Reconstructed and optimized loop period and actual capacity sequence y: (Equation 11) in, To reconstruct and optimize the output, This is the bias term for the decoding network; (Equation 10) Among them, the output For high-dimensional sample data after encoding network , The weight matrix of the encoding network, This refers to the bias term corresponding to the encoding network.

3. The method for assessing the remaining life status of a lithium battery based on an improved TCN according to claim 2, characterized in that: The recursive prediction process described in step S04 is as follows: Define the input and output feature datasets as follows: (Equation 4) in, This represents the model input at time L. Indicates the first Capacity output at any given time; The first Capacity predicted at any time Add to the input capacity sequence at the next time step Continue predicting the capacity of the model; the new feature dataset obtained recursively is: (Equation 5) Once the predicted capacity value is obtained Stop the prediction process; count the number of remaining cycles that will reach the end-of-life (EOL) capacity during the entire prediction process to obtain the remaining battery life. (Equation 3) in, This refers to the battery's rated capacity.

4. The method for assessing the remaining life status of a lithium battery based on an improved TCN according to claim 3, characterized in that: The dilation coefficient of the hybrid dilated convolution is designed to be [1,2,4,8,16,32,64]; The activation function for: (Equation 8) Where a is the input correction coefficient and x is the input layer feature value of the model.

5. The method for assessing the remaining life status of a lithium battery based on an improved TCN according to claim 4, characterized in that: The training and testing process of the A-TCN-DAE model is as follows: Randomly initialize the weight and bias parameters of the network model. Training parameters include the learning rate, ... The training set is input into the A-TCN-DAE model. A mini-batch of the input sequence is used. The hidden layer parameters and output layer output are calculated based on the input data through forward propagation. The loss function is calculated, and backpropagation is performed with the goal of minimizing the loss function. The hidden layer error is calculated and the partial derivative is obtained. The weight parameters and bias parameters in the network are updated by gradient descent. The error between the predicted value and the true value is reduced through continuous iteration until the preset number of iterations is reached. The model parameters are saved to obtain the trained A-TCN-DAE model. Then, the test set is recursively input into the trained A-TCN-DAE model, and subsequent capacities are predicted recursively. Once the predicted capacities are obtained... Capacity value corresponding to the preset battery EOL Eligible conditions: End the recursive process; use the mean absolute error (MAE), root mean square error (RMSE), and R... 2 The accuracy of the model's predictions was statistically verified on the test set to measure the metrics, thereby obtaining the A-TCN-DAE model that can be used online.

6. The method for assessing the remaining life status of a lithium battery based on an improved TCN according to claim 5, characterized in that: The process of obtaining the training set and the test set is as follows: 1) Four identical lithium batteries are charged to full charge using constant current and constant voltage CCCV method. Then, lithium batteries No. 1 and No. 2 are discharged at a constant current of 0.5C, and lithium batteries No. 3 and No. 4 are discharged at a constant current of 1C. This process continues until the battery voltage drops to the minimum threshold voltage of 2.7V, which is considered one cycle. 2) Repeat the above operation to obtain the measured voltage and current data of the lithium battery under cyclic conditions until the current capacity of the lithium battery decays to the EOL state, and obtain the voltage and current curves under cyclic conditions. 3) Linear interpolation was used to supplement missing or outlier points in the measured voltage and current data of the lithium battery under cyclic operating conditions: (Equation 1) Where V t V represents the current missing point or outlier. t-1 V represents the data at the previous time point. t+1 This represents the data at the next point in time. 4) Establish a cycle for all voltage data with a cutoff voltage of 2.7V as the cutoff point for slicing, and number the slices sequentially for each cycle: 5) Integrate the ampere-hours of the discharge process during each battery cycle from full charge to full discharge to obtain the battery capacity in the current cycle. (Equation 2) in This represents the capacity of a specific battery cell at time t-1. This represents the capacity of a specific battery cell at time t; 6) The battery capacity within each cycle is averaged according to the sampling time interval to obtain the actual capacity within each cycle. 7) Summarize the cycle number and actual capacity sequence according to the mapping relationship between the segment number and the actual capacity of each cycle. Then, use the cycle number and actual capacity sequence data of lithium batteries No. 2 and No. 4 as the training set, and the cycle number and actual capacity sequence data of lithium batteries No. 1 and No. 3 as the test set.

Images