Lithium battery state of health prediction method based on lstm-ebrb

Abstract

The application provides a lithium battery health state prediction method based on an LSTM-EBRB, and comprises the following steps: step S1: extracting a health factor of a lithium battery as a premise attribute, and constructing a rule base of an EBRB reasoning model; step S2: inputting training set data, performing rule base reasoning, and calculating corresponding activation weights; step S3: taking the activation weights obtained in step S2 as input, and training a long short-term memory (LSTM) neural network model; step S4: inputting prediction object data, setting a prediction starting point, calculating activation weights, inputting the LSTM model in sequence, performing prediction, and obtaining improved activation weights; and step S5: synthesizing activation rules, and obtaining a prediction result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of lithium battery health state prediction, and particularly relates to a lithium battery health state prediction method based on LSTM-EBRB. BACKGROUND

[0002] The treatment and recycling of retired batteries are the realistic needs of green environmental protection and resource regeneration. At present, lithium battery enterprises are actively moving towards the road of digital and intelligent transformation and upgrading, and continuously improving the safety level of the lithium battery industry chain. However, there are still some difficulties and problems in the field of lithium batteries. There are many factors affecting the service life of lithium batteries, and how to build a lithium battery life prediction model with higher interpretability under data-driven to improve the feasibility and accuracy of lithium battery life prediction has become a research difficulty. According to the existing literature, the current lithium battery health state prediction methods can be roughly divided into two categories: one is the mechanism model method based on the physical, chemical or experience system mechanism model of the object to be predicted; the other is the data-driven method based on historical feature data. At present, there are still some problems in the related research of lithium battery health state prediction. First, in the previous model method, the important information of lithium battery characteristics is easy to be ignored. Second, most of the prediction algorithms have low transparency, and the prediction results lack interpretability. However, the existing battery health state prediction model is a "black box" structure, and it is difficult to obtain the decision reason behind the model, and the interpretability is poor.

[0003] The EBRB model can better handle the uncertainty of information, and has high traceability and interpretability. It can reasonably extend and predict the historical data of health state, and is suitable for evaluating lithium battery health state which has multiple related factors and many uncertain factors. It can effectively reduce the complexity of the prediction model while improving the stability and accuracy of the prediction model. It provides guidance for the design, manufacture and use of lithium batteries, and reduces energy costs and maintenance costs. However, due to the obvious time sequence correlation in the health state degradation process of lithium batteries, the traditional EBRB model cannot well capture the time sequence characteristics in the degradation process of lithium batteries. SUMMARY

[0004] Therefore, aiming at the blank and deficiency of the prior art, the purpose of the present application is to provide a lithium battery health state prediction method based on LSTM-EBRB.

[0005] The technical scheme specifically adopted by the present application to solve its technical problems is:

[0006] A lithium battery health state prediction method based on LSTM-EBRB comprises the following steps:

[0007] Step S1: Extract the health factors of the lithium battery as the premise attribute, and construct the rule base of the EBRB reasoning model;

[0008] Step S2: Input the training set data, perform rule base reasoning, and calculate the corresponding activation weight;

[0009] Step S3: Take the activation weight obtained in step S2 as input to train the LSTM model;

[0010] Step S4: Input the prediction object data, set the prediction starting point, calculate the activation weight, input the LSTM model in sequence, perform prediction, and obtain the improved activation weight;

[0011] Step S5: Synthesize the activation rule to obtain the prediction result.

[0012] Further, step S1 specifically comprises:

[0013] Step S11: Preprocess the lithium battery data set, detect outliers and repair;

[0014] Step S12: Extract the isobaric discharge time in each cycle period of the lithium battery; the isobaric discharge time TF matched with the mth battery charge and discharge cycle period m is expressed as:

[0015] TF m = t rv -t cv

[0016] wherein, t rv is the time corresponding to the rated voltage during the discharging process; t cvthe time corresponding to the voltage drop to the cut-off voltage;

[0017] Step S13: Extract the isochronous average voltage drop of the lithium battery in each cycle period; the isochronous average voltage drop VF corresponding to the mth cycle period of the lithium battery is VF = (V1 - V2) / (t1 - t2) m is expressed as:

[0018]

[0019] wherein V d is the voltage corresponding to each time point; d represents the number of sampling intervals of the mth cycle period; t rv represents the rated voltage of the lithium battery;

[0020] Step S14: Extract the charging power peak time of the lithium battery in each cycle period; define 0 seconds as the starting time, use the measured voltage and current data in the battery during the charging process to obtain the time series of electric power, and extract the time corresponding to the maximum value of the electric power in the sequence; the charging power peak time corresponding to the ith cycle period is expressed as:

[0021] PF i = t (p), p = max (p1, …, p K )

[0022] wherein (p1, …, p K ) represents the electric power corresponding to the Kth time in each cycle;

[0023] Step S15: Take the health factors: isochronous average voltage drop, isochronous voltage drop discharge time difference, charging power peak time and cycle number of the lithium battery as characteristic inputs, and construct the premise attribute of EBRB.

[0024] Further, in step S2, the rule form in the rule base is that, for the kth rule R k :

[0025] R k :

[0026]

[0027] wherein represents the confidence of the jth reference level of the ith premise attribute of the kth rule; represents the utility value of the jth reference level of the ith premise attribute of the kth rule; D n (n = 1, 2, …, N) represents the nth evaluation level of the result attribute D, and N represents the number of evaluation levels;

[0028] represents the confidence of the evaluation level D in the kth rule; i represents that the information contained in the kth rule is complete.

[0029] Further, in step S2:

[0030] The historical sample data of the lithium battery is used as the data for generating the rule base; the evaluation level utility value of the premise attribute U and the result attribute and the number thereof are determined;

[0031] Then, the input-output pair for generating the rule base is continuously converted into the same confidence distribution form as the premise attribute and the result attribute of the EBRB inference model; until all the data generate corresponding rules;

[0032] Finally, the rule weight of each rule is calculated to construct a complete EBRB inference model.

[0033] Further, step S2 specifically includes:

[0034] Step S21: divide the training set and the test set, and input the training set into the EBRB inference model;

[0035] Step S22: calculate the individual matching degree

[0036] Step S23: calculate the activation weight ω of each rule k According to the individual matching degree S i (x i , U k ) obtained by each premise attribute U i corresponding to the input data x i , the rule weight θ k and the attribute weight δ k,j , the activation weight of the kth rule is calculated:

[0037]

[0038] According to the size of the activation weight, it is determined whether the current rule needs to be activated, that is, ω k > 0 indicates that the kth rule is an activated rule;

[0039] Step S24: synthesize the activated rules and obtain the final result.

[0040] Further, in step S22, the calculation of the individual matching degree is specifically:

[0041] The input data is converted into a confidence distribution form by the information conversion method based on the utility:

[0042] ​

[0043] wherein

[0044]

[0045] wherein, x k,i represents the input value in the kth input value vector corresponding to the ith premise attribute; represents the numerical value x k,i is converted into the ith premise attribute U i the matching degree on the jth reference level; J i represents the number of reference levels of the ith premise attribute; u(A i,j ) represents the utility value of the jth reference level of the ith premise attribute; u(A i,j+1 ) represents the utility value of the j+1th reference level of the ith premise attribute;

[0046] Then, the individual matching degree S k (x i , U i ) of the ith premise attribute in the kth extended confidence rule is calculated, and the calculation formula is as follows:

[0047]

[0048] Further, the step S3 specifically comprises:

[0049] Step S31: normalizing data; the obtained activation weight data is normalized to eliminate the influence of dimension, and the normalization formula is as follows:

[0050]

[0051] wherein, μ and σ are the average value and standard deviation of the data, and x and x' are the data before and after normalization, respectively;

[0052] Step S32: inputting sequence data into the LSTM model, and updating the hidden state and cell state at each time step;

[0053] Step S33: generating a sequence at each time step;

[0054] Step S34: converting the model output into an interpretable format, and reversing the normalization as follows:

[0055] x = x standard * σ + μ

[0056] wherein x standard is the normalized data, σ is the standard deviation of the original data, and μ is the average value of the original data.

[0057] Further, the step S4 specifically comprises:

[0058] Step S41: For the prediction object data set, set the prediction starting point, and input the data before the prediction starting point to the EBRB inference model;

[0059] Step S42: Match with the rule base, and calculate the individual matching degree S k (x i ,U i )

[0060] Step S43: Calculate the activation weight ωk of each rule

[0061] Step S44: Input the obtained activation weight in sequence to the LSTM model, start prediction from the prediction starting point, and obtain the improved activation weight.

[0062] Further, in step S5:

[0063] The distributed confidence of the result attribute in all activation rules is synthesized into a new distributed confidence by using the analytical formula of ER:

[0064]

[0065] In the above formula, N represents the number of evaluation grades; L represents the total number of rules;

[0066] Let u(D n ) be the utility value of the nth result grade D n , n = 1, …, N, and the prediction value obtained based on the EBRB inference model is:

[0067]

[0068] Compared with the prior art, the present application and the preferred schemes thereof have at least the following outstanding features and beneficial effects:

[0069] 1. On the EBRB model, an improved activation weight fusing time series prediction is proposed. Compared with the original activation weight, the improved activation weight introduces a long short-term memory (LSTM) model for time series prediction, considers the time series characteristics of the lithium battery health state decay, and can better capture the time series changes of the lithium battery health state, thereby improving the accuracy and reliability of the prediction results.

[0070] 2、The data-driven EBRB inference model adopted can better process the uncertainty of information, has high traceability and interpretability, is suitable for evaluating the lithium battery health state which has multiple related factors, complex internal working conditions and many uncertain factors, can effectively reduce the complexity of the prediction model while improving the stability and accuracy of the prediction model, and improves the robustness, accuracy and interpretability of the prediction model. A scientific basis is provided for the management and maintenance of lithium batteries, and the reliability and safety of lithium batteries are improved. BRIEF DESCRIPTION OF DRAWINGS

[0071] The application will be further described in detail below in combination with the drawings and specific embodiments:

[0072] Figure 1 is an EBRB inference model schematic diagram;

[0073] Figure 2 is a health factor and lithium battery capacity correlation curve in the embodiment of the application;

[0074] Figure 3 is an improved activation weight prediction method diagram of LSTM-EBRB in the embodiment of the application;

[0075] Figure 4 is a method flowchart of the embodiment of the application. DETAILED DESCRIPTION

[0076] In order to make the features and advantages of the patent more obvious and easy to understand, the following specific examples are described in detail as follows:

[0077] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise indicated, all technical and scientific terms used in the specification have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0078] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form, and in addition, it should be understood that when the terms "comprise" and / or "include" are used in the specification, they indicate the presence of a feature, step, operation, device, component and / or combination thereof.

[0079] As shown in Figure 4 The embodiment of the application provides a lithium battery health state prediction method based on LSTM-EBRB, which comprises the following steps:

[0080] Step S1: Extract the health factor of the lithium battery as the premise attribute, and construct the rule base of the EBRB reasoning model

[0081] Step S3: Take the activation weight obtained in S2 as input, and train a long short-term memory (LSTM) neural network model

[0082] Step S4: Input the prediction object data, set the prediction starting point, calculate the activation weight, input the sequence into the LSTM model, and perform prediction to obtain the improved activation weight

[0083] Step S5: Synthesize the activation rules to obtain the prediction result

[0084] As a preferred scheme of the embodiment, first, the health factor of the lithium battery is extracted as the premise attribute, and the specific steps are as follows:

[0085] (1): Preprocess the lithium battery data set, detect and repair outliers.

[0086] (2): Extract the isobaric discharge time in each cycle of the lithium battery. The isobaric discharge time TF matched with the mth battery charge and discharge cycle period is m , which is expressed as:

[0087] TF m = t rv -t cv

[0088] Wherein, t rv is the time when the voltage is the rated voltage during discharging; t cv is the time when the voltage drops to the cutoff voltage.

[0089] (3): Extract the equal-time average voltage drop in each cycle of the lithium battery. Then the equal-time average voltage drop VF corresponding to the mth cycle is m , which is expressed as:

[0090]

[0091] Wherein, V d is the voltage corresponding to each time point; d represents the sampling interval number of the mth cycle; t rv represents the rated voltage of the lithium battery.

[0092] (4): Extract the charging power peak time in each cycle of the lithium battery. Define 0 seconds as the starting time, use the measured voltage and current data in the battery during charging to obtain the electric power time sequence, and extract the time corresponding to the maximum value of the electric power in the sequence. Then the charging power peak time corresponding to the ith cycle is expressed as:

[0093] PF i =t(p), p = max(p1,…,p K )

[0094] Among them, (p1,…,p K ) represents the electrical power at the Kth moment in each cycle.

[0095] (5): The three health factors extracted above—average voltage drop over equal time, equal voltage drop discharge time difference, charging power peak time, and lithium battery cycle number—can be used as feature inputs to construct the prerequisite attributes for ERB.

[0096] In this embodiment, the rule form in the EBRB inference model is constructed, with the k-th rule R... k For example, as shown below:

[0097] R k :

[0098]

[0099] in The confidence level of the j-th reference level for the i-th premise attribute of the k-th rule; D represents the utility value of the j-th reference level of the i-th premise attribute of the k-th rule; n (n = 1, 2, ..., N) represents the nth evaluation level of the result attribute D, and N represents the number of evaluation levels; Represents the relationship between the rating level D and the k-th rule. i Confidence level; This means that the information contained in rule k is complete.

[0100] The specific steps for constructing the rule base of the EBRB inference model are as follows:

[0101] (1) Use historical sample data of lithium batteries as data to generate the rule base;

[0102] (2) Determine the evaluation level utility values ​​and the number of the premise attribute U and the outcome attribute;

[0103] (3) Transform the input-output pairs used to generate the rule base into the same confidence distribution form as the premise and outcome attributes of the EBRB inference model;

[0104] (4) Continue to execute the above steps (3) until all the data generates the corresponding rules;

[0105] (5) Calculate the rule weight of each rule. At this point, a complete EBRB inference model is constructed.

[0106] After the EBRB is constructed, the EBRB inference model can perform rule inference according to new input data, and the traditional EBRB inference process is as shown in Figure 1 If the activation weight sequence data of the lithium battery data matching rule is needed to train the LSTM model, the specific steps of the EBRB-based inference process are as follows:

[0107] (1) Divide the training set and the test set, and input the training set into the EBRB inference model

[0108] (2) Calculate the individual matching degree, specifically:

[0109] First, the input data is converted into a belief distribution form through the information conversion method based on utility

[0110]

[0111] wherein

[0112]

[0113] In the formula, x k,i represents the input value corresponding to the i-th premise attribute in the k-th input value vector; represents the value x k,i in the k-th input value vector converted to the i-th premise attribute U i on the j-th reference level; J i represents the number of reference levels of the i-th premise attribute; u(A i,j ) represents the utility value of the j-th reference level of the i-th premise attribute; and u(A i,j+1 ) represents the utility value of the j+1-th reference level of the i-th premise attribute.

[0114] Then, the individual matching degree S k (x i ,U i ) of the i-th premise attribute in the k-th extended belief rule is calculated, and the calculation formula is:

[0115]

[0116] (3) Calculate the activation weight ω k of each rule, and according to the individual matching degree S i (x i ,U k ) obtained from the input data x i corresponding to each premise attribute U i , the rule weight θ k and the attribute weight δ k,j , the activation weight of the k-th rule is calculated

[0117]

[0118] At this point, the activation weight data is obtained, and the training of the LSTM model is performed, with the following steps:

[0119] (1) Standardize the data

[0120] Standardize the activation weight data to eliminate the dimensional effect, and the standardization formula is as follows:

[0121]

[0122] Where μ and σ are the mean and standard deviation of the data, and x and x' are the data before and after standardization.

[0123] (2) Input the sequence data into the LSTM model, and update the hidden state and cell state at each time step. The core of LSTM is the design of its cell state and the input gate that decides which new information will be stored in the cell state, the forget gate that decides which information will be discarded or retained, and the output gate that decides the next hidden state. These gates control the flow of information.

[0124] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0125] Where f t is the output of the forget gate, σ is the sigmoid activation function, W f and b are the weights and biases of the forget gate, t -1 is the hidden state of the previous time step, and x t is the input of the current time step.

[0126] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0127]

[0128] Where i t is the output of the input gate, is the candidate cell state, which is used to update the cell state. Combining the information of the forget gate and the input gate, the cell state can be updated.

[0129]

[0130] Where is the cell state at the current time step, and is the cell state at the previous time step.

[0131] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0132] h t =o t *tanh(C t )

[0133] Among them, o t It is the output of the output gate, h t It is the hidden state of the current time step.

[0134] (3) Perform prediction and sequence generation at each time step.

[0135] (4) Convert the model output into an interpretable format by denormalizing it using the following formula:

[0136] x = x stabdard *σ+μ

[0137] Where x standard These are the standardized data, where σ is the standard deviation of the original data and μ is the mean of the original data.

[0138] After the LSTM model is trained, for the dataset to be predicted, a prediction starting point is set, and the data before the prediction starting point is input into the ERB inference model to improve the prediction of activation weights. The specific steps are as follows:

[0139] (1) Match with the rule base and calculate the individual matching degree S. k (x i U i )

[0140] (3) Calculate the activation weight ω for each rule. k

[0141] (4) Input the obtained activation weights into the LSTM model in sequence form, start the prediction from the prediction starting point, and obtain the improved activation weights.

[0142] Finally, the activation rule is synthesized to obtain the predicted health status of the lithium battery, specifically:

[0143] Using the analytical formula of ER, the distributed confidence scores of the outcome attributes in all activation rules are combined into a new distributed confidence score:

[0144]

[0145] In the above formula, N represents the number of evaluation grades; L represents the total number of rules;

[0146] Let u(D n ) be the utility value of the nth(n = 1, …, N) result grade D n , and the prediction value obtained based on the EBRB inference model is:

[0147]

[0148] The implementation process and efficacy of the present application scheme are further demonstrated and introduced below through specific test examples:

[0149] This example takes the charge-discharge data of lithium batteries under the full life cycle collected by NASA as an example. According to the above technical scheme, in the process of predicting the state of health of lithium batteries, feature extraction is first needed, that is, to extract health factors. After preprocessing, the lithium battery health factor dataset is obtained to model and predict the EBRB inference model.

[0150] The NASA dataset is in MATLAB data storage format. Four batteries B0005, B0006, B0007 and B0018 are used as data sources. The four battery samples are all subjected to charge-discharge cycles at an ambient temperature of 24 degrees, and different charge-discharge conditions are set. When the lithium battery is in a charging state, the battery monomer is charged in a constant current (CC) mode at 1.5A, and after its voltage reaches the rated voltage of 4.2V, it is charged in a constant voltage (CV) mode until its charging current drops to 20mA. When it is in a discharging state, the battery monomer is discharged in a CC mode at 2A until the voltages of the batteries B0005, B0006, B0007 and B0018 drop to 2.7V, 2.5V, 2.2V and 2.5V, respectively. When the battery reaches its failure threshold, that is, its rated capacity drops to 70% of the original, the test is terminated. The data structure of each group of battery samples is the same, and the charge-discharge data in the lithium battery cycle period is used for research, as shown in Table 1.

[0151] Table 1 Data structure

[0152]

[0153] Before further data analysis and model training, in order to optimize data quality, data preprocessing is needed first, and the data set of the four lithium battery samples is subjected to outlier detection and repair. In this example, the 3sigma outlier detection method is used, so as to retain the normal trend of the battery data in the charge-discharge process. The outlier detection formula of the 3sigma method is as follows:

[0154]

[0155] Wherein, μ represents the mean of window data; σ represents the standard deviation of window data.

[0156] In this embodiment, the health factor of the lithium battery is extracted, and the specific process is as follows:

[0157] The data is visualized and analyzed to obtain four curves as shown in Figure 2 , which are respectively: the curve of the capacity of the lithium battery sample changing with the number of charge and discharge cycles, the curve of the constant pressure drop discharge time changing with the number of cycles, the curve of the internal voltage changing with time during the discharge process, and the curve of the electric power changing with time during the charging process.

[0158] It can be seen that there is a negative correlation between the capacity of the lithium battery and the number of charge and discharge cycles, and the capacity gradually decreases with the increase of the number of cycles; the constant pressure drop discharge time decreases significantly with the increase of the number of charge and discharge cycles; the more the number of charge and discharge cycles, the shorter the time taken for the battery to drop to the lowest point, and the more significant the voltage drop within the same time.

[0159] The extracted health factors are analyzed for correlation. The Spearman correlation coefficient analysis method is used to quantitatively process the correlation between the three input indicators and the health status of the lithium battery. The Spearman correlation coefficient eliminates the data dimension by ranking. The calculation formula is as follows:

[0160]

[0161] Wherein, d i = R i -S i .

[0162] The results of the Spearman correlation coefficient analysis are shown in Table 2.

[0163] Table 2: Spearman correlation analysis between the health status of the lithium battery and the health factors

[0164] Lithium battery model Average voltage drop in equal time Discharge time difference in equal voltage drop Charging power peak time B0005 -0.929 0.925 0.993 B0006 -0.987 0.954 0.999 B0007 -0.918 0.948 0.993 B0018 -0.997 0.904 0.997 Average value -0.958 0.933 0.996

[0165] In summary, the absolute value of the Spearman correlation coefficient between the three health factors and the capacity of the lithium battery is greater than 0.9, verifying the strong correlation between the average pressure drop within the same time, the constant pressure drop discharge time difference, the charging power peak time and the health status of the lithium battery, and the reliability of the three as the premise attribute input.

[0166] In this embodiment, the rule form in the EBRB reasoning model is constructed, and the kth rule R k is taken as an example, as shown below:

[0167] R k :

[0168]

[0169] wherein represents the confidence of the jth reference level of the ith premise attribute of the kth rule; represents the utility value of the jth reference level of the ith premise attribute of the kth rule;D n (n = 1, 2, …, N) represents the nth evaluation level of the result attribute D, and N represents the number of evaluation levels;

[0170] represents the confidence of the evaluation level D i in the kth rule; represents that the information contained in the kth rule is complete.

[0171] In this embodiment, the specific steps of constructing the rule base of the EBRB reasoning model are as follows:

[0172] (1) The historical sample data of the lithium battery is used as the data for generating the rule base;

[0173] (2) Based on expert knowledge and training data, the evaluation level utility value and the number of premise attributes U and result attributes are determined; in this embodiment, the optimal parameters are obtained by adjusting the parameters, and the evaluation level of the premise attribute is set to 4 and the evaluation level of the result attribute is set to 5.

[0174] (3) The input-output pair used to generate the rule base is converted into the same confidence distribution form as the premise attribute and the result attribute of the EBRB reasoning model;

[0175] (4) Step (3) is continuously executed until all data generates corresponding rules;

[0176] (5) The rule weight of each rule is calculated, and thus a complete EBRB reasoning model is constructed.

[0177] After the EBRB is constructed, the EBRB reasoning model can perform rule reasoning according to new input data. To obtain the activation weight sequence data of the lithium battery data matching rule for training the LSTM model, the specific steps of the EBRB reasoning process are as follows:

[0178] (1) Divide the training set and the test set, which is set to 7:3 here, and input the training set into the EBRB reasoning model.

[0179] (2) Calculate the individual matching degree, which is specifically:

[0180] First, the input data is converted into a confidence distribution form by the information conversion method based on the utility

[0181]

[0182] wherein

[0183]

[0184] wherein, x k,i represents the input value in the kth input value vector corresponding to the ith premise attribute; represents the numerical value x k,i in the kth input value vector converted into the ith premise attribute U i the matching degree on the jth reference level; J i represents the number of reference levels of the ith premise attribute; u(A i,j ) represents the utility value of the jth reference level of the ith premise attribute; u(A i,j+1 ) represents the utility value of the j+1th reference level of the ith premise attribute;

[0185] Then, the individual matching degree S k (x i ,U i ) of the ith premise attribute in the kth extended confidence rule is calculated, and the calculation formula is as follows:

[0186]

[0187] (3) The activation weight ω k of each rule is calculated, and the individual matching degree S k (x i ,U i ) obtained by the input data x i corresponding to each premise attribute U i , the rule weight θ k and the attribute weight δ k,j are used to calculate the activation weight ω

[0188]

[0189] At this point, the activation weight data is obtained, and the training of the LSTM model is performed, and the steps are as follows:

[0190] (1) Standardize the data

[0191] Standardize the activation weight data to eliminate the influence of the dimension, and the standardization formula is as follows:

[0192]

[0193] wherein, μ and σ are the mean and standard deviation of the data, and x and x' are the data before and after standardization.

[0194] (2) Input sequence data into the LSTM model and update the hidden state and cell state at each time step. The core of LSTM lies in the design of its cell state and the input gate that decides which new information will be stored in the cell state, the forget gate that decides which information will be discarded or retained, and the output gate that decides the next hidden state, which control the flow of information.

[0195] f t = σ(W f · [h t-1 , x t ] + b f )

[0196] where f t is the output of the forget gate, σ is the sigmoid activation function, W f and b f are the weights and biases of the forget gate, h t-1 is the hidden state of the previous time step, and x t is the input of the current time step.

[0197] i t = σ(W i · [h t-1 , x t ] + b i )

[0198]

[0199] where i t is the output of the input gate, is the candidate cell state, which is used to update the cell state. Combining the information of the forget gate and the input gate, the cell state can be updated.

[0200]

[0201] where C t is the cell state of the current time step, and C t is the cell state of the previous time step.

[0202] o t = σ(W o · [h t-1 , x t ] + b o )

[0203] h t = o t * tanh(C t )

[0204] where o t is the output of the output gate, and h t is the hidden state of the current time step.

[0205] (3) Generate a sequence of predictions at each time step

[0206] (4) Convert the model output into an interpretable format by denormalizing it as follows:

[0207] x = x standard * σ + μ

[0208] where x standard is the standardized data, σ is the standard deviation of the original data, and μ is the mean of the original data.

[0209] In this embodiment, the parameter settings for training the LSTM model are shown in Table 3.

[0210] Table 3 LSTM parameter settings

[0211] Parameter name Parameter value Number of neurons 50 Optimization method Adam Loss function MSE Epoch 50

[0212] After the LSTM model is trained, for the prediction object data set, set the prediction starting point. In this embodiment, refer to Figure 3 , set the prediction starting point to 9. Input the data before the prediction starting point, i.e., the first 8 data, into the EBRB inference model, and then perform prediction of the improved activation weight. The specific steps are as follows:

[0213] (1) Match with the rule base and calculate the individual matching degree S k (x i , U i )

[0214] (3) Calculate the activation weight ω k of each rule

[0215] (4) Input the obtained activation weight in sequence form into the LSTM model, start prediction from the 9th at the prediction starting point, and obtain the improved activation weight.

[0216] Finally, synthesize the activation rules to obtain the lithium battery health state prediction result, which is specifically:

[0217] Use the analytical formula of ER to synthesize the distributed confidence of the result attribute in all activation rules into a new distributed confidence:

[0218]

[0219] In the above formula, N represents the number of evaluation grades; L represents the total number of rules;

[0220] Let u(D n ) be the utility value of the nth(n = 1, …, N) result grade D n , then the prediction value obtained based on the EBRB inference model is:

[0221]

[0222] In this embodiment, different LSTM parameters are set for experiments, which effectively verify that the EBRB model based on improved activation weight of the new invention has high effectiveness and robustness, and is superior to the original EBRB model in performance. The specific parameter settings and evaluation indexes are shown in Table 4:

[0223] Table 4 Model Comparison

[0224]

[0225]

[0226] In summary, in this embodiment, experiments show that the model has better adaptability to changes in different data and conditions, and has significant advantages in prediction accuracy and robustness.

[0227] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0228] The application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0229] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0230] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable devices provide a process for implementing the flowchart Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0231] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the present application should be understood as the usual meaning understood by those skilled in the art to which the present application belongs. The "first", "second" and similar words used in the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Include" or "contain" and similar words mean that the elements or objects before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connected" or "connected" and similar words are not limited to physical or mechanical connection, but can include electrical connection, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to represent relative positional relationship, when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0232] The above is only the preferred embodiment of the present application, and is not intended to limit the other forms of the present application. Any skilled person in the art can use the disclosed technical content to make changes or modifications as equivalent embodiments. However, any simple modification, equivalent change and modification of the above embodiments made without departing from the technical solution of the present application, according to the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

[0233] The present patent is not limited to the above best mode, and anyone can derive other various forms of lithium battery state of health prediction method based on LSTM-EBRB under the inspiration of the present patent. Any equivalent change and modification made according to the scope of the present patent application shall be covered by the present patent.

Claims

1. A method for predicting the state of health of a lithium battery based on LSTM-EBRB, characterized in that, Comprise the following steps: Step S1: extract the health factor of lithium battery as the premise attribute, construct the rule base of EBRB reasoning model; Step S2: input the training set data, perform rule base reasoning, and calculate the corresponding activation weight; Step S3: input the activation weight obtained in step S2 as input to train the LSTM model; Step S4: input the prediction object data, set the prediction starting point, calculate the activation weight, input the LSTM model in sequence, and perform prediction to obtain the improved activation weight; Step S5: synthesize the activation rule to obtain the prediction result; Step S1 specifically comprises: Step S11: pre-process the lithium battery data set, detect outliers and repair; Step S12: Extracting the isobaric discharge time of each cycle of lithium battery; the isobaric discharge time TF matched with the mth battery charge-discharge cycle m is expressed as: TF m = t rv - t cv wherein t rv is the time corresponding to the rated voltage during the discharge process; t cv is the time corresponding to the cut-off voltage. Step S13: Extracting the isochronal average voltage drop of each cycle of lithium battery; then the isochronal average voltage drop VF corresponding to the mth battery charge and discharge cycle m is expressed as: wherein V d is the voltage corresponding to each time point; d represents the number of sampling intervals of the mth cycle; t rv represents the rated voltage of the lithium battery; Step S14: extract the charging power peak time in each cycle of the lithium battery; define 0 seconds as the starting time, use the measured voltage and current data inside the battery during charging to obtain the electric power time series, and extract the time corresponding to the maximum electric power in the sequence; the charging power peak time corresponding to the i th cycle is expressed as: PF i =t(p),p=max(p1,…,p K ) wherein (p1,...,p K ) represents the electrical power corresponding to the Kth time instant at each cycle; Step S15: take the health factors: average pressure drop in the same time, discharge time difference under the same pressure drop, charging power peak time, and cycle number of lithium battery as characteristic input to construct the premise attribute of EBRB; In step S2, the rules in the rule base are in the form that, for the kth rule R k : wherein a confidence value of the jth reference level of the ith premise attribute representing the kth rule; a utility value of the jth reference level of the ith premise attribute representing the kth rule;D n (n = 1, 2, …, N) represents the nth evaluation level of the result attribute D, and N represents the number of evaluation levels; represents the confidence of the evaluation level D in the kth rule; i represents that the information contained in the kth rule is complete;​ In step S2: The historical sample data of the lithium battery is used as the data for generating the rule base; the evaluation level utility value of the premise attribute U and the result attribute and the number thereof are determined; Then constantly convert the input-output pairs used to generate the rule base into the same confidence distribution form as the premise attribute and the result attribute of the EBRB reasoning model; until all data generate corresponding rules; Finally, calculate the rule weight of each rule to construct a complete EBRB reasoning model; Step S2 specifically comprises: Step S21: divide the training set and the test set, and input the training set into the EBRB reasoning model; Step S22: calculate the individual matching degree Step S23: Calculate the activation weight ω for each rule. k According to each prerequisite attribute U i Corresponding input data x i The obtained individual matching degree S k (x i U i ), rule weight θ k and attribute weight δ k,j Calculate the activation weight of the k-th rule: The current rule is determined whether to be activated according to the size of the activation weight, that is, ω k > 0 indicates that the kth rule is an activated rule; Step S24: synthesize the activation rule and obtain the final result; In step S22, the calculation of the individual matching degree is specifically: Convert the input data into a confidence distribution form through the information conversion method based on the utility: Wherein wherein x k,i represents the input value in the kth input value vector corresponding to the ith premise attribute; represents the numerical value x k,i is converted to the ith premise attribute U i the matching degree on the jth reference level; J i represents the number of reference levels of the ith premise attribute; u(A i,j ) represents the utility value of the jth reference level of the ith premise attribute; u(A i,j+1 ) represents the utility value of the j+1th reference level of the ith premise attribute; Then the individual matching degree S of the ith premise attribute in the kth extended confidence rule is calculated k (x i ,U i ), the calculation formula is:

2. The LSTM-EBRB-based lithium battery state of health prediction method of claim 1, wherein: Step S3 specifically comprises: Step S31: standardize the data; standardize the obtained activation weight data to eliminate the dimension effect, and the standardization formula is as follows: Wherein, μ and σ are the average value and standard deviation of the data, and x and x' are the data before and after standardization respectively; Step S32: input the sequence data into the LSTM model, and update the hidden state and cell state at each time step; Step S33: perform prediction to generate a sequence at each time step; Step S34: convert the model output into an interpretable format, and reverse standardize it according to the following formula: x = x standard *σ + μ where x standard is the normalized data, σ is the standard deviation of the original data, and μ is the mean of the original data. 3.The LSTM-EBRB-based lithium battery state of health prediction method of claim 1, wherein: Step S4 specifically comprises: Step S41: for the prediction object data set, set the prediction starting point, and input the data before the prediction starting point to the EBRB reasoning model; Step S42: Matching with the rule base, calculating individual matching degree S k (x i ,U i ) Step S43: Calculate the activation weight ω of each rule k Step S44: input the obtained activation weight in sequence to the LSTM model, and start prediction from the prediction starting point to obtain the improved activation weight.

4. The LSTM-EBRB-based lithium battery state of health prediction method of claim 3, wherein: In step S5: The distributed confidence of the result attribute in all the activation rules is synthesized into a new distributed confidence by using the analytical formula of ER: In the above formula, N represents the number of evaluation grades; L represents the total number of rules; Let u(D n ) be the utility value of the nth result rank D n , n = 1, …, N, then the prediction value based on the EBRB inference model is:

Images