CN112036084B - A kind of similar product lifetime migration screening method and system - Google Patents

Abstract

The invention discloses a similar product life migration prediction method and system, and relates to the technical field of similar product migration learning. training data; by performing minimum screening of curve shape, capacity degradation rate similarity, life distribution similarity and distance metric minimum, obtain transferable sample data for life prediction of similar products across recipes, and use the transferable sample data adapted to the transferable sample data. The life prediction model performs life migration prediction for similar products across formulas, and obtains life prediction results; the invention realizes the accurate prediction of the remaining life of lithium ion batteries across formulas, and the prediction accuracy can reach up to 99.9%, which can effectively save the design and development process of lithium batteries In the test time and cost, it has considerable economic benefits and application value.

Description

A kind of similar product lifetime migration screening method and system

technical field

The invention relates to the technical field of migration life prediction of similar products, in particular to a similar product life migration screening method and system.

Background technique

With the development of energy storage technology and energy industry, lithium-ion batteries are widely used in military electronic products, avionics, electric vehicles and various portable electronic devices (such as notebooks) due to their advantages of light weight, low discharge rate and long life. Main energy storage devices for computers, digital cameras, tablet computers, mobile phones, etc. Cycle life is an important design property of lithium battery products. In the design and development process of lithium batteries, in order to accurately obtain the life of batteries with different formulas in the design matrix and provide feedback for formula selection and design optimization, it is necessary to carry out cycle life tests for batteries with each formula in the design matrix. The test needs to continue until the battery capacity retention rate reaches a specified threshold, that is, the end-of-life point of the lithium battery. However, due to the large number of lithium battery recipes in the design matrix, the existing cycle life test time and capital cost are too high, especially for power lithium batteries with a life cycle of many years, the design and development efficiency is too low, and enterprises unbearable.

Lithium battery life prediction is usually used in the use stage of lithium batteries, mainly based on a small amount of known historical data, to predict the remaining life of the battery at the current moment. Since the number of test cycles of the predicted lithium battery should be as small as possible, it is difficult to obtain enough battery test data to meet the design and development needs of the remaining life prediction model. Therefore, using the massive historical cycle life test data of other recipes of the same battery platform of the battery enterprise can provide data support for the remaining life prediction model required for design and development. However, how to define and measure data transferability, and use it to design a screening strategy for transferable samples, to obtain the most similar samples from a large number of other battery data with differentiated formulations, and how to use the most similar samples for life prediction of similar products is of great significance and application requirements.

SUMMARY OF THE INVENTION

In order to solve the problems existing in the prior art, the present invention provides a transferable sample screening method and system that is more suitable for battery life prediction between different formulations in the battery design and development process by using short-term measured data and a migration learning prediction method. The migration sample screening method and system provided can improve the accuracy of life prediction, effectively avoid energy consumption and resource waste caused by long-term testing, and have high prediction accuracy and strong universality.

In order to achieve the technical purpose of the present invention, one aspect of the present invention provides a method for predicting the lifespan of similar products, which is characterized in that, comprising:

Preprocess the short-term cycle life test data of products with similar formula to be tested to obtain target sample data, and preprocess the full-life test capacity data of other products with similar formula to obtain multiple training data;

By performing curve shape screening, the first training data that is similar to the curve type of the target sample data is screened from the plurality of training data;

Through the capacity degradation rate similarity screening, the second training data that is similar to the capacity degradation trend of the target sample data is screened from the first training data;

Through the similarity screening of the lifespan distribution, the third training data with a lifespan distribution similar to the target sample data is selected from the second training data;

Through the minimum distance metric screening, the fourth training data with the smallest distance metric from the target sample data is selected from the third training data;

Use the fourth training data as transferable sample data for life prediction of similar products across recipes;

Use a lifetime prediction model adapted to the transferable sample data to perform lifetime migration predictions for similar products across formulations.

Wherein, using the lifespan prediction model adapted to the transferable sample data to predict the lifespan migration of similar products across recipes includes:

Use the transferable sample data to fine-tune and train models previously trained for other products with similar formulations to obtain the lifespan prediction model;

Using the life prediction model to perform life prediction processing on the current test data of the products with similar formula to be tested, to obtain the remaining service life RUL of the products with similar formula to be tested.

Wherein, the preprocessing includes normalizing the target sample data and the training data;

Described using the life prediction model to perform life prediction processing on the current test data of the products with similar formulations to be tested, and obtaining the remaining service life RUL of the similar products with similar formulations to be tested includes:

Input the current test data of the product with similar formula to be tested into the life prediction model, and output the RUL label value of the remaining service life of the product with similar formula to be tested;

The RUL label value of the remaining service life of the product with a similar formula to be tested is reverse-normalized, and the predicted value of the remaining service life RUL of the product with a similar formula to be tested is obtained.

Wherein, filtering out the first training data with a curve type similar to the target sample data from multiple training data includes:

The target sample data and multiple training data are graphed into target sample data curves and multiple training data curves respectively;

Divide the target sample data curve and multiple training data curves into three categories: straight line, concave curve and convex curve;

Screening is performed according to the types of straight lines, concave curves and convex curves, and the training data curves different from the target sample data curve types are excluded to obtain the first training data.

Wherein, filtering out the second training data that is similar to the capacity degradation trend of the target sample data from the first training data includes:

Calculate the rate of change of the capacity curve when the first training data degenerates from the initial state to the end of the test, and retain several first training data closest to the target sample data;

The reserved first training data is used as the second training data.

Wherein, filtering out the third training data that is similar to the target sample data lifespan distribution from the second training data includes:

Compare the lifetime distribution of the second training data by measuring the number of cycles when running to the trial stop threshold, and retain several second training data that are closest to the lifetime distribution of the target sample data;

The retained second training data that is closest to the lifetime distribution of the target sample data is used as the third training data.

Wherein, the fourth training data with the smallest distance metric from the target sample data selected from the third training data includes:

Select the Chebyshev distance to filter the capacity curve, and calculate the Chebyshev distance between the degradation curves of several third training data and the capacity degradation curve of the target sample data;

The third training data with the smallest Chebyshev distance is selected as the fourth training data.

Wherein, the preprocessing also includes:

Eliminate unstable initial data and data that do not reflect the degradation trend in the short-term test sample data of products with similar formula to be tested and the full-life test capacity database of other products with similar formula;

Smoothing the removed short-term test sample data and the data in the full-life test capacity database to obtain target sample data and prediction training data.

Wherein, the similar product is a lithium battery.

Further, the similar product is a lithium-ion battery.

In order to achieve the technical purpose of the present invention, another aspect of the present invention provides a life migration prediction system for similar products across recipes, which is characterized in that it includes a processor and a memory, which are stored on the memory and run on the processor. A program and a data bus for implementing connection and communication between the processor and the memory, the program when executed by the processor implements a cross-formulation similar product life migration prediction method.

Beneficial effects:

Due to the short test data of the predicted battery, it is difficult to effectively highlight its capacity degradation law, which makes it difficult to effectively train the life prediction model and give accurate prediction results. In order to solve this problem, the present invention provides a transfer learning based method, which adopts a four-time screening method for transferable samples, and obtains the highest similarity with the predicted battery capacity degradation law from the historical full-life test data of batteries with different formulas. The data is transferred to the training of the predicted battery life prediction model, and the accurate prediction of the remaining life of lithium power batteries across recipes is realized. The prediction accuracy can reach 99.9%, which can effectively save the test time in the design and development process of lithium batteries. And cost, with considerable economic benefits and application value.

Description of drawings

1 is a flowchart of a method for predicting the lifespan migration of similar products provided in Embodiment 1 of the present invention;

2 is a flow chart of a method for predicting the lifespan of a lithium-ion battery provided by Application Example 1 of the present invention;

Fig. 3 is the construction and testing process diagram of the Tr-LSTM-based RUL transfer prediction model provided by application embodiment 1 of the present invention;

4 is a schematic diagram of the RNN structure provided by Application Example 1 of the present invention;

5 is a schematic diagram of the structure of the LSTM unit provided by Application Example 1 of the present invention;

6 is a diagram of a Tr-LSTM-based RUL prediction model and its transfer learning strategy provided by Application Example 1 of the present invention;

Fig. 7 is the linear fitting of the prediction curve carried out by slope and intercept provided by the application example 1 of the present invention;

Fig. 8 is the experimental process design optimization flow that the application example 1 of the present invention provides;

Figure 9 is a graph of the data preprocessing results of battery degradation data under three temperature conditions of 25°C, 45°C and 60°C provided by Test Example 1, wherein Figure 9(a) is the data curve at 25°C, and Figure 9(b) is The data curve at 45°C, Figure 9(c) is the data curve at 60°C;

Figure 10 is a schematic diagram of the screening results of migration samples provided in Test Example 1 under three temperature conditions of 25°C, 45°C and 60°C.

Detailed ways

The method and system of the present invention will be described in more detail below with reference to the schematic diagrams, which represent preferred embodiments of the present invention, and it should be understood that those skilled in the art can modify the present invention described herein and still achieve the advantageous effects of the present invention. Therefore, the following description should be construed as widely known to those skilled in the art and not as a limitation of the present invention.

The invention is described in more detail by way of example in the following paragraphs with reference to the accompanying drawings. The advantages and features of the present invention will become apparent from the following description and claims. It should be noted that, the accompanying drawings are all in a very simplified form and in inaccurate scales, and are only used to facilitate and clearly assist the purpose of explaining the embodiments of the present invention.

The experimental methods used in the following examples are conventional methods unless otherwise specified. The structures, materials, etc. used in the following examples can be obtained from commercial sources unless otherwise specified.

Example 1

As shown in Figure 1, a similar product migration sample screening method provided by the present invention includes:

Step S101 preprocesses short-term cycle life test data of similar products with a formula to be tested to obtain target sample data, and preprocesses battery life test capacity data of other formulas to obtain a plurality of training data;

Specifically, the preprocessing includes:

Eliminate unstable initial data and data that do not reflect the degradation trend in the short-term test sample data of the formula battery to be tested and the full-life test capacity database of other formula batteries;

Smoothing the removed short-term test sample data and the data in the full-life test capacity database to obtain target sample data and prediction training data.

Further, before the smoothing process, it also includes:

Normalize the deleted short-term test sample data and the data in the full-life test capacity database.

In order to ensure the consistency of data scales in cross-recipe prediction, the present invention first smoothes the original data curve of the lithium battery capacity retention rate based on the method of local weighted regression to obtain normalized preprocessing data. The invention selects 80% of the initial capacity of the lithium battery as the life end point (failure threshold), and normalizes its capacity data and cycle life data, the initial capacity value is 1, and the capacity value when running to the failure threshold is 0, the input data of the prediction model and the corresponding cycle life label are obtained. The specific calculation steps are as follows:

1) Data normalization

82% of the initial capacity of the lithium battery is selected as the failure threshold, and the capacity value of the lithium battery is normalized to 1-0 (the initial capacity is 1, and 82% of the initial capacity is 0). At the same time, the RUL label represents the RUL value of the battery corresponding to each test cycle, which also needs to be normalized (the remaining life corresponding to the initial capacity is 1, and the remaining life corresponding to 82% of the initial capacity is 0). The process is shown below.

2) Smoothing preprocessing of lithium battery capacity curve based on local weighted regression

Due to the influence of human interference, abnormal stop and other factors during the test, the original data of the capacity curve of some batteries often have sudden changes and abnormal values, which will lead to deviations in the results. Therefore, it is necessary to smooth the original battery data.

The present invention uses a local weighted regression algorithm to smooth the capacity curve. The local weighted regression algorithm (LWR, Locally Weighted Regression) is a regression algorithm that improves the effect of ordinary regression algorithms, which is to perform polynomial weighted fitting on local observation data, And use the least squares method to estimate, and finally get the points that need to be fitted. The specific principle is:

For each point qi, determine a window range, on all q _k in the window, _k =1,2,...,n, the weight α _k (q _i ) can be obtained from the weight function, using the weighted The weighted least squares method of the value α _k (q _i ) performs a d-order polynomial fit (Equation 1) on q _i to obtain the fitted value p _i . Using α _k (q _i ) to obtain p _i is called locally weighted regression.

p _i =α ₀ (q _i )+α ₁ (q _i )q _i +...+α _m (q _i )q _i ^β +ε _i ,i=1,2,...,n

where α ₀ (q _i ), α ₁ (q _i ),...,α _m (q _i ) are parameters unknown to q _i , ε _i ,i=1,2,...,n are independent and identically distributed random Error term, β is a value given in advance.

LWR has the following advantages: adaptively filter out noise interference between local data and maintain the characteristics of the original signal; improve prediction accuracy by reducing noise interference, and effectively avoid over-fitting and under-fitting problems.

In order to ensure the representativeness and accuracy of the sample data and improve the prediction accuracy of the remaining life of the battery, the invention calculates four similarities of the capacity degradation curves of the target battery and other batteries with the same temperature and the same rate and different formulas based on a certain test data length (capacity curve shape, capacity degradation rate, lifetime distribution and Chebyshev distance similarity) are screened. The battery with the highest similarity to the target battery is selected from a large number of historical batteries as a final transferable sample for migration life prediction, and the specific steps are shown in S102-S105.

Step S102 is to screen out the first training data that is similar to the target sample data curve type from a plurality of training data by performing curve shape screening;

Specifically, the filtering of the first training data with a curve type similar to the target sample data from the plurality of training data includes:

The target sample data and multiple training data are graphed into target sample data curves and multiple training data curves respectively;

Divide the target sample data curve and multiple training data curves into three categories: straight line, concave curve and convex curve;

Screening is performed according to the types of straight lines, concave curves and convex curves, and the training data curves different from the target sample data curve types are excluded to obtain the first training data.

The present invention performs the first screening according to the curve type, excludes training curves different from the target curve type, and narrows the range of similarity measure efficiently. The type of original curve can be judged according to the statistical law of quadratic slope: the quadratic slope of a straight line is equal to 0, the quadratic slope of a concave curve is greater than 0, and the quadratic slope of a convex curve is generally less than 0.

S ₁ =f″(x)

Among them, S ₁ represents the first screening.

In step S103, through the capacity degradation rate similarity screening, the second training data that is similar to the capacity degradation trend of the target sample data is selected from the first training data;

Specifically, the filtering of the second training data that is similar to the capacity degradation trend of the target sample data from the first training data includes:

Calculate the rate of change of the capacity curve when the first training data degenerates from the initial state to the end of the test, and retain several first training data closest to the target sample data;

The reserved first training data is used as the second training data.

Among them, the calculation formula of the second screening is:

Where x ₀ , x _EOT represent the rated capacity value in the initial state and the capacity value when the test is stopped, Cycles _EOT represents the number of test cycles running to the test stop threshold EOT, and S ₁ represents the second screening.

Step S104 is to filter out the third training data that is similar to the life distribution of the target sample data from the second training data by screening the lifespan distribution similarity;

Specifically, the filtering of the third training data from the second training data that is similar to the life distribution of the target sample data includes:

Compare the lifetime distribution of the second training data by measuring the number of cycles when running to the trial stop threshold, and retain several second training data that are closest to the lifetime distribution of the target sample data;

The retained second training data that is closest to the lifetime distribution of the target sample data is used as the third training data.

Step S105 selects the fourth training data with the smallest distance metric from the target sample data from the third training data, and uses the fourth training data as the transferable sample data for cross-recipe similar product life prediction;

Specifically, selecting the fourth training data with the smallest distance metric from the target sample data from the third training data includes:

Select the Chebyshev distance to filter the capacity curve, and calculate the Chebyshev distance between the degradation curves of several third training data and the capacity degradation curve of the target sample data;

The third training data with the smallest Chebyshev distance is selected as the fourth training data.

The Chebyshev Distance is: if two points p and q have coordinates p _i and q _i , respectively, then the Chebyshev distance D _Chebyshev (p, q) between them is defined as their respective The maximum value of the coordinate value difference, in the following form:

The inventor found through a large number of experiments that the convergence of the similarity curves calculated based on the Chebyshev distance is better, and the screened curves have more similar degradation trends. Therefore, the present application combines the similarity measurement method based on the Chebyshev distance with the above-mentioned screening method, The prediction accuracy of deep learning and collaborative filtering is improved.

Step S106 uses the lifespan prediction model adapted to the transferable sample data to predict the lifespan migration of similar products across recipes

Wherein, the preprocessing includes normalizing the target sample data and the training data;

Described using the life prediction model to perform life prediction processing on the current test data of the products with similar formulations to be tested, and obtaining the remaining service life RUL of the similar products with similar formulations to be tested includes:

Input the current test data of the product with similar formula to be tested into the life prediction model, and output the RUL label value of the remaining service life of the product with similar formula to be tested;

The RUL label value of the remaining service life of the product with similar formula to be tested is reverse-normalized, and the predicted value of the remaining service life RUL of the product with similar formula to be tested is obtained.

Specifically, the similar product is a lithium battery.

Further, the similar product is a lithium-ion battery.

Application Example 1

The present invention uses the migration prediction of similar product life provided in the above-mentioned embodiment 1 for lithium-ion battery life prediction, and the method flow is shown in Figure 2. The prediction method flow can be divided into data preprocessing, migration sample selection, remaining life Prediction and experimental optimization in four parts. Specific steps include:

1) Data preprocessing: In order to ensure the consistency of data scales in cross-recipe prediction, the original data curve of the lithium battery capacity retention rate is first smoothed based on the method of local weighted regression, and the normalized preprocessing data is obtained. Determine the failure threshold of the battery in this study, select 80% of the initial capacity of the lithium battery as the end of life, and normalize its capacity data and cycle life data, the initial capacity value is 1, and when it runs to the failure threshold, the The capacity value is changed to 0, and the input data of the prediction model and the corresponding cycle life label are obtained;

2) Selection of transferable samples based on four screenings: In order to ensure the representativeness and accuracy of sample data and improve the accuracy of battery remaining life prediction, the target battery and the same temperature and same rate with different formulas are calculated based on a certain length of test data. Four similarities (capacity curve shape, capacity degradation rate, lifetime distribution and Chebyshev distance similarity) of other battery capacity degradation curves are screened. Select the battery with the highest similarity to the target battery from a large number of historical batteries, as the final transferable sample for migration life prediction, to solve the problem of what to migrate;

3) Remaining life prediction based on Tr-LSTM: First, the Tr-LSTM model is initialized by inheriting the pre-trained LSTM model structure and weight parameters through model migration; then, the Tr-LSTM is fine-tuned through the filtered transferable samples, and input to the target battery The short-term test data of the target battery is obtained, and the predicted life label of the target battery is obtained; finally, the RUL of the target battery is obtained by using the inverse normalization rule.

4) Cycle life test optimization: when the cycle life test of the target battery runs to a preset test stop threshold, the test is stopped. Using a small amount of data of the target battery obtained from the test, the predicted RUL value is obtained to replace the real value obtained by the cycle life experiment, so as to reduce the cycle life test time, improve the efficiency, and reduce the cost of lithium battery design and development.

1. Data preprocessing

Due to the differences in the capacity degradation curves of different batteries, in order to achieve a higher accuracy life migration prediction, the sample data is first preprocessed. mainly include:

1) Data normalization

82% of the initial capacity of the lithium battery is selected as the failure threshold, and the capacity value of the lithium battery is normalized to 1-0 (the initial capacity is 1, and 82% of the initial capacity is 0). At the same time, the RUL label represents the RUL value of the battery corresponding to each test cycle, which also needs to be normalized (the remaining life corresponding to the initial capacity is 1, and the remaining life corresponding to 82% of the initial capacity is 0). The process is shown below.

2) Smoothing preprocessing of lithium battery capacity curve based on local weighted regression

Due to the influence of human interference, abnormal stop and other factors during the test, the original data of the capacity curve of some batteries often have sudden changes and abnormal values, which will lead to deviations in the results. Therefore, it is necessary to smooth the original battery data.

This paper uses the local weighted regression algorithm to smooth the capacity curve. The local weighted regression algorithm (LWR, Locally Weighted Regression) is a regression algorithm that can improve the effect of ordinary regression algorithms. The local weighted regression algorithm was proposed by C1leveland. After the efforts of Cleveland and Develin, the algorithm was extended and applied to the case of multiple independent variables. fitted points. The specific principles are as follows:

For each point qi, determine a window range, on all q _k in the window, _k =1,2,...,n, the weight α _k (q _i ) can be obtained from the weight function, using the weighted The weighted least squares method of the value α _k (q _i ) performs a d-order polynomial fit (Equation 1) on q _i to obtain the fitted value p _i . Using α _k (q _i ) to obtain p _i is called locally weighted regression.

p _i =α ₀ (q _i )+α ₁ (q _i )q _i +...+α _m (q _i )q _i ^β +ε _i ,i=1,2,...,n

where α ₀ (q _i ), α ₁ (q _i ),...,α _m (q _i ) are parameters unknown to q _i , ε _i ,i=1,2,...,n are independent and identically distributed random Error term, β is a value given in advance.

LWR has the following advantages: adaptively filter out noise interference between local data and maintain the characteristics of the original signal; improve prediction accuracy by reducing noise interference, and effectively avoid over-fitting and under-fitting problems.

2. Transferable sample selection based on four screenings

The degree of similarity between the transferable samples and the target battery directly affects the prediction accuracy, so selecting the most similar samples from the historical sample library is of great significance to improve the prediction accuracy. The historical battery data of different formulations were optimized and screened four times based on the shape of the capacity curve, the capacity degradation rate, the life distribution and the similarity of the Chebyshev distance, and the transferable samples were obtained.

2.1 Primary Screening: Migration of Sample Selection Based on Capacity Curve Morphology

The capacity curves of batteries with different formulations show different degradation trends. According to the different curve shapes, the battery capacity curves are divided into three types: straight lines, concave curves and convex curves. The first screening is performed according to the curve type, and the training curves that are different from the target curve type are excluded, and the scope of the similarity measure is efficiently narrowed. The type of original curve can be judged according to the statistical law of quadratic slope: the quadratic slope of a straight line is equal to 0, the quadratic slope of a concave curve is greater than 0, and the quadratic slope of a convex curve is generally less than 0.

S ₁ =f″(x)

2.2. Secondary screening: transferable sample selection based on capacity degradation rate

Calculate the rate of change of the capacity curve of different batteries from the initial state degradation to the end of the test, and keep the 10 samples closest to the target battery.

Where x ₀ , x _EOT represent the rated capacity value of the initial state and the capacity value when the test is stopped, Cycles _EOT represents the number of test cycles that run to the test stop threshold EOT.

2.3. Tertiary Screening: Transferable Sample Selection Based on Lifetime Concentration

By measuring the number of cycles when running to the test stop threshold TS _EOT , the lifetime distributions of different lithium batteries were compared and the 5 candidate batteries closest to the target battery were retained.

2.4 Quaternary Screening: Transferable Sample Selection Based on Capacity Curve Distance Metrics

Select the Chebyshev distance to screen the capacity curve for the fourth time, calculate the Chebyshev distance between the historical battery and the target battery capacity degradation curve, and select the sample battery with the smallest Chebyshev distance (the highest similarity) as the final transferable sample .

Chebyshev Distance: If two points p and q have coordinates p _i and q _i respectively, then the Chebyshev distance D _Chebyshev (p,q) between the two points is defined as its coordinates The maximum value of the numerical difference, in the form:

Similar curves calculated based on Chebyshev distance have better convergence, and the selected curves have more similar degradation trends. Therefore, the similarity measurement method based on Chebyshev distance can be used to improve the prediction accuracy of deep learning and collaborative filtering.

3. RUL prediction method based on Tr-LSTM

The construction and testing process of the Tr-LSTM-based RUL transfer prediction model is shown in Figure 3, which includes three steps: model initialization, fine-tuning training, and RUL prediction.

The Tr-LSTM model is initialized with a strategy based on model transfer learning. First, the structural parameters of a pre-trained or historical LSTM model are inherited, and the Tr-LSTM model is initially constructed; then, the corresponding layers are initialized by reusing the weight parameters of the previous layers of the pre-trained model. By inheriting the empirical knowledge learned and extracted from previous models, the training efficiency and prediction accuracy of the Tr-LSTM model are improved; finally, the remaining Tr-LSTM layers are randomly initialized.

During the fine-tuning training phase, the filtered transferable samples and their RUL labels are used to fine-tune the weights of the Tr-LSTM model. By fine-tuning the training process, the nonlinear mapping relationship between the normalized capacity data and RUL labels can be more accurately established, the trend sensitivity of the prediction model can be improved, and better prediction performance can be obtained; finally, the short-term test data of the target battery can be obtained. Feed into the Tr-LSTM model to predict the RUL of the target battery.

3.1 LSTM model

The deep learning model has the ability to extract abstract features through multi-layer nonlinear mapping, and can abstract and extract features from the input signal layer by layer, so as to dig out deeper potential laws. Recurrent Neural Network (RNN) introduces the concept of timing in the design of network structure, considers the timing dependence of input and output variables, and uses multi-step time series as input variables. Each neuron in the network not only receives the previous layer. The neuron outputs and receives the output signal of the previous moment at the same time, forming a ring-shaped self-circulating connection structure in the hidden layer ^[2] , thus obtaining the "memory" ability, as shown in equations (1)-(2). x(t) represents the input of the time series signal at time t, h(t) represents the state of the hidden layer of the model at time t, y(t) represents the output at time t, and the RNN forward propagation process is shown in the figure

h(t)=tanh(Wh(t-1)+Ux(t)+b) (1)

y(t)=softmax(Vh(t)+c) (2)

Among them, b and c are the bias vectors of the input layer and the hidden layer, respectively, and U, W, and V are the weight matrices transferred between the input layer-hidden layer, the hidden layer-hidden layer, and the hidden layer-output layer, respectively. Then, the loss of the model at the current position can be quantified by the loss function L(t), and the gradient of the loss function can be calculated. The BPTT (back-propagation through time) algorithm is directly used in the back-propagation process of the network. Different time steps in the training process The cell units update the same weight parameters, which is called weight sharing.

When the standard RNN has a deep network layer, using the BPTT algorithm to update the weight will cause the problem of gradient disappearance or gradient explosion, that is, the gradient will gradually decay or gradually enlarge along each layer, and the amplitude will be transmitted to the first layer. Too large or too small to reasonably update the network weight parameters. The Long Short-Time Memory (LSTM) proposed by Hochreiter and Schmidhuber, as an improved recurrent neural network, can effectively solve the problem of gradient disappearance and gradient expansion of RNN algorithm in long-period signal processing. The influence of longer distance parameters on learning results has been widely used.

The lstm unit in the long short-term memory (LSTM) migration prediction model selected by the present invention introduces the ingenious idea of self-circulation. Through the gate control units such as forget gate, input gate, and output gate, the neuron can determine the self-cycle according to the context information. The weight of the cyclic update realizes the memory and forgetting mechanism of the lstm unit. The main hidden layer unit in the Lstm model is called "memory block", which contains one or more memory cell units, as shown in Figure 5.

Each memory cell unit obtains the input x(t) at time t, and at the same time chooses whether to add the state h(t-1) at time t-1, and the final output depends on the relationship between the two. In the deep network structure LSTM, features can flow and transfer from layer to layer along the network structure, realize multiple combinations of features, and obtain more advanced and high-quality feature expression, so it has a strong fitting ability.

In the first step, a forget gate (Gers, 2000, learning to forget) controls the weights for self-recurrent updates. The forget gate uses the sigmoid layer to map the input vector x(t) at time t and the output h(t-1) at time t-1 to a weight value between 0 and 1, 0 means completely discarded, 1 means completely reserved , as the weight f(t) of the state information update at the previous moment, as shown in formula (3), so as to realize the discarding of some information from the historical memory;

f(t)=sigmoid(W _f ·[h(t-1),x(t)]+b _f ) (3)

where W _f and b _f are the weight and bias term of the forget gate, respectively;

LSTM细胞单元的状态将更新为C(t)，包含由输入门控制的当前状态信息和由遗忘门控制的记忆状态信息，如式(6)所示；Similarly, the input gate is mapped in the same way, as the update weight i(t) of the current input information to decide to select part of the current input information to participate in the cell unit state update, as shown in formula (4); then through the tanh layer generation formula ( 5) The new candidate state vector represented by

The state of the LSTM cell unit will be updated to C(t), which contains the current state information controlled by the input gate and the memory state information controlled by the forget gate, as shown in equation (6);

i(t)=sigmoid(W _i ·[h(t-1),x(t)]+b _i ) (4)

Wi and _bi are the weights and _bias terms of the input gate, respectively;

W _c , b _c are the weights and bias terms of the tanh layer, respectively;

Finally, the output gate determines the output part of the information. The internal state information C(t) is mapped to between -1 and 1 through the tanh layer, and multiplied with the output vector o(t) of the output gate sigmoid layer (equation (7)) to obtain the current state output h(t ), as shown in formula (8).

o(t)=sigmoid(W _o ·[h(t-1),x(t)]+b _o ) (7)

W _o , b _o are the weights and bias terms of the output gate, respectively;

h(t)=o(t)*tanh(C(t)) (8)

3.2 Tr-LSTM model and transfer learning strategy

The network structure of the Tr-LSTM prediction model consisting of two LSTM layers and two fully connected layers is shown in Figure 6. The input and output of the model are the normalized capacity and the corresponding RUL label, respectively. The LSTM layer is used to extract the temporal features hidden in the volume sequence. Typically, the first few LSTM layers extract basic features of capacity. The higher the LSTM layer, the more abstract features are extracted and the more task-related. A fully connected layer is used to fit the mapping relationship between capacity features and RUL labels. During the model testing process, the RUL label value output from the test is denormalized to obtain the predicted value of the remaining service life.

This application proposes two hybrid transfer learning strategies to realize the RUL transfer prediction process based on the Tr-LSTM model. In terms of model layer migration, both strategies inherit the structural parameters of each layer of the pre-training model and transfer the historical experience learned from the pre-training model; in terms of feature layer migration, according to different migration strategies, the new models inherit and Freeze the weight parameters of the specified layer of the pre-trained model to speed up the training of the model. Finally, the weight parameters of the remaining layers are randomly initialized, and the transferable samples are input to fine-tune the remaining layers of the Tr-LSTM model.

For the battery at 25°C, the migration strategy 1 is adopted, and only the weight parameters of the first layer of the model are frozen, which helps to inherit the general features extracted by the pre-training model, and only adjust the second layer features of the model according to the particularity of the target battery.

For the battery at 60 °C, the migration strategy 2 is adopted. Due to the accelerated degradation caused by the increase in temperature, the amount of data is reduced, which is not enough to effectively train the model. Therefore, the weight parameters of the first two layers are frozen to improve the training efficiency of the model and help improve the prediction accuracy. Spend.

For the battery at 45°C, the lifetime distribution is between 25°C and 60°C. Therefore, migration strategies 1 and 2 are used to predict its RUL, respectively, and the arithmetic mean of the predicted results is used as the final predicted value of RUL.

Since the predicted RUL curve shows an approximate linear downward trend, we linearly fit the predicted curve by estimating the slope and intercept, which helps to improve the prediction accuracy, as shown in Figure 7.

4. Optimization of lithium battery cycle life test based on life prediction

Aiming at the cycle life test of lithium-ion batteries, this paper proposes an experimental optimization method based on remaining life prediction, which provides a guarantee for reducing the cycle life test time of lithium batteries and improving the design and development efficiency of lithium batteries.

4.1 Introduction of lithium battery cycle life experiment

(1) Test platform

The lithium battery cycle life experiments conducted in this application were conducted on a lithium ion battery cycle life test bench.

(2) Lithium-ion battery formula

In the present invention, enterprise test data is used to verify the feasibility and effectiveness of the proposed lithium-ion battery life prediction method (Note: the battery used in this experiment is different from the battery in the actual product, but a special soft pack in the product design stage. Battery). In this experiment, 10 groups of lithium batteries with different formulations of the same battery platform were selected, which are A, B,... , but battery anode materials and electrolyte solutions vary.

(3) Experimental temperature T

In this experiment, the standard experimental temperature conditions (25 °C) and the high temperature experimental conditions (45 °C and 60 °C) were considered, and the cycle life experiments were carried out on multiple batteries at each experimental temperature to obtain the cycle life of the corresponding lithium batteries.

(4) Charge and discharge process

The capacity of the lithium battery used in the experiment is about 2070mAh, and all experiments are carried out at the same charging rate, that is, the standard constant current/constant voltage method, which is a constant current rate of 1C until the voltage reaches 4.2V, and then maintains the voltage value of 4.2V until The charging current drops below 0.1A. The cyclic discharge process was carried out with a constant discharge current of 2A, and the discharge cut-off voltage was 2.78V.

(5) Failure threshold TS _Failure (end of life condition)

Due to the limitation of the experimental conditions and the influence of the characteristics of the lithium battery, the experiment was stopped when the end-of-life condition of the lithium battery was reached, that is, the battery capacity was degraded to 82% of the initial capacity. Therefore, 82% is considered as the failure threshold TS _Failure for the target and reference cells.

(6) Target battery experiment termination threshold TS _obj

In order to fully reduce the experimental error and obtain enough data to ensure the validity of the proposed prediction method, it is necessary to optimize the termination threshold TS _obj of the target battery test data. In the experiment, 90% of the initial capacity is selected as the termination threshold of the target battery test data. TS _obj , predicts the remaining life of the lithium battery through experimental measurement data.

4.2 Optimization of experimental process design

The detailed experimental process after optimization is shown in Figure 8. The main steps are as follows:

a. The target battery cycle life experiment is carried out at the temperature T _obj , and the experiment is stopped when the target battery experiment termination threshold TS _obj is reached. The target battery capacity degradation data obtained from experiments are normalized by data preprocessing.

b. Select a lithium battery from the historical database that belongs to the same battery platform as the target battery and obtains complete capacity degradation data through cycle life experiments at temperature T _obj , and normalizes its capacity degradation data and RUL label.

c. Through the four screening process of capacity curve shape, capacity curve distance, future trend of capacity curve and battery life concentration, select the sample data most similar to the target battery as the reference battery to predict the remaining life of the target battery.

d. Standardize the capacity degradation data and RUL label of the selected reference battery, input it into the LSTM deep learning model, and train the model.

e. When the capacity of the target battery reaches the battery cycle life experiment termination threshold TS _obj , input some known capacity data of the target battery into the trained LSTM model, and then the predicted RUL value of the target battery can be obtained. The predicted cycle life value can replace the real life value obtained from the cycle life experiment, thereby reducing the cycle life experiment time, improving the experiment efficiency, and reducing the cost of lithium battery design and development.

Test Example 1

1. Test and data introduction

In the present invention, the test data of a battery company is used to verify the feasibility and effectiveness of the proposed lithium-ion battery life prediction method (Note: the battery used in the test is a soft-pack battery specially used in the product design stage, which is consistent with the company's actual The battery used in the product varies). The selected test database contains 10 kinds of lithium-ion batteries with different formulations, which are A, B, ..., I, J groups. The tests were carried out at 25°C, 45°C and 60°C at the same time, and three temperature conditions were obtained. battery degradation data under.

2. Data preprocessing results

In the present invention, the battery capacity degradation data is used as the performance index reflecting the system degradation. Through the preliminary analysis of the data, the failure threshold is set as 0.82, that is, when the capacity degrades to 82% of the initial capacity, it is considered that the battery has reached the end of life. .

According to the data normalization standard, the battery capacity degradation data is normalized to 1-0 (the initial capacity is 1, and 82% of the initial capacity is 0), and the corresponding remaining life is also normalized to 1-0 (the initial capacity is 0). The remaining life corresponding to the capacity is 1, and the remaining life corresponding to 82% of the initial capacity is 0). The normalized original data curve of battery capacity degradation has large jump fluctuations, which is caused by the complex chemical reaction process during the discharge process of lithium batteries. In order to facilitate data analysis, the method of local weighted regression is used to analyze the original data. for smoothing. The effect diagram of the smoothing preprocessing of the lithium battery capacity based on the local weighted regression method is shown in Figure 9.

After smoothing the lithium battery capacity curve based on the local weighted regression method, after screening, a total of 147 different batteries were selected for research in this study, as shown in Table 1:

Table 1 Battery Screening Results

3. Transferable sample selection results

The research of the present invention is limited to the migration prediction of battery life under the same temperature and the same rate and different formulations, and different experimental termination thresholds are set, such as TS _failure = 90%, which means that the test is stopped when the capacity degrades to 90% of the initial capacity. Through the four-time screening method, compare the similarity between the target battery capacity degradation curve and the curve of other sample batteries (different formulations) at the same temperature and the same rate when the capacity degrades to 90% of the length, and select the optimal sample battery as the transferable sample to predict the target battery. The remaining cycle life is improved, and the accuracy of life prediction is improved. The schematic diagram of the migration sample screening results is shown in Figure 10.

4. Lifetime migration prediction based on LSTM model

1) Parameter settings of the LSTM model

Combined with the data characteristics of the battery, taking into account the optimization of computing resources and time, the LSTM model structure used includes a bottom data input layer, two hidden layers and a top data prediction layer, and the number of neurons in each layer is 100. ,50,50,1, the activation transfer function is a sigmoid function, the learning rate is 0.3, the noise reduction occlusion ratio is 0.15, and the optimizer is adam. The size of each sub-input block of deep learning is 100. In order to ensure the adequacy of feature self-learning, the cyclic execution steps of unsupervised learning and back-propagation process are epochs=8. See Table 3 for details. In order to avoid the influence of a single data anomaly on the predicted label, the input of the deep learning model is based on the data interval, and the one-dimensional data is adjusted into a data column in the form of a sliding window and input to the LSTM model (the number of rows in each column is the input layer. number of neurons).

2) Lifetime prediction results of the LSTM model

The target battery and the corresponding reference battery obtained through four screenings are used for life prediction, still taking 25 °C as an example, and the prediction results are shown in Table 4:

Table 2 Prediction results of battery life at 25℃

Table 3 Prediction results of battery life at 45℃

Table 4 Prediction results of battery life at 60℃

According to the results in Table 2, it can be seen that the accuracy of life prediction using the migrated samples provided by the present invention can reach up to 99.9%, which saves the cost of analyzing the performance of battery formulations. The data similarity analysis of batteries with different formulations can save a lot of test costs. , to achieve effective data sharing. It brings a new option for battery cycle life test, with good economy and practicability.

Although the present invention has been described in detail above, the present invention is not limited thereto, and various modifications can be made by those skilled in the art in accordance with the principles of the present invention. Therefore, all modifications made in accordance with the principles of the present invention should be understood as falling within the protection scope of the present invention.

Claims (10)

1. a kind of cross-formulation similar product life migration prediction method, is characterized in that, comprises: Preprocess the short-term cycle life test data of products with similar formula to be tested to obtain target sample data, and preprocess the full-life test capacity data of other products with similar formula to obtain multiple training data; By performing curve shape screening, the first training data that is similar to the curve type of the target sample data is screened from the plurality of training data; Through the capacity degradation rate similarity screening, the second training data that is similar to the capacity degradation trend of the target sample data is screened from the first training data; Through the similarity screening of the lifespan distribution, the third training data with a lifespan distribution similar to the target sample data is selected from the second training data; Through the minimum distance metric screening, the fourth training data with the smallest distance metric from the target sample data is selected from the third training data; Use the fourth training data as transferable sample data for life prediction of similar products across recipes; Use a lifetime prediction model adapted to the transferable sample data to perform lifetime migration predictions for similar products across formulations. 2. The method of claim 1, wherein using a life prediction model adapted to the transferable sample data to predict the lifespan migration of similar products across recipes comprises: Use the transferable sample data to fine-tune and train models previously trained for other products with similar formulations to obtain the lifespan prediction model; Using the life prediction model to perform life prediction processing on the current test data of the products with similar formula to be tested, to obtain the remaining service life RUL of the products with similar formula to be tested. 3. The method of claim 2, wherein the preprocessing comprises normalizing the target sample data and the training data; Described using the life prediction model to perform life prediction processing on the current test data of the products with similar formulations to be tested, and obtaining the remaining service life RUL of the similar products with similar formulations to be tested includes: Input the current test data of the product with similar formula to be tested into the life prediction model, and output the RUL label value of the remaining service life of the product with similar formula to be tested; The RUL label value of the remaining service life of the product with a similar formula to be tested is reverse-normalized, and the predicted value of the remaining service life RUL of the product with a similar formula to be tested is obtained. 4. The method according to claim 1, wherein screening out the first training data similar to the target sample data curve type from the plurality of training data comprises: The target sample data and multiple training data are graphed into target sample data curves and multiple training data curves respectively; Divide the target sample data curve and multiple training data curves into three categories: straight line, concave curve and convex curve; Screening is performed according to the types of straight lines, concave curves and convex curves, and the training data curves different from the target sample data curve types are excluded to obtain the first training data. 5. The method according to claim 4, wherein screening out the second training data similar to the capacity degradation trend of the target sample data from the first training data comprises: Calculate the rate of change of the capacity curve when the first training data degenerates from the initial state to the end of the test, and retain several first training data closest to the target sample data; The reserved first training data is used as the second training data. 6. The method of claim 5, wherein filtering out the third training data that is similar to the target sample data lifespan distribution from the second training data comprises: Compare the lifetime distribution of the second training data by measuring the number of cycles when running to the trial stop threshold, and retain several second training data that are closest to the lifetime distribution of the target sample data; The retained second training data that is closest to the lifetime distribution of the target sample data is used as the third training data. 7. The method according to claim 6, wherein screening out the fourth training data with the smallest distance metric from the target sample data from the third training data comprises: Select the Chebyshev distance to filter the capacity curve, and calculate the Chebyshev distance between the degradation curves of several third training data and the capacity degradation curve of the target sample data; The third training data with the smallest Chebyshev distance is selected as the fourth training data. 8. The method of claim 3, wherein the preprocessing further comprises: Eliminate unstable initial data and data that do not reflect the degradation trend in the short-term test sample data of products with similar formula to be tested and the full-life test capacity database of other products with similar formula; Smoothing the removed short-term test sample data and the data in the full-life test capacity database to obtain target sample data and prediction training data. 9. The method of any one of claims 1-8, wherein the similar product is a lithium battery. 10. A life migration prediction system for similar products across recipes, comprising a processor, a memory, a program stored on the memory and executable on the processor, and a program for implementing the processor and the processor. A data bus is connected and communicated between the memories, and when the program is executed by the processor, the method for predicting the lifespan migration of similar products across recipes according to any one of the above claims 1-9 is implemented.

Images