CN112036084B

CN112036084B - Similar product life migration screening method and system

Info

Publication number: CN112036084B
Application number: CN202010889701.0A
Authority: CN
Inventors: 马剑; 尚芃超; 邹新宇; 丁宇; 吕琛
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2020-08-28
Filing date: 2020-08-28
Publication date: 2022-08-02
Anticipated expiration: 2040-08-28
Also published as: CN112036084A

Abstract

The invention discloses a similar product life migration prediction method and a similar product life migration prediction system, which relate to the technical field of similar product migration learning and comprise the steps of preprocessing short-term cycle life test data of a formula similar product to be tested and life test data of other formula batteries to obtain target sample data and a plurality of training data; obtaining migratable sample data for predicting the service life of the cross-formula similar product by performing minimum screening on curve morphology, capacity degradation rate similarity, service life distribution similarity and distance measurement, and performing service life migration prediction on the cross-formula similar product by using a service life prediction model suitable for the migratable sample data to obtain a service life prediction result; the invention realizes the accurate prediction of the residual life of the lithium ion battery across the formula, the highest prediction accuracy can reach 99.9 percent, the test time and the cost in the design and development process of the lithium ion battery can be effectively saved, and the invention has considerable economic benefit and application value.

Description

Similar product life migration screening method and system

Technical Field

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

Background

With the development of energy storage technology and energy industry, lithium ion batteries are widely used as main energy storage devices of military electronic products, avionic devices, electric vehicles and various portable electronic devices (such as notebook computers, digital cameras, tablet computers, mobile phones and the like) due to the advantages of light weight, low discharge rate, long service life and the like. Cycle life is an important design property for lithium battery products. In the design and development process of the lithium battery, in order to accurately obtain the service life conditions of batteries with different formulas in a design matrix and provide feedback for formula selection and design optimization, a cycle life test needs to be carried out on the batteries with various formulas in the design matrix. This test is continued until the battery capacity retention rate reaches a prescribed threshold, i.e., the end-of-life point of the lithium battery. However, the number of lithium battery formulas in the design matrix is large, so that the existing cycle life testing time and capital cost are too high, and particularly for power lithium batteries with the life cycle of more than a year, the design and development efficiency is too low, and enterprises are difficult to bear.

Lithium battery life prediction is generally used in the use stage of a lithium battery, and the residual life of the battery at the current moment is predicted mainly according to a small amount of known historical data. As the number of test cycles of the predicted lithium battery is as small as possible, enough battery test data volume is difficult to obtain to meet the design and development requirements of the residual life prediction model. Therefore, by using other formula mass historical cycle life test data of the same battery platform of a battery enterprise, data support can be provided for a residual life prediction model required by design and development. However, how to define and measure data migratability and to design a screening strategy for migratable samples, and to obtain the most similar samples from other large quantities of differentiated formula battery data, how to use the most similar samples has great significance and application requirements for life prediction of similar products.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a transferable sample screening method and a transferable sample screening system which are more suitable for predicting the service life of batteries among different formulas in the battery design and development process by utilizing short-term actual measurement data and a transfer learning prediction method.

To achieve the technical object of the present invention, in one aspect, the present invention provides a method for predicting life migration of a similar product, including:

preprocessing short-term cycle life test data of similar products of a formula to be tested to obtain target sample data, and preprocessing full life test capacity data of similar products of other formulas to obtain a plurality of training data;

screening first training data similar to the curve type of the target sample data from a plurality of training data by screening curve forms;

screening out second training data similar to the capacity degradation trend of the target sample data from the first training data through capacity degradation rate similarity screening;

screening third training data similar to the life distribution of the target sample data from the second training data through life distribution similarity screening;

screening fourth training data with the minimum distance measurement from the target sample data from the third training data through the minimum distance measurement screening;

taking the fourth training data as migratable sample data of the cross-formula similar product life prediction;

and performing life migration prediction on the similar products across the formula by utilizing a life prediction model adaptive to the migratable sample data.

Wherein performing life migration prediction on the cross-recipe similar product using a life prediction model adapted to the migratable sample data comprises:

performing fine tuning training on a model which is trained aiming at other formula similar products in advance by using the migratable sample data to obtain the life prediction model;

and performing life prediction processing on the current test data of the similar product of the formula to be tested by using the life prediction model to obtain the residual service life RUL of the similar product of the formula to be tested.

The preprocessing comprises the step of carrying out normalization processing on target sample data and training data;

the step of performing life prediction processing on the current test data of the similar product of the formula to be tested by using the life prediction model to obtain the remaining service life RUL of the similar product of the formula to be tested comprises the following steps:

inputting the current test data of the similar product of the formula to be tested into the life prediction model, and outputting the residual service life RUL label value of the similar product of the formula to be tested;

and performing reverse normalization processing on the residual service life RUL label value of the similar product of the formula to be detected to obtain a residual service life RUL predicted value of the similar product of the formula to be detected.

The method for screening out the first training data similar to the curve type of the target sample data from the plurality of training data comprises the following steps:

respectively imaging the target sample data and the plurality of training data into a target sample data curve and a plurality of training data curves;

dividing a target sample data curve and a plurality of training data curves into three types, namely a straight line, a concave curve and a convex curve;

and screening according to the straight line type, the concave curve type and the convex curve type, and eliminating a training data curve different from the target sample data curve type to obtain first training data.

Screening out second training data similar to the capacity degradation trend of the target sample data from the first training data comprises the following steps:

calculating the change rate of a capacity curve when the first training data is degraded from an initial state to the end of the test, and reserving a plurality of first training data which are closest to target sample data;

and using the retained first training data as second training data.

Screening out third training data similar to the life distribution of the target sample data from the second training data comprises the following steps:

comparing the life distribution of the second training data by measuring the number of cycles when the test is run to the test stop threshold, and reserving a plurality of second training data which are closest to the life distribution of the target sample data;

and using the second training data which is kept to be closest to the life distribution of the target sample data as third training data.

Wherein, the step of screening out fourth training data with the minimum distance metric with the target sample data from the third training data comprises the following steps:

selecting the Chebyshev distance to screen the capacity curve, and calculating the Chebyshev distance between the degradation curves of the 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 pre-processing further comprises:

removing short-term test sample data of the similar product of the formula to be tested and unstable initial data and data which do not show a degeneration trend in a full-life test capacity database of the similar product of other formulas;

and 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 object of the present invention, in another aspect, the present invention provides a cross-recipe similar product life migration prediction system, which includes a processor, a memory, a program stored on the memory and operable on the processor, and a data bus for implementing connection communication between the processor and the memory, wherein the program, when executed by the processor, implements a cross-recipe similar product life migration prediction method.

Has the advantages that:

due to the fact that the short test data of the predicted battery is difficult to effectively highlight the capacity degradation rule of the battery, the service life prediction model is difficult to effectively train and give an accurate prediction result. Aiming at the problem, the invention provides a concept based on transfer learning, a transferable sample screening method adopting four-time screening is adopted, data with the highest similarity to the predicted battery capacity degradation rule is obtained from historical life test data of batteries with different formulas, and the data is transferred and applied to training of a predicted battery life prediction model, so that accurate prediction of the residual life of the lithium power battery across formulas is realized, the prediction accuracy can reach 99.9% to the maximum, the test time and cost in the design and development process of the lithium battery can be effectively saved, and the method has considerable economic benefit and application value.

Drawings

FIG. 1 is a flowchart of a method for predicting life migration of similar products according to embodiment 1 of the present invention;

fig. 2 is a flow of a lithium ion battery life migration prediction method provided in application example 1 of the present invention;

FIG. 3 is a diagram illustrating a process of constructing and testing a Tr-LSTM-based RUL branch prediction model according to example 1;

FIG. 4 is a schematic diagram of an RNN structure provided in application example 1 of the present invention;

FIG. 5 is a schematic diagram of the LSTM unit structure provided in application example 1 of the present invention;

FIG. 6 is a diagram illustrating a Tr-LSTM-based RUL prediction model and its branch learning strategy provided in example 1;

FIG. 7 is a linear fit of a prediction curve by slope and intercept provided by application example 1 of the present invention;

FIG. 8 is a flow of experimental process design optimization provided in application example 1 of the present invention;

fig. 9 is a graph showing the result of preprocessing of data on the degradation data of the battery under the three temperature conditions of 25 ℃, 45 ℃ and 60 ℃ provided in experimental example 1, in which fig. 9(a) is a data curve at 25 ℃, fig. 9(b) is a data curve at 45 ℃, and fig. 9(c) is a data curve at 60 ℃;

FIG. 10 is a graph showing the results of screening samples that can migrate under the three temperature conditions of 25 ℃, 45 ℃ and 60 ℃ as provided in test example 1.

Detailed Description

The method and system of the present invention will now be described in more detail, with reference to the schematic drawings in which preferred embodiments of the invention are shown, it being understood that one skilled in the art could modify the invention herein described while still achieving the advantageous results of the invention. Accordingly, the following description should be construed as broadly as possible to those skilled in the art and not as limiting the invention.

The invention is more particularly described in the following paragraphs with reference to the accompanying drawings by way of example. Advantages and features of the present invention will become apparent from the following description and from the claims. It is to be noted that the drawings are in a very simplified form and are not to precise scale, which is merely for the purpose of facilitating and distinctly claiming the embodiments of the present invention.

The experimental procedures used in the following examples are all conventional procedures unless otherwise specified. The structures, materials, and the like used in the following examples are commercially available unless otherwise specified.

Example 1

As shown in fig. 1, the method for screening transferable samples of similar products provided by the present invention comprises:

step S101, preprocessing short-term cycle life test data of a similar product of a formula to be tested to obtain target sample data, and preprocessing full-life test capacity data of other formula batteries to obtain a plurality of training data;

specifically, the pretreatment comprises the following steps:

removing short-term test sample data of the formula battery to be tested and unstable initial data and data which do not show a degeneration trend in the database of the total life test capacity of other formula batteries;

Further, before the smoothing, the method further includes:

and normalizing the removed short-term test sample data and the data in the life test capacity database.

In order to ensure the consistency of data scale during the cross-formula prediction, the invention firstly carries out smoothing treatment on the original data curve of the lithium battery capacity retention rate based on a local weighted regression method to obtain the normalized preprocessing data. The method selects 80% of the initial capacity of the lithium battery as a service life end point (failure threshold), and carries out normalization processing on the capacity data and the cycle life data, wherein the initial capacity is 1, and the capacity when the lithium battery runs to the failure threshold is 0, so as to obtain input data of a prediction model and a corresponding cycle life label, and the specific calculation steps are as follows:

1) data normalization

And selecting 82% of the initial capacity of the lithium battery as a failure threshold value, and normalizing the capacity value of the lithium battery to be 1-0 (the initial capacity is 1, and 82% of the initial capacity is 0). Meanwhile, the RUL label represents the RUL value of the battery corresponding to each test cycle, and it is also required to perform normalization processing (the remaining life corresponding to the initial capacity is 1, and the remaining life corresponding to 82% of the initial capacity is 0), and the normalization process is as follows.

2) Lithium battery capacity curve smoothing pretreatment based on local weighted regression

Due to the influences of factors such as artificial interference and abnormal stop in the test process, sudden change and abnormal values often exist in the capacity curve raw data of part of batteries, and the result is deviated, so that the battery raw data needs to be subjected to smooth preprocessing.

The invention uses a local Weighted Regression algorithm to carry out smoothing processing on a capacity curve, the local Weighted Regression algorithm (LWR, Localy Weighted Regression) is a Regression algorithm for realizing effect improvement on a common Regression algorithm, namely polynomial Weighted fitting is carried out on local observation data, and a least square method is used for carrying out estimation, so that points needing fitting are finally obtained, and the specific principle is as follows:

for each point q _i Determining a window range within which all q's are within _k Where k is 1,2, …, n, the weight α is obtained from the weight function _k (q _i ) Using a weight α _k (q _i ) Weighted least squares of (Q) _i D-order polynomial fitting (formula 1) is carried out to obtain a fitting value p _i . Using alpha _k (q _i ) To obtain p _i It is called local 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

Wherein alpha is ₀ (q _i ),α ₁ (q _i ),…,α _m (q _i ) Is relative to q _i Unknown parameter, ∈ _i I is 1,2, …, n is an independent and equally distributed random error term, and β is a predetermined value.

LWR has the following advantages: adaptively filtering noise interference between local data and keeping the characteristics of original signals; the prediction precision is improved by reducing noise interference, and the problems of over-fitting and under-fitting are effectively avoided.

In order to ensure the representativeness and the accuracy of sample data and improve the prediction accuracy of the residual life of the battery, the invention calculates four similarities (capacity curve morphology, capacity degradation rate, life distribution and Chebyshev distance similarity) of capacity degradation curves of a target battery and other batteries with the same temperature, the same multiplying power and different formulas by taking a certain test data length as a reference for screening. The battery with the highest similarity to the target battery is selected from the large number of historical batteries to serve as a migratable sample for migration life prediction, and specific steps are shown as S102-S105.

S102, screening first training data similar to the curve type of target sample data from a plurality of training data by screening curve forms;

specifically, the screening out first training data similar to the curve type of the target sample data from the plurality of training data includes:

The method carries out the first screening according to the curve type, excludes the training curve with the type different from that of the target curve, and efficiently reduces the range of similarity measurement. The type of the original curve can be judged according to the statistic rule of the secondary slope: the linear quadratic slope is equal to 0, the concave curve quadratic slope is greater than 0, and the convex curve quadratic slope is generally less than 0.

S ₁ ＝f″(x)

Wherein S is ₁ Represents the first screening.

Step S103, screening out second training data similar to the capacity degradation trend of the target sample data from the first training data through capacity degradation rate similarity screening;

specifically, the screening out, from the first training data, second training data similar to the capacity degradation trend of the target sample data includes:

and using the retained first training data as second training data.

Wherein, the calculation formula of the second screening is as follows:

wherein x ₀ ，x _EOT Cycles, nominal capacity value indicating initial condition and capacity value at test stop _EOT Denotes the number of test cycles, S, run to the test stop threshold EOT ₁ Representing a second screening.

Step S104, screening third training data similar to the life distribution of the target sample data from the second training data through life distribution similarity screening;

specifically, screening out third training data similar to the life distribution of the target sample data from the second training data includes:

comparing the life distributions of the second training data by measuring the number of cycles when the test is run to the test stop threshold, and reserving a plurality of second training data which are closest to the life distribution of the target sample data;

Step S105, screening fourth training data with the minimum distance measurement with the target sample data from the third training data through the minimum distance measurement screening, and taking the fourth training data as migratable sample data for predicting the service life of the cross-formula similar product;

specifically, the step of screening out fourth training data with the minimum distance metric from the target sample data from the third training data includes:

The Chebyshev Distance (Chebyshev Distance) is: if two points p and q have coordinates p _i And q is _i The Chebyshev distance D between the two _Chebyshev (p, q), defined as the maximum of their respective coordinate value differences, of the form:

the inventor finds that similar curves calculated based on the Chebyshev distance are better in convergence and the screened curves are more similar in degradation trend through a large number of experiments, so that the similarity measurement method based on the Chebyshev distance is combined with the screening method, and the prediction accuracy of deep learning and collaborative filtering is improved.

Step S106, life migration prediction is carried out on the similar products across the formula by using a life prediction model suitable for the migratable sample data

carrying out reverse normalization processing on the residual service life RUL label value of the similar product of the formula to be detected to obtain the residual service life RUL predicted value of the similar product of the formula to be detected

In particular, the similar product is a lithium battery.

Further, the similar product is a lithium ion battery.

Application example 1

The method for predicting the service life of the lithium ion battery by using the similar product service life migratable prediction provided by the embodiment 1 is shown in fig. 2, and the flow of the prediction method can be divided into four parts, namely data preprocessing, migratable sample selection, residual service life prediction and experimental optimization. The method comprises the following specific steps:

1) data preprocessing: in order to ensure the consistency of data scale during the cross-formula prediction, firstly, the original data curve of the lithium battery capacity retention rate is subjected to smoothing treatment based on a local weighted regression method to obtain normalized preprocessing data. Determining a failure threshold of the battery in the research, selecting 80% of the initial capacity of the lithium battery as a service life end point, and performing normalization processing on capacity data and cycle life data of the lithium battery, wherein the initial capacity is 1, and the capacity when the lithium battery is operated to the failure threshold is 0, so as to obtain input data of a prediction model and a corresponding cycle life label;

2) migratable sample selection based on four screenings: in order to ensure the representativeness and the accuracy of sample data and improve the prediction accuracy of the residual life of the battery, four similarities (capacity curve morphology, capacity degradation rate, life distribution and Chebyshev distance similarity) of capacity degradation curves of a target battery and other batteries with the same temperature, the same multiplying power and different formulas are calculated by taking a certain test data length as a reference for screening. Selecting a battery with the highest similarity to a target battery from a large number of historical batteries, and using the battery as a migratable sample for migration life prediction to solve the problem of what is migrated;

3) Tr-LSTM-based prediction of remaining life: firstly, inheriting a pre-trained LSTM model structure and weight parameters through model migration, and initializing a Tr-LSTM model; then, inputting short-term test data of the target battery through the screened migratable sample fine tuning Tr-LSTM to obtain a predicted life label of the target battery; and finally, obtaining the RUL of the target battery by utilizing an inverse normalization rule.

4) Optimizing the cycle life test: and stopping the test when the cycle life test of the target battery runs to a preset test stop threshold value. By using a small amount of data of the target battery obtained by testing, the RUL predicted value is obtained to replace a true value obtained by a cycle life experiment, so that the cycle life experiment time is reduced, the efficiency is improved, and the design and development cost of the lithium battery is reduced.

1. Data pre-processing

Due to the difference of the degradation curves of different battery capacities, in order to realize the life migration prediction with higher accuracy, sample data is preprocessed firstly. The method mainly comprises the following steps:

1) data normalization

The capacity curve is smoothed by using a local Weighted Regression algorithm (LWR), which is a Regression algorithm for improving the effect of a general Regression algorithm. The local weighted regression algorithm is proposed by C1eveland, and the algorithm is popularized and applied to the conditions of a plurality of independent variables through the efforts of Cleveland and Develin, and the principle is that polynomial weighted fitting is carried out on local observation data, the estimation is carried out by using a least square method, and finally, points needing to be fitted are obtained. The specific principle is as follows:

for each point q _i Determining a window range within which all q's are within _k Where k is 1,2, …, n, the weight α is obtained from the weight function _k (q _i ) Using a weight α _k (q _i ) Weighted least squares of _i D-order polynomial fitting (formula 1) is carried out to obtain a fitting value p _i . Using alpha _k (q _i ) To obtain p _i It is called local weighted regression.

2. Migratable sample selection based on four screenings

The similarity degree between the transferable sample and the target battery directly influences the prediction precision, so that the selection of the most similar sample from the historical sample library has important significance for improving the prediction accuracy. And performing four times of optimized screening on historical battery data of different formulas based on capacity curve forms, capacity degradation rates, service life distribution and Chebyshev distance similarity to obtain a transferable sample.

2.1 primary screening: migratable sample selection based on capacity curve morphology

The capacity curves of batteries with different formulas show different degradation trends, and the battery capacity curves are divided into a straight line, a concave curve and a convex curve according to different curve forms. And performing first screening according to the curve type, eliminating a training curve with a different type from the target curve type, and efficiently reducing the range of similarity measurement. The type of the original curve can be judged according to the statistic rule of the secondary slope: the linear quadratic slope is equal to 0, the concave curve quadratic slope is greater than 0, and the convex curve quadratic slope is generally less than 0.

S ₁ ＝f″(x)

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

The rate of change of the capacity curve from the initial state degradation to the end of the test was calculated for the different cells, leaving the 10 samples closest to the target cell.

Wherein x ₀ ，x _EOT Cycles, nominal capacity value indicating initial condition and capacity value at test stop _EOT Representing the number of test cycles run to the test stop threshold EOT.

2.3. And (3) screening for three times: migratable sample selection based on lifetime concentration

By measuring the movementGo to test stop threshold TS _EOT And comparing the service life distribution of different lithium batteries according to the cycle number, and reserving 5 candidate batteries closest to the target battery.

2.4 four screenings: migratable sample selection based on capacity curve distance metric

And selecting the Chebyshev distance to screen the capacity curve for the fourth time, calculating the Chebyshev distance between the historical battery and the target battery capacity degradation curve, and selecting the sample battery with the minimum Chebyshev distance (highest similarity) as the final transferable sample.

Chebyshev Distance (Chebyshev Distance): if two points p and q have coordinates p _i And q is _i Then the Chebyshev distance D between the two _Chebyshev (p, q), defined as the maximum of their respective coordinate value differences, of the form:

similar curves calculated based on the Chebyshev distance are better in convergence, and the screened curves are more similar in degradation trend, so that the prediction accuracy of deep learning and collaborative filtering can be improved by using a similarity measurement method based on the Chebyshev distance.

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 FIG. 3, and comprises three steps: model initialization, fine tuning training, and RUL prediction.

The Tr-LSTM model is initialized based on a strategy of model transfer learning. Firstly, inheriting a structural parameter of a pre-trained or historical LSTM model, and preliminarily constructing a Tr-LSTM model; then, the corresponding layer is initialized by reusing the weight parameters of the previous layers of the pre-training model. The training efficiency and the prediction accuracy of the Tr-LSTM model are improved by inheriting the experience knowledge of the previous model learning and extraction; finally, the remaining Tr-LSTM layers are randomly initialized.

In the fine training stage, the weights of the Tr-LSTM model are fine-tuned by using the migratable samples obtained by screening and the RUL labels thereof. Through fine tuning of the training process, the nonlinear mapping relation between the normalized capacity data and the RUL label can be established more accurately, the trend sensitivity of the prediction model is improved, and better prediction performance is obtained; and finally, sending the short-term test data of the target battery into a Tr-LSTM model to predict the RUL of the target battery.

3.1LSTM model

The deep learning model has the capability of extracting abstract features through multilayer nonlinear mapping, can abstract input signals layer by layer and extract features, and excavates deeper potential rules. A cyclic Neural Network (RNN) introduces a time sequence concept in the design of a Network structure, considers the time sequence dependency relationship of input and output variables, takes a multi-step time sequence as the input variable, each neuron in the Network not only receives the output of a previous layer of neurons, but also receives the output signal of the previous moment, and a ring-shaped self-cyclic connection structure is formed in a hidden layer ^[2] Thus, the memory capability is obtained, as shown in formulas (1) to (2). x (t) represents the input of a time sequence signal at the time t, h (t) represents the state of a hidden layer of a model at the time t, y (t) represents the output at the time t, and the RNN forward propagation process is shown as

h(t)＝tanh(Wh(t-1)+Ux(t)+b) (1)

y(t)＝softmax(Vh(t)+c) (2)

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

When the number of network layers is deep, the problem of gradient disappearance or gradient explosion occurs when the BPTT algorithm is adopted for weight updating of the standard RNN, namely, the gradient is gradually attenuated or gradually amplified along each layer, and the amplitude is particularly large or small when the gradient is transmitted to the initial layer, so that the network weight parameters cannot be updated reasonably. The Long Short-Time Memory network (LSTM) proposed by Hochreiter and Schmidhuber is used as an improved recurrent neural network, can effectively solve the problems of gradient disappearance and gradient expansion of an RNN algorithm in Long-period signal processing, maintains the influence of a longer-distance parameter on a learning result, and is widely applied.

The LSTM unit in the long-short term memory (LSTM) migration prediction model introduces a self-circulation ingenious concept, and the neuron can determine self-circulation updating weight according to context information through the forgetting gate, the input gate, the output gate and other gate control units, so that a memory and forgetting mechanism of the LSTM unit is realized. The main hidden layer unit in the Lstm model is called a "memory block" and comprises one or more memory cell units, as shown in FIG. 5.

Each memory cell unit obtains input x (t) at time t, simultaneously selects whether state h (t-1) at time t-1 is added or not, and finally outputs depend on the relation of the two. In the depth network structure LSTM, the features can flow and transfer between layers along the network structure, so that multiple combinations of the features are realized, a higher-level and high-quality feature expression mode is obtained, and the depth network structure LSTM has strong fitting capability.

The first step is to control the weight of the self-loop update by a forgetting gate (Gers, 2000, learning to get). The forgetting gate utilizes a sigmoid layer to map an input vector x (t) at the moment t and an output h (t-1) at the moment t to a weight value between 0 and 1, wherein 0 represents complete discarding and 1 represents complete reservation as a weight f (t) for updating state information at the previous moment, as shown in a formula (3), so that partial information is discarded from history memory;

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

in the formula W _f ，b _f Respectively, the weight of the forgetting gate and the bias term;

similarly, the input gates are mapped in the same way as the update weights i (t) of the current input information to determine to select a part of the current input information to participate in the cell state update, as shown in equation (4)) Shown; then, a new candidate state vector represented by formula (5) is generated by the tanh layer

The state of the LSTM cell unit is updated to C (t), and the current state information controlled by the input door and the memory state information controlled by the forgetting door are included, as shown in the formula (6);

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

W _i ，b _i weight of input gate, bias term, respectively;

W _c ，b _c respectively, the weight of the tanh layer, the bias term;

and finally, the output gate determines the output part information. And mapping the internal state information C (t) to a range from-1 to 1 through the tanh layer, and multiplying the internal state information C (t) by an output vector o (t) (formula (7)) of the output gate sigmoid layer to obtain an output h (t) of the current state, wherein the output h (t) is shown in a formula (8).

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

W _o ，b _o Weight of the output gate, bias term, 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, which is composed of two LSTM layers and two fully connected layers, is shown in fig. 6. The input and output of the model are normalized capacity and corresponding RUL label, respectively. The LSTM layer is used to extract the temporal features hidden in the capacity sequence. Typically, the first few LSTM layers extract essential features of the capacity. The higher the LSTM layer, the more abstract the extracted features and the greater the task relevance. And fitting the mapping relation between the capacity characteristics and the RUL labels by utilizing a full connection layer. And in the model testing process, the RUL label value output by the test is subjected to inverse normalization, and the predicted value of the residual service life is obtained.

The application provides two mixed migration learning strategies to realize the RUL migration prediction process based on the Tr-LSTM model. In the aspect of model layer migration, both strategies inherit the structural parameters of each layer of the pre-training model, and learned historical experience is migrated from the pre-training model; in the aspect of feature layer migration, according to different migration strategies, the new model respectively inherits and freezes the weight parameters of the pre-training model designated layer, and training of the model is accelerated. And finally, randomly initializing the weight parameters of the rest layers, and inputting a transferable sample to perform fine tuning training on the rest layers of the Tr-LSTM model.

For a battery at 25 ℃, a migration strategy 1 is adopted, only the first-layer weight parameters of the model are frozen, which is helpful for inheriting the general characteristics extracted by the pre-training model, and only the second-layer characteristics of the model are adjusted according to the specificity of the target battery.

For the battery with the temperature of 60 ℃, the migration strategy 2 is adopted, the accelerated degradation is caused by the temperature rise, the data volume is reduced, and the model cannot be effectively trained, so that the weight parameters of the first two layers are frozen to improve the training efficiency of the model, and the prediction accuracy is improved.

For a 45 ℃ battery, the life distribution is between 25 ℃ and 60 ℃, therefore,

migration strategies

1 and 2 are used to predict the RUL, respectively, with the arithmetic mean of the prediction results as the final prediction value of RUL.

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

4. Lithium battery cycle life test optimization based on life prediction

Aiming at the cycle life test of the lithium ion battery, the experiment optimization method based on the residual life prediction is provided, and guarantees are provided for reducing the cycle life test time of the lithium ion battery and improving the design and development efficiency of the lithium ion battery.

4.1 introduction to cycle life experiment of lithium cell

(1) Test platform

The lithium battery cycle life experiment that this application was carried out is gone on lithium ion battery cycle life test bench.

(2) Lithium ion battery formula

The feasibility and the effectiveness of the provided lithium ion battery service life prediction method are verified by adopting enterprise test data (note that the experiment is different from a battery in an actual product, but is a special soft package battery in a product design stage). In the experiment, 10 groups of lithium batteries with the same battery platform and different formulas are selected, namely A, B, … … I and J groups in sequence, each group represents a lithium battery formula, the different groups have the same cathode and separation materials, but the anode materials and the electrolyte solutions of the batteries are different.

(3) Experiment temperature T

The experiment considers the standard experiment temperature condition (25 ℃) and the high-temperature experiment condition (45 ℃ and 60 ℃), and the cycle life experiment is carried out on a plurality of batteries at each group of experiment temperature to obtain the cycle life of the lithium batteries with respective corresponding formulas.

(4) Charging and discharging process

The capacity of the lithium battery used in the experiment is about 2070mAh, all experiments are carried out under the same charging rate, namely a standard constant current/constant voltage mode, the constant current rate is 1C before the voltage reaches 4.2V, and then the voltage value is kept at 4.2V until the charging current is reduced to be below 0.1A. The cyclic discharge process was performed at a constant discharge current 2A, and the discharge cut-off voltage was 2.78V.

(5) Failure threshold TS _Failture (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, i.e., the degradation of the battery capacity to 82% of the initial capacity, was reached. Therefore, 82% is considered as the failure threshold TS of the target battery and the reference battery _Failture 。

(6) Target Battery test termination threshold TS _obj

In order to sufficiently reduce experimental error, enoughThe data that are redundant ensure the validity of the proposed prediction method, it is necessary to test the termination threshold TS of the data for the target battery _obj Optimizing, and taking 90% of initial capacity as target battery test data termination threshold TS in experiment _obj And predicting the residual life of the lithium battery through test measurement data.

4.2 Experimental Process design optimization

The detailed experimental process after optimization is shown in fig. 8, and the main steps are as follows:

a. target Battery cycle Life test at temperature T _obj Proceeding until a target battery experiment termination threshold TS is reached _obj The experiment was stopped. The target battery capacity degradation data obtained from the experiment was normalized by data preprocessing.

b. Selecting the battery platform which belongs to the same battery platform as the target battery and is at the temperature T from the historical database _obj And (3) obtaining the lithium battery with complete capacity degradation data through a cycle life experiment, and standardizing the capacity degradation data and the RUL label.

c. And selecting sample data most similar to the target battery as a reference battery through four screening processes of capacity curve form, capacity curve distance, capacity curve future trend and battery life concentration for predicting the residual life of the target battery.

d. And standardizing the capacity degradation data of the selected reference battery and the RUL label, inputting the standard into an LSTM deep learning model, and training the model.

e. When the capacity of the target battery reaches the battery cycle life experiment termination threshold TS _obj And inputting part of known capacity data of the target battery into the well-trained LSTM model to obtain the predicted RUL value of the target battery. The cycle life predicted value can replace a real life value obtained by a cycle life experiment, so that the cycle life experiment time can be reduced, the experiment efficiency is improved, and the cost of lithium battery design and development is reduced.

Test example 1

1. Test and data presentation

The feasibility and the effectiveness of the lithium ion battery service life prediction method provided by the invention are verified by adopting test data of a certain battery enterprise (note that the battery used in the test is a soft package battery specially used in a product design stage, which is different from the battery used in a company real product). The selected test database contains 10 lithium ion batteries with different formulas, namely A, B, … …, I and J, and tests are carried out at 25 ℃, 45 ℃ and 60 ℃ to obtain battery degradation data under three temperature conditions.

2. Result of data preprocessing

According to the invention, the battery capacity degradation data is used as a performance index for reflecting system degradation, and the failure threshold value is set to be 0.82 through preliminary analysis of the data, namely when the capacity is degraded to 82% of the initial capacity, the battery is considered to reach the end point of the service life.

According to the data normalization standard, battery capacity degradation data is normalized to 1-0 (initial capacity is 1, 82% of initial capacity is 0), and the corresponding remaining life is also normalized to 1-0 (remaining life corresponding to initial capacity is 1, remaining life corresponding to 82% of initial capacity is 0). The normalized battery capacity degradation original data curve has large jump fluctuation, which is caused by a complex chemical reaction process in the discharging process of the lithium battery, and the original data is smoothed by adopting a local weighted regression method in order to facilitate data analysis. The effect graph of the lithium battery capacity smoothing pretreatment based on the local weighted regression method is shown in fig. 9.

After smoothing the capacity curve of the lithium battery based on the local weighted regression method, screening, 147 different batteries were selected for the study, and the study is specifically shown in table 1:

TABLE 1 Battery screening results

3. Migratable sample selection results

The research of the invention limits the migration prediction of the battery life under the conditions of the same temperature, the same multiplying power and different formulas, and sets different experimentsAn end-of-experience threshold, e.g. TS _{Fail to work} The test was stopped when the capacity had degraded to 90% of the initial capacity. Through a four-time screening method, the similarity of the target battery capacity degradation curve and the curve when the capacity of the rest sample batteries (different formulas) is degraded to 90% length at the same temperature and the same multiplying power is compared, the optimal sample battery is selected as a transferable sample to predict the residual cycle life of the target battery, and the life prediction precision is improved. A schematic diagram of the results of the migratable sample screening is shown in fig. 10.

4. Lifetime migration prediction based on LSTM model

1) Parameter setting of LSTM model

The optimization of computing resources and time is comprehensively considered by combining the data characteristics of the battery, the adopted LSTM model structure comprises a data input layer at the bottom layer, two hidden layers and a data prediction layer at the top layer, the number of neurons of each layer is respectively 100,50,50 and 1, the activation transfer function is a sigmoid function, the learning rate is 0.3, the noise reduction shielding ratio is 0.15, and the optimizer is selected as adam. The size of each sub-input block is 100 for deep learning. To ensure the sufficiency of the feature self-learning, the loop execution step of the unsupervised learning and back propagation process is epochs-8. See table 3 for details. In order to avoid the influence of single data anomaly on the prediction label, the input of the deep learning model is that one-dimensional data is normalized into data columns in the form of a sliding window and input into the LSTM model by taking a data interval as a reference (the number of rows in each column is the number of neurons in an input layer).

2) Life prediction results for LSTM model

The target battery and the corresponding reference battery obtained by four screening processes are subjected to life prediction, and the prediction results are shown in table 4, taking 25 ℃ as an example:

TABLE 225 ℃ Battery Life prediction results

TABLE 345 ℃ Battery Life prediction

TABLE 460 ℃ Battery Life prediction

According to the results in the table 2, the maximum precision of the service life prediction by using the transferable sample provided by the invention can reach 99.9%, the cost for analyzing the formula performance of the battery is saved, and a large amount of test cost can be saved and effective data sharing can be realized by data similarity analysis of batteries with different formulas. The method brings a new choice for the battery cycle life test, and has good economical efficiency and practicability.

Although the present invention has been described in detail hereinabove, the present invention is not limited thereto, and various modifications can be made by those skilled in the art in light of the principle of the present invention. Thus, modifications made in accordance with the principles of the present invention should be understood to fall within the scope of the present invention.

Claims

1. A method for predicting life migration of a cross-formula similar product is characterized by comprising the following steps:

preprocessing short-term cycle life test data of a to-be-tested formula similar product to obtain target sample data, and preprocessing full-life test capacity data of other formula similar products to obtain a plurality of training data;

2. The method of claim 1, wherein conducting life migration predictions for cross-recipe similar products using a life prediction model adapted to the migratable sample data comprises:

3. The method of claim 2, in which the pre-processing comprises normalising target sample data and training data;

4. The method of claim 1, wherein screening out first training data from the plurality of training data that is of a similar type as a target sample data curve comprises:

5. The method of claim 4, wherein screening out second training data from the first training data that is similar to a capacity degradation trend of target sample data comprises:

and using the retained first training data as second training data.

6. The method of claim 5, wherein screening out third training data from the second training data that is similar to the target sample data lifetime distribution comprises:

7. The method of claim 6, wherein the screening out fourth training data from the third training data having a smallest distance metric to the target sample data comprises:

8. The method of claim 3, wherein the pre-processing further comprises:

9. The method of any of claims 1-8, wherein the similar product is a lithium battery.

10. A cross-recipe similar product life migration prediction system comprising a processor, a memory, a program stored on the memory and executable on the processor, and a data bus for enabling connection communication between the processor and the memory, the program when executed by the processor implementing the cross-recipe similar product life migration prediction method of any of claims 1-9.