CN113536671A

CN113536671A - Lithium battery life prediction method based on LSTM

Info

Publication number: CN113536671A
Application number: CN202110778444.8A
Authority: CN
Inventors: 马剑; 徐沛洋; 许庶; 陶来发; 程玉杰; 丁宇; 王超; 索明亮; 吕琛
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2021-07-09
Filing date: 2021-07-09
Publication date: 2021-10-22
Anticipated expiration: 2041-07-09
Also published as: CN113536671B

Abstract

The invention provides a lithium battery life prediction method based on LSTM, which comprises the following steps: acquiring a capacity degradation data set of a lithium battery; preprocessing the volume degradation data set; constructing a residual life prediction model based on LSTM; after constructing a residual life prediction model based on LSTM, further constructing three lithium battery local life prediction models and a central server end global life prediction model, wherein the LSTM is of a network structure, and any one of LSTM units comprises a forgetting gate, an input gate and an output gate; the three lithium battery local life prediction models have the same structure, and the model structure comprises two LSTM layers, two Dropout layers for preventing overfitting and a top prediction output layer.

Description

Lithium battery life prediction method based on LSTM

Technical Field

The invention relates to the technical field of fault testing and prediction of high-end equipment, in particular to a lithium battery service life prediction method based on LSTM.

Background

In recent years, with the improvement of industrialization degree, the group products gradually develop from simplification and simplification to intellectualization and complication, and have more and more important application values in many fields, which also puts certain requirements on the reliability of the group products. The fault Prediction and Health Management (PHM) provides a feasible solution for improving the reliability of group products, and the fault prediction and Health Management technology uses technologies such as data mining and information fusion, etc. to research and analyze real-time state data of products, so that the product maintenance cost can be reduced, and the reliability of the products can be improved.

The artificial intelligence technology is adopted in the field of fault diagnosis and prediction to challenge and have opportunities, and the research hotspot is formed by how to quickly diagnose the fault and accurately give the prediction result by using the artificial intelligence technology. The group products are of various types, the failure modes of each type of group products are complex and various, it is impractical to develop diagnosis and prediction models for all possible failure modes, and it is very costly to collect all state data of the product in the whole life cycle. The similarity of product operation state data collected on group products is limited, the direct integration of the data is not beneficial to the accuracy of a prediction result, great difficulty is caused in joint modeling, the difficulty of data island and related laws issued in recent years about data privacy protection problems further limit the freedom of data exchange among enterprises, and the problem of data island and privacy protection seriously hinders the further development of artificial intelligence technology.

The research on the related technologies of fault prediction and health management has great significance to group products. On the one hand, the waste of maintenance time is avoided, and the economic cost of maintenance is reduced. On the other hand, compared with the traditional regular maintenance and post-fault maintenance, the PHM for the product can effectively improve the reliability and safety of the product and reduce the occurrence of catastrophic accidents. However, in practical application, the privacy of data needs to be protected, the data which can be fully collected is insufficient, the cost of the collected data is huge, and in addition, products of different enterprises are in different working condition environments, the direct integration modeling effect is poor.

The Life prediction includes End of Life (EOL) prediction and Remaining Life (RUL) prediction, and the relationship between them can be expressed as RUL ═ EOL-T, where T is a given time. The RUL prediction method based on the analytic model is to construct a mathematical model for describing a product degradation process by researching a physical mechanism of product failure, model parameters are required to be continuously updated according to state data of a product in the process, and finally the aim of predicting the residual life of the product by using historical data is achieved. The method realizes the prediction of the service life by researching and analyzing the information contained in the product failure through an advanced data mining technology, and utilizes a probability statistical method or a machine learning method to construct a prediction model, thereby reducing the requirements on the physical and mathematical abilities of research personnel on the premise of ensuring higher accuracy of the prediction result. Artificial neural networks are most frequently used in machine learning based life prediction methods. Deep learning models such as a recurrent neural network have strong characteristic self-learning capacity, and how to apply the deep learning models to the aspect of life prediction is also a research hotspot. In general, the life prediction technology based on data driving does not need to research the physical mechanism behind the product failure, reduces the requirements on model designers, and is widely applied to the field of residual life prediction of various products.

In summary, due to the fact that the group products are complex and diverse, the failure mechanisms underlying the group products are also complex and diverse, and the residual life prediction method based on the analytic model needs to research the corresponding failure mechanisms and establish a proper model for each failure mode, so that the method is not suitable. Based on the method, the data-driven LSTM (long-short memory network) residual life prediction method is selected for prediction, the defect that the method based on probability statistics seriously depends on a large amount of data is overcome, the residual life of group products is accurately predicted, the maintenance of the products is effectively guided, and the reliability of the products is provided.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a residual life prediction method based on LSTM.

According to one aspect of the invention, the method comprises: a lithium battery life prediction method based on LSTM comprises the following steps:

acquiring a capacity degradation data set of a lithium battery; pre-processing the volume degradation data set, the pre-processing comprising: and (3) data normalization processing, wherein the data normalization method is min-max normalization: dividing a data set to divide the data set into a training set and a testing set before model training; constructing a residual life prediction model based on LSTM; the residual life prediction model is based on a neural network RNN and a Long Short Term Memory (LSTM) network that process time series data, the LSTM being used to eliminate gradient explosion and gradient disappearance problems existing in the RNN network; after constructing the LSTM-based residual life prediction model, further constructing three lithium battery local life prediction models and a central server end global life prediction model, wherein; the LSTM is a network structure, and any LSTM unit comprises a forgetting gate, an input gate and an output gate; the three lithium battery local life prediction models have the same structure, and the model structure comprises two LSTM layers, two Dropout layers for preventing overfitting and a top prediction output layer; and finally, sending the training set into the constructed local life prediction model of the lithium battery and the overall life prediction model of the central server, carrying out model training and outputting the predicted residual life of the lithium battery.

According to the embodiment of the present invention, the activation functions of the two LSTM layers are hyperbolic tangent tanh functions, the activation function of the last output layer sense layer is a linear function, and Dropout rate is set to 0.3.

Further, in the forgetting gate, the following operations are performed:

the information of the previous time is selectively discarded before being transmitted to the next time: h is to be_t-1And x_tThe following formula is substituted to calculate a value belonging to [0, 1]]The value of the vector represents the cell state C_t-1How much information is retained or discarded; 0 means no reservation, 1 means all reservations;

f_t＝σ(W_f·[h_t-1，x_t]+b_f)。

further, it is further decided what new information to add to the cell state: i.e. i_tIs the weight coefficient of the updated information, h_t-1And x_tSubstituting the following formula to obtain the i_tThen by activating the function tanh, using h_t-1And x_tGenerating new candidate state vectors

Among them are:

i_t＝σ(W_i·[h_t-1，x_t]+b_i)

further, the status information may be updated by the following formula, wherein:

cell export also needs to be according to h_t-1And x_tTo judge, firstly h is_t-1And x_tBringing in

Obtaining the judgment condition, and then substituting Ct into the tanh activation function to calculate a value of [ -1, 1 [ ]]Multiplying the vector with the judgment condition to obtain the final output, wherein:

o_t＝σ(W_o[h_t-1，x_t]+b_o)

h_t＝o_t*tanh(C_t)。

further, in the model training process, the batch training sample batch _ size is 32, the length of a single sample data is 50, the model training selection optimizer Adam has a learning rate of 0.001, and the training times epochs is 20.

Further, after model training, the lithium battery local life prediction model sends a weight coefficient and a loss value to the central server end global life prediction model, and the central server obtains the weight transmitted by the lithium battery product end and further processes the weight to weight the weight value.

Further, the weight value is determined by:

w_iweight, loss, representing the ith product-side prediction model_iThe loss value of the ith product end prediction model is represented, w represents the weight of the central server prediction model, and the w calculated by the central server is sent to the single product end for updating the weight, wherein the communication frequency of the single equipment end and the central server end is set to be 20.

Further, after model training, determining the initialization parameters of the local prediction model of the next iteration of the local life prediction model of the lithium battery; the initialization parameters are determined by the following method: directly adopting the structure of the local life prediction model in the previous round to ensure that the model structure is unchanged, freezing the weight parameters and the structure parameters of the first four layers, randomizing the weights of the other layers, performing the next round of training, inputting the state data of the previous 50 moments, and outputting the state data of the next moment.

Further, by adopting the initialization parameter determination method, the average accuracy of the prediction result is 95.48%.

Drawings

Various embodiments or examples ("examples") of the disclosure are disclosed in the following detailed description and accompanying drawings. The drawings are not necessarily drawn to scale. In general, the operations of the disclosed methods may be performed in any order, unless otherwise specified in the claims. In the drawings:

FIG. 1 illustrates the general framework of the present invention for predicting remaining life for a group product;

FIG. 2 illustrates the Federal migration learning process for group product life prediction of the present invention;

FIG. 3 is a design of a single product end local life prediction model and a central server end global life prediction model;

FIG. 4 is a schematic diagram of an LSTM network structure;

FIG. 4A is an LSTM forgetting gate unit;

FIG. 4B is an LSTM input gate unit;

FIG. 4C illustrates an update operation of the LSTM;

FIG. 4D is a graph showing the state of output cells;

FIG. 5 is an LSTM life prediction model;

FIG. 6 is the original data of a lithium battery at 25 ℃;

FIG. 7 is a graph of normalized results of data for a 25 ℃ lithium battery;

FIG. 8 is a comparison graph of data smoothness of lithium batteries;

FIG. 9 is a model structure for predicting the end-of-life of a single lithium battery product based on LSTM;

FIG. 10 is a graphical illustration of a local prediction result for a single product end;

FIG. 11 is an LSTM-based learning strategy;

FIG. 12 is a graphical illustration of a scenario one predicted result;

FIG. 13 is a graphical illustration of a prediction result for scenario two;

FIG. 14 is a graphical illustration of the results of a solution three prediction.

Detailed Description

Before explaining one or more embodiments of the present disclosure in detail, it is to be understood that the embodiments are not limited in their application to the details of construction and to the procedures or methods set forth in the following description or illustrated in the drawings.

The residual life prediction of the group products is researched, and an overall framework aiming at the residual life prediction of the group products is designed firstly. The client-server architecture is the most common one, and is relatively simple. In this framework, multiple participants (also referred to as users or clients) with local data collaboratively train a machine learning model that is applicable to all participants, with the help of a central server (also referred to as a parameter server or aggregation server). The specific workflow under this framework is such that: (1) the multiple participants train the model locally, obtain parameters such as the weight of the model, and then send the parameters to the central server for further processing. (2) And the central server performs weighted average processing after obtaining the parameters uploaded by the participants. (3) And the central server transmits the result after weighted average processing to each participant. (4) And each participant updates the local model by using the parameters sent by the central server. The above steps are continued until a given number of iterations is reached, and the final model is obtained.

Specifically, the overall framework designed by the present invention for predicting remaining life of a group of products is shown in FIG. 1. As shown in fig. 1, the overall framework of predicting remaining life of the group product of the present invention includes a plurality of single product ends and a central server end, and three single products are exemplarily shown. Each single product end obtains the original data of a group product (such as a lithium battery) by means of a sensor, the original data is preprocessed firstly, the preprocessing comprises data normalization, data smoothing, data set division and the like, then training is carried out on the local single product end, and the obtained weight parameters are uploaded to a central server end. In the single product-server architecture shown in fig. 1, three exemplary single product ends do not have any interaction with each other, and only the weight parameters of the residual life prediction model trained at the single product end need to be uploaded to the server end, then the server end performs weighted average processing on the weight parameters and sends the processed weight parameters to the three single product ends, and the three single product ends adopt a transfer learning idea to select partial parameters for updating the model, and the above operations are repeated until the global model of the residual life prediction at the server end converges.

FIG. 2 is a flow chart of Federal transfer learning for group product remaining life prediction designed according to the present invention. As shown in fig. 2, the learning flowchart includes: at the beginning, acquiring the original data of group products; sending the original data into a residual life prediction model of the single product end, training, and obtaining a weight parameter of the residual life prediction model of the single product end after training; the single product end sends the trained weight parameters to the server end; the server side collects the weight and loss value of all local life prediction models, and sends the weight to each single product side after weighting processing; after receiving the weight parameters sent by the server, each single product end selects a proper migration learning strategy to update the local life prediction model; and the monomer product end receives the original data again on the basis of the updated local life prediction model, trains the model, and uploads the weight parameters of the residual life prediction model of the local monomer product end. The process is a complete cycle, whether the cycle times reach the preset iteration times or not is judged, if not, the training of the single product local life prediction model is continued, otherwise, the model at one product end is arbitrarily selected to inherit all the weight parameters, and the final whole model is obtained.

After a general framework for predicting the residual life of the group products is designed, the tasks are completed according to the framework content. Firstly, a monomer local life prediction model is designed, and the model mainly comprises the following two parts: data analysis and pretreatment, monomer local life prediction model construction, and then central server end global life prediction model construction. The specific flow is shown in fig. 3. The product operating state data collected by the sensors often needs to be preprocessed before being used in the training and testing of the subsequent life prediction model. The application chooses to use a machine learning based approach to build a product-side local life prediction model.

An exemplary module of the present invention will be described in detail with reference to fig. 3.

Step 1: single product end data analysis and preprocessing

Due to the influence of non-steady state in the product running process or abnormal conditions when the sensor collects data, the product running state data collected by the sensor often has large fluctuation values or data abnormality, and the model prediction effect obtained by directly using the collected data for the training of the model is not good, so that preprocessing is often needed before the data is used for predicting the service life. The data preprocessing mainly comprises the following three parts: data normalization processing, data smoothing processing and data set division.

Step 1.1: data normalization processing

The range and the scale of the data acquired every time are different, if one data with a very large scale range exists, the influence of the data with a small scale range on model training can be ignored, and therefore normalization is needed. Investigation has found that there are two main methods for normalizing the data: 0 mean Normalization (Z-score Normalization), Min-Max Normalization (Min-Max Normalization).

The 0-mean normalization is performed on each data by using the following formula, and the distribution of the processed data conforms to the standard normal distribution, that is, the mean value is 0 and the standard deviation is 1. The sample data is represented by the following formula, and μ represents the mean value of all sample data, and is the standard deviation of all sample data.

min-max normalization refers to mapping sample values to [0, 1] via transformation]Sample data X (X)₁，x₂,.. the following equation.

The two data normalization methods have advantages and disadvantages respectively. The process of 0-mean normalization is relatively complex, requiring a priori computation using samples, and is relatively complex. For min-max normalization, the processing procedure is relatively simple, the sample data is changed into decimal between [0 and 1], and the min-max normalization method is simpler and more convenient when the distance measurement is not involved. In practice, the method is selected according to specific situations.

Step 1.2: data smoothing

The collected data often has burrs and noises, and the effect of directly using the data for prediction analysis is not good, so that the original sample data needs to be subjected to smoothing pretreatment. The Local Weighted Regression (LWR) algorithm is often used in data smoothing processing, and its work flow is as follows: firstly, dividing a sample into a plurality of intervals, carrying out polynomial fitting on a local sample, then estimating a fitting value by using a least square method, and finally, finishing the smoothing treatment on the original sample. The main idea of the local weighted regression algorithm is to calculate different weight coefficients according to the distances between other points and observation sample points, multiply the sample points by corresponding fitting weights and add the sample points to obtain the fitting values of the observation points, and all the sample points can obtain smooth data with noise removed through the processing. The specific mathematical principle is as follows:

a range scale of 2K is set in advance, Q ═ Q for the sample set₁，q₂，...，q_NAny one sample point q in_i(i ═ 1, 2, …, N), the weighted fit is obtained using the following equation.

Wherein the weight coefficient w_i(q_k) Is determined by the formula, from which it can be seen that for the distance observation point q_kThe farther away the sample point q_iCorresponding to the weight coefficient w_i(q_k) The smaller the value of (a), and for the distance observation point q_kThe closer sample point q_iCorresponding to the weight coefficient w_i(q_k) The larger the value of (A), the better the noise point of anomaly in the original data set can be removed.

The local weighted regression algorithm can effectively remove noise in the original data set, so that the original data drawing curve becomes smooth, and the occurrence of over-fitting or under-fitting conditions is avoided.

Step 1.3: data set partitioning

After data normalization and data smoothing, the data set needs to be divided into a training set and a test set before model training. The K-fold cross validation is a method for dividing a data set, and is completed by extracting certain data in the data set as a test set and taking the rest of the data as a training set, which are not repeated each time, so that the data can be fully utilized, and the method is suitable for being used in a classification task. The leave-in method is also a method frequently used for dividing the data set, is simple and convenient to operate, and is suitable for being used when the sample data size is sufficient. Particularly, when the method is applied, the proper method is selected according to the task requirement and the size of the data volume.

Step 2: construction of residual life prediction model based on LSTM

With the development of sensor technology, more and more product running state data are obtained through sensors, and the traditional shallow machine learning algorithm is not careful in processing mass data. And the strong nonlinear mapping capability and the high-dimensional feature extraction capability of deep learning make the deep learning very suitable for the situation. RNNs are neural networks that process time series data, and what is being investigated here is a group product whose operating state data is slowly decaying over time, belonging to time series data, and thus suitable for application of RNNs to solve. However, the problem of gradient explosion and gradient disappearance exists in pure RNN, and long-short term memory (LSTM) solves the problem through selective forgetting. LSTM is therefore chosen herein for predicting the remaining life of a group product, and fig. 4 is a schematic diagram of a LSTM network structure according to the present invention.

As shown in fig. 4. An LSTM cell contains three gates to control cell states, called forgetting gate, input gate, and output gate.

The information of the previous moment is usually discarded selectively before being transmitted to the next moment, and the forgetting gate can complete the operation. H is to be_t-1And x_tSubstituting the formula shown in the following figure to calculate a value belonging to [0, 1]]The value of the vector represents the cell state C_t-1How much information is retained or discarded. 0 means no reservation and 1 means all reservations. The forgetting gate is shown in fig. 4A, which includes:

f_t＝σ(Wf·[h_t-1，x_t]+b_f)

the next step is to decide which new information to add to the cell state. i.e. i_tIs the weight coefficient of the updated information, h_t-1And x_tSubstituting into the first formula as shown in the following figure may result, then using h by the activation function tanh_t-1And x_tGenerating new candidate state vectors

These two steps are depicted in FIG. 4B, in which:

i_t＝σ(W_i·[h_t-1，x_t]+b_i)

the state information can be updated through the formula shown in the following figure, and the updating covers part of the state information at the previous moment and also comprises part of the state information at the moment according to the formula. The update operation is illustrated in FIG. 4C, where:

Obtaining a judgment condition, and adding C_tSubstituting into tanh activation function to calculate a value of [ -1, 1 [)]The vectors in the space between the two are multiplied by the judgment condition to obtain the final output. This step is illustrated in fig. 4D, where:

o_t＝σ(W_o[h_t-1，x_t]+b_o)

h_t＝o_t*tanh(C_t)

and step 3: building of single product end local life prediction model and central server end global life prediction model

After the residual life prediction model based on the LSTM is built, a single product end local life prediction model and a central server end global life prediction model are further built.

The building of the local and global life prediction models is completed by using an LSTM layer, a Dropout layer and a Dense layer, generally, in the building process, an original model can be designed according to experience, then, the number of LSTM layers is increased or decreased according to the accuracy of a prediction result and the length of prediction time by inputting samples, and generally 1 to 3 LSTM layers can meet the accuracy requirement of the prediction result. Fig. 5 is an exemplary LSTM lifetime prediction model according to the present invention, which includes 2 LSTM layers and 2 sense layers, as shown in fig. 5. The input is status data of the product and the example output is a predicted life of the product.

In the invention, in order to facilitate the interactive update of the parameters of the central server and the single product, the service life prediction model structure of the central server is the same as that of the single product, so that the convergence of the global model can be ensured, and the service life prediction model structure of the central server and the service life prediction model of the single product are completed in one construction process when the models are constructed.

The LSTM life prediction model is a model structure of which a single product end life prediction model needs to be fixed, and the single product end weight parameter part after summary processing is used as an initialization parameter. Therefore, the relation is to summarize and process the monomer product end weight parameters, and consider which part is selected as the monomer product end prediction model initialization parameter of the next iteration.

Firstly, parameter summarization processing of a local life prediction model of a single product end is carried out. The existing processing mode is to weight the weight of the model according to the number of samples at each product end, the weight coefficient of the weight of the product with more samples is larger, but the influence of the loss value of the model on the weight coefficient is ignored, and the prediction accuracy of the global model is not high enough. The invention comprehensively carries out weighting processing on the weight value according to the loss value and the sample number, as shown in a formula (1.1). w is a_iWeight, loss, representing the ith product-side prediction model_iAnd representing the mean square error value of the ith product-side prediction model, and comprehensively considering the influence of the sample number and the model loss value on the weighting coefficient. w represents the weight of the central server prediction model, and k represents the number of the single product ends. The idea of doing so is that the single product end model with a large loss value accounts for a larger proportion of the global model, and the local data thereof has a large influence on the global model. And w calculated by the central server is sent to the single product end for updating the weight.

The above process is a complete iteration of information interaction between the single product side and the central server side, and the process needs to be repeated for many times in order to achieve a satisfactory prediction effect of the global model. Generally, a fixed number m of iterations is preset, and the process is stopped when the number m is reached repeatedly.

Next, it is described which part is selected as the initialization parameter of the local prediction model for the next iteration. Since the global life prediction model herein is LSTM-based, there is a need to master the implementation of LSTM in combination with model-based migration learning. The method for realizing the combination of the LSTM and the model-based transfer learning comprises the steps of fixing the structural parameters and the weight parameters of the LSTM layer by layer, then inputting data again to train other layer numbers to obtain new weight parameters, and selecting the optimal freezing layer number and the optimal structural parameters according to the accuracy of the prediction result of the model. The structural parameters refer to learning rate, activation function, optimizer and other parameters of the model, and the weight parameters refer to bias and weight values of the model. The principle is that the first layers of the trained model have a good function of capturing the input data feature relationship, the process of retraining the model again to obtain the captured input features can be omitted by directly transferring the layers, and actually, the transfer of the layers is determined according to the accuracy of the output prediction result in specific application. Therefore, the invention can complete the migration learning by freezing part or the whole of the global life prediction model of the central server.

Example embodiments: prediction of remaining life of lithium battery

The lithium batteries of different models have certain similarity and certain difference, can be regarded as a group product, have larger difference of degradation trend of the lithium batteries of different models, and are higher in cost for acquiring the capacity data of the lithium batteries and generally unwilling to share the data among different enterprises due to competition, so that the method is suitable for predicting the service life of the lithium batteries based on the LSTM.

The data set used in the invention is acquired through a cycle life experiment, the lithium ion battery cycle life test bed is provided for battery research and development enterprises, the temperature condition acquired by the data set is 25 ℃, a voltage-limiting constant-current charging and discharging mode is utilized, and the stopping condition of the experiment is that the capacity of the lithium ion battery is degraded to 82% of the initial value.

The data set includes 10 different models of lithium ion battery capacity degradation data, 3 of which are used herein (groups a, B, and C). The total of 13 groups of data are obtained, and each group of data averagely comprises 1000 capacity lithium battery data. The cathode materials are the same between different groups, but the anodes differ. The degradation data of the batteries of different types at 25 ℃ selected by the invention is shown in table 1.

Table 125 ℃ battery grouping

The A, B, C lithium batteries were combined and used as a data set for model training, and the specific combinations were as shown in table 2, for a total of 35 combinations.

TABLE 2 data set grouping case

Because a plurality of groups of lithium ion batteries with different models are selected, the trend difference of the failure process between different groups is large, and the original degradation data (residual capacity) of the lithium ion batteries with different models is shown in fig. 6. It can be seen from the figure that the raw data before being processed has different failure process trends, which is not favorable for convergence of the global life prediction model at the subsequent central server side.

In the process of testing the cycle life of the lithium battery, because materials of different battery models are different, the capacity data of the lithium battery needs to be normalized. Since no distance metric and covariance calculation are involved here, a simpler min-max normalization method is employed. Setting the failure threshold value as 82% of the initial capacity of the lithium battery, recording the initial capacity as 1, recording the failure threshold value as 0, and sampling data X (X)₁，x₂，...，x_n) The normalization method of (a) is shown in formula (1), and the normalization result of the sample data of the lithium battery at the temperature of 25 ℃ is shown in fig. 7. It can be seen that normalizationThe samples are distributed in the same way, and the accuracy of the life prediction result is improved.

As can be seen from fig. 7, the normalized data plotted curve still has spike-like fluctuation, so that the original battery data needs to be subjected to smoothing preprocessing and processed by using a local weighted regression algorithm. FIG. 8 is a comparison of data before and after smoothing of a lithium battery sample. In the figure, a black curve is drawn by original degradation data, and a blue curve is drawn by smoothed data. It can be seen from the figure that the original data degradation black curve has burr-like fluctuation, and the smooth data blue curve eliminates noise interference on the premise of keeping the curve trend similar to that of the original data black curve through local weighted regression smoothing processing, which is beneficial to improving the accuracy of the prediction result.

After normalization and smoothing, the data set needs to be divided, and the sample data volume of each group is sufficient, so that a reservation method is directly used. Here, 60% of the data set is divided into training sets and the remaining data is divided into test sets.

The single product end local prediction is that each single product trains a local life prediction model by using self data locally, and then the weight parameters of the model are uploaded to a central server end. The local prediction model of the single product end utilizes an LSTM model, and three single product ends all adopt the same model structure. The model structure includes two LSTM layers, two Dropout layers to prevent overfitting and one top-level predicted output layer. The activation function of the first two LSTM layers is tanh (hyperbolic tangent), the activation function of the last output layer sense layer is linear, and Dropout rate is set to 0.3. The Dropout rate is set through multiple trials, the Dropout rate is increased when the model is over-fitted, the Dropout rate is decreased when the model is under-fitted, and finally the fitting effect of the model is determined to be better when the Dropout rate is 0.3, and the model structure is shown in fig. 9 specifically. In the model training process, the batch training sample batch _ size is 32, the input data length of a single sample is 50, the output data length is 1, namely the 51 st data is predicted by the previous 50 data, the model training selection optimizer rmsprop has the learning rate of 0.001 and the training times epochs of 20.

The model is used for predicting the service life of A, B, C types of lithium battery data collected under the temperature condition of 25 ℃, the accuracy of the prediction result is recorded, the degradation data of one sample of one type of battery is input each time, and the obtained partial result is shown in figure 10. The prediction effect is recognized from the degree of coincidence of the two curves, and the prediction effect is generally acceptable from the viewpoint of the image, but there are a few cases where the prediction result error is large, such as B _2006 and C _ 3006.

The model of the central server side is the final global prediction model, and in order to enable the global model to have a good prediction effect on the data of each single product, the global life prediction model is still built by using the LSTM network.

In order to facilitate final convergence of the global model and process parameters such as weight transmitted by the single equipment, the global model adopts a model structure similar to that of the single equipment-side prediction model and comprises two LSTM layers, two Dropout layers for preventing overfitting and a top prediction output layer. In the model training process, the batch training sample batch _ size is 32, the length of a single sample data is 50, the model training selection optimizer Adam has a learning rate of 0.001, and the training times epochs are 20.

According to the disclosure of the present invention, the monomer product end weight parameters are summarized and then considered which part is selected as the monomer product end prediction model initialization parameter of the next iteration.

Firstly, a summarizing process of the weight parameters of the single product end is introduced, the central server acquires the weight transmitted by the single product and needs to further process the weight, and the weight value is weighted by using the loss value according to the content. As shown in formula (2), w_iWeight, loss, representing the ith product-side prediction model_iRepresents the loss value of the ith product-side predictive model,w represents the weight of the central server prediction model. The w calculated by the central server is then sent to the monomer product end for updating the weight, and in the actual operation, it is found that the better prediction effect can be achieved by setting the number of times of communication between the monomer equipment end and the central server end to 20.

The method comprises the steps of carrying out collocation combination on lithium battery data of 1 sample of type A, 7 samples of type B and 5 samples of type C, counting 35 combination modes, training 3 samples of each combination as monomer equipment data at a local prediction model of a monomer product end, and obtaining a weight parameter w_iAnd loss parameter loss_iAnd obtaining the processed weight w after weighted average processing, and then issuing the weight parameters to each monomer product for updating the model.

The following three schemes are described, which is selected as the initialization parameter of the local prediction model in the next iteration, and are specifically shown in fig. 11. The first method is to directly adopt the structure of the local life prediction model in the previous round, ensure the structure of the model to be unchanged, freeze the weight parameters and the structure parameters of the first four layers, randomize the weights of the other layers, perform the next round of training, input the state data of the previous 50 moments, and output the state data of the next moment. And secondly, adopting the structure of the previous local life prediction model to ensure that the model structure is unchanged, freezing the weight parameters and the structure parameters of the third layer and the fourth layer, and then randomizing the weights of the other layers to obtain a new model. The third method is to adopt the structure of the local service life prediction model in the previous round, ensure the structure of the model to be unchanged, freeze the weight parameters and the structure parameters of the first two layers, and randomize the weights of the other layers to obtain a new model.

According to the idea of the first scheme, the structure of the local life prediction model in the previous round is directly adopted, the structure of the model is guaranteed to be unchanged, the weight parameters and the structure parameters of the first four layers are frozen, then the weights of the other layers are randomized, the next round of training is carried out, the average accuracy of the prediction result is calculated to be 95.48%, and the partial drawing of the prediction result is shown as 12.

According to the idea of the second scheme, the structure of the local service life prediction model in the previous round is adopted, the structure of the model is guaranteed to be unchanged, the weight parameters and the structure parameters of the third layer and the fourth layer are frozen, and then the weights of the rest layers are randomized to obtain a new model. The average accuracy of the predicted result is 95.55%, compared with the first scheme, the accuracy is slightly improved, and the partial predicted result is plotted as shown in fig. 13.

And (3) adopting the idea of a third scheme, keeping the structure of the local life prediction model of the previous round unchanged, freezing the weight parameters and the structure parameters of the first two layers, and randomizing the weights of the other layers to obtain a new model. The average accuracy of the obtained prediction results is 95.72%, and the accuracy is improved slightly compared with the second scheme. The prediction results are partially plotted as shown in fig. 14.

The accuracy of the statistical central server-side global model prediction result and the single product-side local model prediction result is shown in the following table. According to the following table, the average accuracy of the prediction results of the local models of the single product end is 95.17%, and the average accuracy of the prediction results of the global model of the central server end is 95.48%, 95.55% and 95.72%, respectively.

TABLE 5.3 prediction accuracy statistics

In this embodiment, the LSTM model and the construction method disclosed by the present invention predict the remaining life of the lithium battery sample, and it is obvious that the accuracy of the prediction result can be effectively improved by predicting the remaining life of the lithium battery sample using the LSTM-based prediction model according to the image and the accuracy record of the prediction result.

The invention researches an LSTM method for predicting the service life of a group product. Aiming at the problem of predicting the residual life of group products, the invention provides a specific implementation method. The method includes the steps that firstly, local life prediction models are predicted at each monomer product end, then parameters of the local models are processed, then next iteration is carried out on the parameters processed by the monomer product end migration part, and finally a global life prediction model is obtained.

The existing FedAvg algorithm is improved at the global life prediction model of the central server, weight coefficients are distributed according to loss values of the life prediction model of the product end instead of the data amount of the product end, and the weight coefficients of the life prediction model with larger loss values are correspondingly larger, so that the accuracy of the global life prediction model is improved.

In conclusion, although the present invention has been described with reference to the embodiments shown in the drawings, equivalent or alternative means may be used without departing from the scope of the claims. The components described and illustrated herein are merely examples of systems/devices and methods that may be used to implement embodiments of the present disclosure and may be replaced with other devices and components without departing from the scope of the claims.

Claims

1. A lithium battery life prediction method based on LSTM comprises the following steps:

acquiring a capacity degradation data set of a lithium battery;

pre-processing the volume degradation data set, the pre-processing comprising:

and (3) data normalization processing, wherein the data normalization method is min-max normalization:

dividing a data set to divide the data set into a training set and a testing set before model training;

constructing a residual life prediction model based on the LSTM, wherein,

the residual life prediction model is based on a neural network RNN and a Long Short Term Memory (LSTM) network that process time series data, the LSTM being used to eliminate gradient explosion and gradient disappearance problems existing in the RNN network;

after constructing the residual life prediction model based on the LSTM, further constructing three lithium battery local life prediction models and a central server end global life prediction model, wherein;

the LSTM is a network structure, and any one of the LSTM units comprises a forgetting gate, an input gate and an output gate;

the three lithium battery local life prediction models have the same structure, and the model structure comprises two LSTM layers, two Dropout layers for preventing overfitting and a top prediction output layer;

and finally, the training set is sent into the constructed local life prediction model of the lithium battery and the overall life prediction model of the central server, model training is carried out, and the predicted residual life of the lithium battery is output.

2. The LSTM-based lithium battery life prediction method of claim 1, wherein the activation function of the two LSTM layers is a hyperbolic tangent tanh function, the activation function of the last output layer sense layer is a linear function, and Dropout rate is set to 0.3.

3. The LSTM-based lithium battery life prediction method of claim 1, wherein in the forgetting gate, the following operations are performed:

f_t＝σ(W_f·[h_t-1，x_t]+b_f)。

4. the LSTM-based lithium battery life prediction method of claim 3 further deciding what new information to add to the cell state: i.e. i_tIs the weight coefficient of the updated information, h_t-1And x_tSubstituting the following formula to obtain the i_tThen by activating the function tanh, using h_t-1And x_tGenerating new candidate state vectors

Among them are:

i_t＝σ(W_i·[h_t-1，x_t]+b_i)

5. the LSTM-based lithium battery life prediction method of claim 4 where the status information can be updated by the following formula where:

Obtaining the judgment condition, then substituting Ct into the tanh activation function to calculate a value of [ -1, 1 [ ]]Multiplying the vector with the judgment condition to obtain the final output, wherein:

o_t＝σ(W_o[h_t-1，x_t]+b_o)

h_t＝o_t*tanh(C_t)。

6. the LSTM-based lithium battery life prediction method of any of claims 1-5, wherein during model training, the batch training sample batch _ size is 32, the single sample data length is 50, the model training selection optimizer Adam has a learning rate of 0.001, and the training times epochs are 20.

7. The LSTM-based lithium battery life prediction method of any of claims 1-5, wherein after model training, the lithium battery local life prediction model sends a weight coefficient and a loss value to the central server side global life prediction model, and the central server obtains the weight transmitted by the lithium battery product side and further processes the weight to weight the weight value.

8. The LSTM-based lithium battery life prediction method of claim 7, wherein the weight value is determined by the following equation:

w_iweight, loss, representing the ith product-side prediction model_iThe loss value of the ith product side prediction model is represented, w represents the weight of the central server prediction model, and w calculated by the central server is sent to the single product side for updating the weight, wherein the communication frequency of the single equipment side and the central server side is set to be 20.

9. The LSTM-based lithium battery life prediction method of claim 8, wherein after model training, the local prediction model initialization parameters for the next iteration of the local life prediction model of the lithium battery are determined; the initialization parameters are determined by the following method: directly adopting the structure of the local life prediction model in the previous round to ensure that the model structure is unchanged, freezing the weight parameters and the structure parameters of the first four layers, randomizing the weights of the other layers, performing the next round of training, inputting the state data of the previous 50 moments, and outputting the state data of the next moment.

10. The LSTM-based lithium battery life prediction method of claim 9, wherein the average accuracy of the prediction results using the initialization parameter determination method is 95.48%.