CN115128468A

CN115128468A - Chemical energy storage battery PHM undervoltage fault prediction method

Info

Publication number: CN115128468A
Application number: CN202210748454.1A
Authority: CN
Inventors: 常伟; 潘多昭
Original assignee: Shanghai Lejia Smart Energy Technology Co ltd
Current assignee: Shanghai Lejia Smart Energy Technology Co ltd
Priority date: 2022-06-28
Filing date: 2022-06-28
Publication date: 2022-09-30

Abstract

The invention discloses a chemical energy storage battery PHM undervoltage fault prediction method, which comprises the steps of firstly preprocessing original undervoltage data; secondly, data exploration is carried out, distribution conditions of all characteristic values are explored based on the undervoltage fault data, and preliminary judgment is carried out on the undervoltage state according to the vehicle state of undervoltage occurrence, the SOC value and the distribution of the lowest highest temperature; extracting features by adopting a sliding window method; and constructing a PHM undervoltage fault prediction model of the energy storage battery by adopting a long-short term memory neural network (LSTM) algorithm. Through the data test of the real vehicle, the research result shows that the model has more than 90% of under-voltage fault prediction accuracy and recall rate of the chemical energy storage battery, and has more stable prediction accuracy. The method can help the target system to provide decision reference for preventive maintenance and repair, reduce maintenance cost and reduce the probability of fatal faults, and has important theoretical significance and application value for chemical energy storage battery research.

Description

Chemical energy storage battery PHM undervoltage fault prediction method

Technical Field

The invention belongs to the technical field of chemical energy storage battery fault prediction, and particularly relates to a PHM undervoltage fault prediction method for a chemical energy storage battery.

Background

In recent years, energy safety and environmental pollution have become factors restricting economic development, the development of energy storage technology has become a great trend of new energy development, and lithium ion batteries in chemical energy storage batteries have become research hotspots in recent years. System faults of the energy storage battery can be classified into three types according to the positions where the faults occur, namely battery body faults, signal transmission faults and line faults. Because the signal fault and the line fault belong to sudden faults, the fault cannot be accurately predicted by combining limited characteristic information. There are three types of energy storage battery body faults: open circuit fault, temperature imbalance and single voltage imbalance of the positive and negative electrode relays. The method is mainly used for detecting the condition that the voltage of the single body is inconsistent, namely the extreme difference and standard deviation of the voltage of the single body are more than 5%, and the faults can directly or indirectly influence the discharging efficiency of the energy storage battery and damage the service life of the energy storage battery.

At present, some prediction methods commonly existing in PHM (diagnostics and Health management) fault prediction comprise a deep learning method, a particle swarm optimization, a nonlinear correction ARIMA and LSTM combined data fusion method, an LSTM, an RAdam optimizer and the like, accurate data measured by a laboratory are required to be obtained for accurate prediction, vehicle collected data under actual working conditions are difficult to apply, and the following technical scheme is provided for solving the problems.

Disclosure of Invention

The invention aims to establish a chemical energy storage battery under-voltage fault prediction model with high precision and real-time performance.

The purpose of the invention can be realized by the following technical scheme:

a PHM undervoltage fault prediction method for a chemical energy storage battery comprises the following steps:

s1, preparing data, and acquiring use related data of the energy storage battery of the real vehicle and monitoring data of the running state of the vehicle;

the use related data of the energy storage battery of the real vehicle comprises full voltage difference data in the normal use process;

the monitoring data of the vehicle running state comprises a vehicle charging state and a vehicle running state;

s2, preprocessing data, cleaning the data of the full voltage difference, converting and extracting the monomer voltage and the temperature of a temperature measuring point, and correcting the current, the total voltage, the temperature and the early warning field;

s3, data exploration, wherein the characteristic analysis is carried out on the vehicle state and battery state data collected before and after the undervoltage occurrence, and the potential factors influencing the undervoltage fault of the energy storage battery are obtained by combining a visual data chart;

s4, extracting characteristics, namely extracting the characteristics by adopting a sliding window according to the potential factors of the undervoltage fault of the energy storage battery obtained by analyzing in the step S3;

s5, establishing a model, establishing a corresponding relation data set between the battery pressure difference state and the corresponding characteristic variable for the extracted characteristics, selecting an LSTM model in deep learning to sequentially perform data extraction, data division and component neural network, and establishing a PHM undervoltage prediction model;

the data extraction is to uniformly extract normal samples and abnormal samples from all vehicle data; the data division comprises dividing the sample data into a training set and a verification set; the component neural network comprises the steps of determining the number of layers and the number of nodes of the LSTM model;

s6, sample testing, verifying the test set data on the basis of establishing the prediction model, and determining the accuracy rate, recall rate and precision rate of the model;

and S7, real vehicle verification.

As a further aspect of the present invention, the data cleansing in step S2 includes: and carrying out numerical conversion on the abnormal value and the null value in the original data field.

As a further aspect of the present invention, the vehicle states before and after the occurrence of the under-voltage include both charged and uncharged states, and the battery state data includes an SOC value distribution, a lowest temperature distribution, and a highest temperature distribution.

As a further aspect of the present invention, the characteristics in step S4 include a lateral pressure difference, a difference between a longitudinal lowest voltage and a first window, an abnormal value characteristic, an under-voltage warning characteristic, and an SOC.

As a further aspect of the present invention, the abnormal value is characterized by marking data that has exceeded a threshold value and data that is abnormal as 1; the undervoltage early warning feature marks the data with the lowest voltage smaller than a preset value as 1.

As a further aspect of the present invention, the sliding window in step S4 is a sliding window of 30 data widths.

As a further aspect of the present invention, the LSTM model in step S5 is composed of two LSTM layers and 2 fully-connected layers and one output layer; the first LSTM layer has 256 nodes, the dropout ratio is 0.3, the second LSTM layer has 64 nodes, and the dropout ratio is 0.2; the first full-connection layer has 64 nodes, the second full-connection layer has 16 nodes, the drop ratio of the two full-connection layers is 0.4, and the activation functions are both tanh; the drop ratio of the second fully connected layer to the output layer is 0.4, and the activation function of the output layer is sigmoid.

The invention has the beneficial effects that:

(1) the method mainly aims at the undervoltage fault of the energy storage battery to establish a machine learning model to perform early warning of the undervoltage fault, predicts the time of the battery reaching the failure moment or a set threshold value from the current moment by using the known running state information, accurately realizes PHM prediction, can provide decision reference for preventive maintenance and repair for a target system, reduces the maintenance cost and reduces the probability of fatal fault;

(2) the undervoltage fault prediction method provided by the invention has the advantages that a large amount of real vehicle data are collected to serve as a data source, characteristics are extracted from battery state index parameters such as voltage, current, SOC (system on chip), insulation resistance and the like by utilizing a sliding window, and a positive sample test result and a negative sample test result are combined, so that the successful prediction alarm can be carried out on the common undervoltage fault problem of the chemical energy storage battery;

(3) aiming at the characteristics of a dynamic nonlinear electrochemical system, the method has the effects of high prediction precision and adaptation to the characteristic extraction and denoising conversion algorithm of vehicle data under the actual working condition;

(4) the method is practiced in an application scene of PHM undervoltage fault prediction of the real vehicle battery by using an LSTM algorithm in the traditional machine learning method, and the hidden battery undervoltage fault and the evolution law are mined from the battery rated information and the state monitoring data, so that the method has stronger theoretical significance and application value;

(5) the invention adopts a sliding window method to extract the features, which is superior to the feature dimension reduction and the traditional feature extraction method.

Drawings

The invention will be further described with reference to the accompanying drawings.

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a vehicle state ratio plot when under-voltage occurs;

FIG. 3 is a distribution diagram of SOC values when under-voltage occurs;

FIG. 4 is a graph of the lowest temperature profile when differential pressure occurs;

FIG. 5 is a graph of the maximum temperature profile when a pressure differential occurs;

FIG. 6 is a confusion matrix;

FIG. 7 is a normalized confusion matrix;

FIG. 8 shows the results of positive and negative sample tests;

fig. 9 is a predictive visualization.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

A chemical energy storage battery PHM undervoltage fault prediction method is shown in figure 1 and comprises the following steps:

s1, preparing data, wherein the step is located before the whole processing scheme and aims to acquire original data related to the under-voltage state of the chemical energy storage battery;

in this step, the original data related to the under-voltage state of the energy storage battery includes full voltage difference data of m actual running automobiles and monitoring data of a vehicle running state, the full voltage difference data is generated in the automobile charging process, and the monitoring data of the vehicle running state includes variable data such as a vehicle charging state and a vehicle running state;

the real vehicle battery use data and the vehicle running state monitoring data are streaming data based on time sequence;

in an embodiment of the present invention, the value of m is 45, the above-mentioned acquisition time span of the full voltage difference data is every month and every hour of a year, the same vehicle type, 45 vehicles, data lasting for 1 year are provided in total at this time as a data source for predicting the PHM undervoltage state;

s2, preprocessing data, and cleaning original data related to the under-voltage state of the energy storage battery and vehicle running state recorded data to improve data quality and facilitate subsequent data exploration and modeling;

in the collected real vehicle data set, because the original PHM undervoltage data has a lot of worthless data and can generate adverse effects on data exploration and modeling, a series of preprocessing work needs to be carried out on the original data to improve the data quality; the data set used by the invention is full voltage difference data of 45 vehicles, the main pretreatment process comprises deletion of null value, repeated value and abnormal time value, correction of alarm field, extraction of monomer voltage and temperature of temperature measuring point, and the like, and the specific treatment process is as follows:

processing null values; performing null check processing on all monomer voltage data and temperature measurement point data in the original data, and deleting data containing null values;

secondly, carrying out duplicate removal operation on the whole body; when each value of a row of data is repeated, the data is regarded as repeated data, and only one of all the repeated data is reserved in the deduplication operation;

processing time abnormal data; when the original time data is abnormal, such as a month is not within 1-12, an hour is not within 0-24, and the like, directly deleting the time data which is not within a certain range;

alarming and correcting; the alarm fields in the original data are all displayed in a decimal mode, but the alarm fields displayed in the available alarm condition comparison table which is taken at present are in a binary digit mode, so that the numerical values of the original alarm fields are converted into binary numerical values, and then all digits which are 1 in the binary system are counted;

extracting the monomer voltage and the temperature of a temperature measuring point; disassembling and extracting the monomer voltage and temperature data of the temperature measuring point in the stored list in the original data, and converting the monomer voltage and the temperature data into an expanded DataFrame form;

correcting the extreme value of the temperature and the voltage; considering that extreme values such as the highest temperature, the lowest temperature, the highest voltage, the lowest voltage and the like may not be consistent with the temperature of all the monomer voltages and the temperature measuring points, the following correction is carried out on the extreme values: replacing a voltage temperature extreme value in the original data by calculating a monomer voltage maximum value, a monomer voltage minimum value, a temperature measurement point maximum value and a temperature measurement point minimum value of each row of data;

correcting current, total voltage and temperature; the current, the total voltage and the temperature in the original data are not true values, and a certain conversion is needed to obtain an actual true value, and the conversion rule is as follows: dividing the current data by a value in the range of 10 to 1000, dividing the total voltage data by 10, and dividing the temperature data by 40;

s3, data exploration, wherein the step mainly comprises the steps of carrying out preliminary analysis on preprocessed data information, summarizing the characteristics and the influence factors of the undervoltage fault of the energy storage battery through data visualization analysis, and preparing for subsequent characteristic engineering;

in the step, data needs to be visualized, the visualization results are shown in fig. 2 to fig. 5, and it is obtained on the basis of the visualization of the graphs that the pressure difference fault of the energy storage battery mainly occurs in the first quarter, and the undervoltage is unrelated to the temperature; the distribution interval of the lowest voltage in the undervoltage process is basically between 1500 and 2800;

FIG. 2 shows that the undervoltage fault occurs primarily in the flameout condition, and a small number during charging; wherein the left rectangular area represents a charged state, the middle rectangular area represents an uncharged state, and the right rectangular area represents a state where charging is completed;

FIG. 3 shows that the SOC (important pressure differential signature) is mostly 0 when an under-voltage fault occurs;

fig. 4 and 5 show that the maximum temperature and the minimum temperature do not change significantly when the pressure difference occurs, so that the temperature characteristics are difficult to reflect whether the pressure difference occurs or not, and the temperature consideration can be reduced in the characteristic extraction process;

therefore, the summary of the undervoltage fault mainly includes the following two aspects:

firstly, a certain monomer voltage starts to abnormally decline at a certain time point, the monomer voltage starts to decline slowly and then declines more and more quickly until the monomer voltage falls below an undervoltage threshold value, and the undervoltage threshold value is a preset value;

one or more single voltage suddenly jumps to a very low value;

s4, extracting characteristics, summarizing and extracting the data obtained after preprocessing, and obtaining characteristic factors influencing the undervoltage fault;

in the step, as single data is easy to fluctuate and has low stability, the method adopts a sliding window mode to extract the characteristics, and the size of the sliding window selected by the differential pressure model is 30 data (the time span is about 5 minutes) by considering the time sequence characteristics of the data and the total amount of the vehicle differential pressure data;

preliminary analysis finds that the influence of the actual working condition on the PHM undervoltage fault can be effectively reflected by extracting the following characteristics: the difference value between each transverse pressure difference and the first window of the longitudinal lowest voltage, the abnormal value characteristic, the under-voltage early warning characteristic and the SOC. The transverse voltage difference refers to the reduction of the maximum voltage to the minimum voltage at the same moment; the abnormal value characteristic refers to marking the data which exceeds a threshold value and the data with obvious data abnormality as 1; the undervoltage early warning feature marks data with the lowest voltage of <3300mv as 1;

s5, establishing a model, namely establishing a chemical energy storage battery PHM undervoltage fault prediction model based on the data after feature extraction;

in one embodiment of the invention, an LSTM model in deep learning is selected for under-voltage prediction, on one hand, the LSTM can be well used for predicting sequence data, and on the other hand, due to the specific structure of the model in the deep learning field, the prediction accuracy is higher;

firstly, data extraction is needed, fragments with rising temperature difference before a fault occurs are extracted, all data are marked as 1, the fault fragments are removed (because the fault occurs and does not need prediction), and other data are marked as 0; the algorithm aims to correctly classify all data, if a certain segment has a large number of labels with the algorithm prediction of 1, the segment is the segment about to fail, and since the positive and negative samples in the real vehicle data are extremely unbalanced in distribution, a proper proportion of data samples are extracted, so that normal samples and abnormal samples are uniformly extracted from all vehicles; the number of normal samples is 30 ten thousand, and the number of abnormal samples is 2 ten thousand;

after data extraction, data division is needed, the extracted samples are divided into a training set and a verification set, and the proportion is as follows: the verification set is 8: 2;

constructing a neural network after data division: the LSTM model is composed of two LSTM layers, 2 full-connection layers and an output layer, the first LSTM layer is provided with 256 nodes, the dropout proportion is 0.3, the second LSTM layer is provided with 64 nodes, the dropout proportion is 0.2, the first full-connection layer is provided with 64 nodes, the second full-connection layer is provided with 16 nodes, the drop proportion of the two full-connection layers is 0.4, and the activation functions are tanh;

the drop ratio of the second full connection layer to the output layer is 0.4, and the activation function of the output layer is sigmoid;

s6, sample testing, random selection 100: 1, testing positive and negative samples;

in this step, 100 random extractions are performed first (negative samples are extracted 1/5 in all negative samples each time, positive samples are extracted 100 times of data in all positive samples), then accuracy and recall tests are performed, the result of each time is counted, and finally the average result, the maximum result and the minimum result are counted, wherein the confusion matrix and the standardized confusion matrix are shown in fig. 6 to 7, and the positive and negative sample test results are shown in fig. 8;

and S7, according to the final algorithm model, carrying out experimental tests on the actual 45 vehicles according to the sliding window, and analyzing the results. After the vehicle data used for training is removed, the part carries out the verification of the real vehicle data, and the specific results are shown in the following table and fig. 9;

index (I)	Mean value	Minimum value	Maximum value
				Rate of accuracy	0.987	0.986	0.988
Recall rate	0.988	0.976	0.996

The left rectangular area shown in fig. 9 represents an algorithm prediction area, and the right rectangular area is an actual window area where a differential pressure fault occurs, and the model provided by the invention can predict the undervoltage fault of the energy storage battery in advance by about 20 minutes. The conclusion is drawn by combining the above table, and the average value of the prediction accuracy can reach 98.7%.

In the description herein, references to the description of "one embodiment," "an example," "a specific example" or the like are intended to mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

The foregoing is illustrative and explanatory only and is not intended to be exhaustive or to limit the invention to the precise embodiments described, and various modifications, additions, and substitutions may be made by those skilled in the art without departing from the scope of the invention or exceeding the scope of the claims.

Claims

1. A chemical energy storage battery PHM undervoltage fault prediction method is characterized by comprising the following steps:

the data extraction is to uniformly extract normal samples and abnormal samples from all vehicle data; the data division comprises the steps of dividing sample data into a training set and a verification set; the component neural network comprises the steps of determining the number of layers and the number of nodes of the LSTM model;

s6, testing the sample, verifying the data of the test set on the basis of establishing the prediction model, and determining the accuracy rate, the recall rate and the precision rate of the model;

and S7, real vehicle verification.

2. The PHM undervoltage fault prediction method of the chemical energy storage battery according to claim 1, wherein the data cleaning in the step S2 comprises: and carrying out numerical conversion on the abnormal value and the null value in the original data field.

3. The PHM undervoltage fault prediction method of the chemical energy storage battery as claimed in claim 1, wherein the vehicle states before and after the undervoltage occurrence include a charged state and a non-charged state, and the battery state data includes an SOC value distribution, a lowest temperature distribution and a highest temperature distribution.

4. The PHM undervoltage fault prediction method of the chemical energy storage battery according to claim 1, wherein the characteristics in the step S4 include a transverse pressure difference, a longitudinal lowest voltage and first window difference value, an abnormal value characteristic, an undervoltage early warning characteristic and an SOC.

5. The PHM undervoltage fault prediction method of the chemical energy storage battery according to claim 4, characterized in that the abnormal value is characterized by marking data which has exceeded a threshold value and data which is abnormal as 1; the undervoltage early warning feature marks the data with the lowest voltage smaller than a preset value as 1.

6. The PHM undervoltage fault prediction method of the chemical energy storage battery according to claim 5, wherein the sliding window in the step S4 is a sliding window with 30 data widths.

7. The PHM undervoltage fault prediction method of the chemical energy storage battery according to claim 6, wherein the LSTM model in step S5 is composed of two LSTM layers, 2 full-connection layers and one output layer; the first LSTM layer has 256 nodes, the dropout ratio is 0.3, the second LSTM layer has 64 nodes, and the dropout ratio is 0.2; the first full-connection layer has 64 nodes, the second full-connection layer has 16 nodes, the drop ratio of the two full-connection layers is 0.4, and the activation functions are both tanh; the drop ratio of the second fully connected layer to the output layer is 0.4, and the activation function of the output layer is sigmoid.