A new energy vehicle battery spontaneous combustion warning method and device based on deep learning

Abstract

The embodiment of the present invention discloses a new energy vehicle battery spontaneous combustion warning method based on deep learning, including: according to the historical operation data of the new energy vehicle, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation; according to the operation data of the predetermined period, an LSTM classifier is established and trained, and the optimal model parameters are obtained through cross-validation; the total voltage, the highest battery value, the highest temperature value, and the lowest temperature value of the target new energy vehicle are obtained; the change of each eigenvalue relative to the mean of the training data is calculated, and each is used as a new eigenvalue; the LSTM classifier outputs the abnormality based on a continuous period of eigenvalue data before the current eigenvalue data, and the abnormal value appears continuously and the abnormality continues to rise to a predetermined number, then it is determined whether the battery is abnormal. A CNN model can also be established and trained, and CNN is used to determine whether the battery is abnormal according to the abnormality change diagram. It can accurately and timely predict battery failures and solve the problem of inaccurate results caused by less battery spontaneous combustion data.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy vehicle battery detection, and in particular to a new energy vehicle battery spontaneous combustion early warning method and device based on deep learning. Background Art

[0002] The market size of new energy vehicles is getting bigger and bigger. The number of vehicles in use is huge and growing rapidly. The key component of new energy vehicles is the power battery. The health of the power battery has a significant impact on the vehicle's power performance and the driver's experience. Therefore, the correct prediction of the power battery health can help to judge the vehicle's power and economy. At present, in the field of power batteries, in order to predict the health of power batteries, the battery performance analysis and simulation are carried out by building a battery physical model. In the face of many external interference factors, it can meet the needs when predicting the health of vehicle power batteries under realistic and complex road conditions.

[0003] With the widespread use of new energy vehicles, the safety issues of power batteries have become more prominent. It is very time-consuming and laborious to use existing technical detection methods to provide early warning and early prevention of power battery failures, and it is difficult to implement in practice. The traditional power battery failure early warning method is to detect obvious changes in the voltage, current, pressure difference, and internal resistance of the battery cell itself through hardware, and issue an alarm when the warning threshold is reached, which has poor applicability for early prevention.

[0004] Industry professionals have proposed a variety of solutions. For example, Solution 1: "A Health Prediction Method and System for New Energy Vehicle Batteries" provides a method that can not only efficiently clean, convert and reduce the real-time vehicle operating data, but also mine the potential relationship between battery health data and vehicle operating data, and build a battery monitoring health prediction model related to vehicle operating data, so as to achieve dynamic prediction of the battery health status of new energy vehicles without being restricted by the high complexity of road conditions. The main method is to obtain vehicle operating data based on big data analysis; pre-process the data, and then use the data mining model built with random forest and RBF neural network algorithms to predict battery health.

[0005] Solution 2: "A method and device for online battery fault detection and analysis of new energy vehicles" uses historical data to train deep learning algorithms LSTM (Long Short-Term Memory) and support vector machines SVM (Support Vector Machine), as well as softmax multi-classifiers, to determine whether the battery is faulty and the type of fault. When this technology is put into use, the vehicle sensor data is input, and the fault diagnosis is first performed through the LSTM model, and then the sensor fault and vehicle battery fault are distinguished through the SVM model. Subsequently, the sensor or vehicle battery fault type is classified and located through two softmax multi-classification models, thereby achieving online and accurate fault detection.

[0006] Solution 3: "A battery failure warning method and system based on big data"

[0007] Capture massive data of the target platform; process and store the massive data based on a distributed system architecture; retrieve and analyze the massive data to obtain data analysis results for at least one battery fault category; establish at least one battery fault warning target model based on the data analysis results; obtain target data of a single vehicle, and perform predictive analysis on the target data through at least one battery fault warning target model to obtain a battery fault prediction result for the single vehicle; obtain battery warning information based on the battery fault prediction result, and send the battery warning information to the terminal.

[0008] In the prior art, insufficient data on spontaneous combustion of new energy vehicle batteries results in a large difference in the amount of data between battery abnormalities and non-abnormal data. The difference in data often leads to extremely unbalanced positive and negative sample data in the training data, which in turn leads to poor prediction performance of the battery spontaneous combustion warning device. Therefore, there is an urgent need to provide a warning solution that can accurately and timely predict battery failures. Summary of the invention

[0009] In view of this, an embodiment of the present invention provides a new energy vehicle battery spontaneous combustion warning method and device based on deep learning to overcome some defects and shortcomings in the prior art and to accurately and timely predict battery failures.

[0010] In order to achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0011] An embodiment of the present invention provides a new energy vehicle battery spontaneous combustion warning method based on deep learning, comprising the following steps:

[0012] S10: Based on the historical operation data of new energy vehicles, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation; based on the operation data of the scheduled period, an LSTM classifier is established and trained, and the optimal model parameters are obtained through cross-validation;

[0013] S11: Obtain the current operating data of the target new energy vehicle and extract the required four characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0014] S12: Determine whether the four characteristic value data are valid. If they are invalid data, remove them and re-acquire new data; if they are valid data, proceed to the next step;

[0015] S13: Calculate the change of each eigenvalue relative to the mean of the training data and use them as new eigenvalues ​​to obtain eight eigenvalues;

[0016] S14: using the decision tree to determine whether the characteristic value data is extreme data, if it is extreme data, it means there is no abnormality, and the extreme data is data with low characteristic values;

[0017] S15: Read a continuous section of feature value data before the current feature value data, and input it into the LSTM classifier. The LSTM classifier outputs an abnormality degree. If the abnormal values ​​output by the LSTM classifier appear continuously and the abnormality degree continues to increase to a predetermined amount, the battery is determined to be abnormal.

[0018] Another new energy vehicle battery spontaneous combustion warning method based on deep learning provided by an embodiment of the present invention includes the following steps:

[0019] S20a: Obtain historical operating data of the target new energy vehicle and establish and train a decision tree model based on the historical operating data, and obtain optimal model parameters through cross-validation;

[0020] S20b: According to the operation data of the predetermined period, establish and train the LSTM classifier, and obtain the optimal model parameters through cross-validation;

[0021] S20c: According to the abnormality change graph obtained by the LSTM classifier, a CNN model is established and trained, and the optimal model parameters are obtained through cross-validation;

[0022] S21: Obtain the operating data of the target new energy vehicle at the current moment, and extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0023] S22: Determine whether the data of the four characteristic values ​​are valid, if they are valid data, proceed to the next step;

[0024] S23: Calculate the change of each eigenvalue relative to the mean of the training data and use them as new eigenvalues ​​to obtain eight eigenvalues;

[0025] S24: using a decision tree to determine whether the characteristic value data is extreme data, if it is extreme data, it means there is no abnormality, and the extreme data is data with low characteristic values;

[0026] S25: read a section of continuous data before the current data, input it into the LSTM classifier, calculate the abnormality and construct an abnormality change graph;

[0027] S26: Use CNN to determine whether the battery is abnormal based on the abnormality change graph. If it is abnormal, it is determined that the vehicle is at risk of battery spontaneous combustion at the current moment and a warning signal is issued.

[0028] An embodiment of the present invention provides a new energy vehicle battery spontaneous combustion warning device based on deep learning, comprising:

[0029] The decision tree module builds and trains a decision tree model based on the historical operation data of new energy vehicles, and obtains the optimal model parameters through cross-validation; it is used to determine whether the operation data is extreme data;

[0030] LSTM classifier, used to establish and train LSTM classifier according to historical operation data of a predetermined period, and obtain the optimal model parameters through cross-validation; used to calculate the abnormality at a certain moment and construct an abnormality change graph;

[0031] The data acquisition module is used to obtain the operating data of the target new energy vehicle at the current moment and extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0032] A data identification module is used to determine whether the data of the four characteristic values ​​is valid. If it is invalid data, the data is removed and the previous step is re-executed to obtain new data; if it is valid data, the next step is continued;

[0033] A calculation module is used to calculate the change of each eigenvalue relative to the mean of the training data, and use them as new features to obtain eight eigenvalues;

[0034] If the decision tree module determines that the eigenvalue data is extreme data, it means that there is no abnormality, and the extreme data is data with low eigenvalues;

[0035] A continuous segment of feature value data before the current feature value data is input into the LSTM classifier, and the LSTM classifier outputs an abnormality degree. If the abnormal values ​​output by the LSTM classifier appear continuously and the abnormality degree continues to rise to a predetermined amount, the battery is determined to be abnormal.

[0036] Another new energy vehicle battery spontaneous combustion warning device based on deep learning provided by an embodiment of the present invention includes:

[0037] Decision tree module: Based on the historical operation data of new energy vehicles, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation; it is used to determine whether the operation data is extreme data;

[0038] LSTM classifier: Based on the historical operation data of the predetermined period, the LSTM classifier is established and trained, and the optimal model parameters are obtained through cross-validation; it is used to calculate the abnormality at a certain moment and construct the abnormality change graph;

[0039] The CNN module builds and trains the CNN model based on the abnormality change graph obtained by the LSTM classifier, and obtains the optimal model parameters through cross-validation to determine whether the battery is abnormal;

[0040] The data acquisition module obtains the operating data of the target new energy vehicle at the current moment and extracts the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0041] The data identification module determines whether the data of the four characteristic values ​​is valid. If it is invalid data, the data is removed and the previous step is re-executed to obtain new data; if it is valid data, the next step is continued;

[0042] The calculation module calculates the change of each eigenvalue relative to the mean of the training data and uses them as new features to obtain eight eigenvalues.

[0043] If the decision tree determines that the eigenvalue data is extreme data, it means that there is no abnormality, and the extreme data is data with low eigenvalues;

[0044] Input a section of continuous data before the current data into the LSTM classifier, calculate the anomaly degree and construct an anomaly degree change graph;

[0045] The CNN module determines whether the battery is abnormal based on the abnormality change diagram. If it is abnormal, it determines that the vehicle is at risk of battery spontaneous combustion at the current moment and sends a warning signal.

[0046] The embodiments of the present invention have the following advantages:

[0047] The present invention can solve the problem of inaccurate results caused by less battery spontaneous combustion data by using a single classification model, which makes it easier to put the invention into use. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0049] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0050] Figure 1 A flow chart of a new energy vehicle battery spontaneous combustion warning method provided by an embodiment of the present invention;

[0051] Figure 2 A flow chart of another new energy vehicle battery spontaneous combustion warning method provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the LSTM classifier principle of an embodiment of the present invention;

[0053] Figure 4 The total voltage change before removing invalid data for a vehicle with normal battery in an embodiment of the present invention;

[0054] Figure 5 The total voltage change of the vehicle with normal battery in the embodiment of the present invention;

[0055] Figure 6 The maximum battery value change of the vehicle with normal battery in the embodiment of the present invention;

[0056] Figure 7 The maximum temperature change of the vehicle with normal battery in the embodiment of the present invention;

[0057] Figure 8 The minimum temperature value change of the vehicle with normal battery in the embodiment of the present invention;

[0058] Fig. 9 The abnormality degree change of the battery abnormal vehicle in the embodiment of the present invention;

[0059] Fig.10 The abnormality degree change of the vehicle with normal battery in the embodiment of the present invention;

[0060] Fig.11A schematic diagram of the architecture of a new energy vehicle battery spontaneous combustion warning device provided by an embodiment of the present invention;

[0061] Fig.12 A schematic diagram of the architecture of another new energy vehicle battery spontaneous combustion warning device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] Reference Figure 1 , an embodiment of the present invention provides a new energy vehicle battery spontaneous combustion warning method based on deep learning, comprising the following steps:

[0064] S10: Based on the historical operation data of new energy vehicles, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation; based on the operation data of the scheduled period, an LSTM (Long Short-Term Memory) classifier is established and trained, and the optimal model parameters are obtained through cross-validation;

[0065] S11: Obtain the current operating data of the target new energy vehicle and extract the required four characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0066] S12: Determine whether the four characteristic value data are valid. If they are invalid data, remove them and reacquire new data; if they are valid data, proceed to the next step;

[0067] S13: Calculate the change of each eigenvalue relative to the mean of the training data and use them as new eigenvalues ​​to obtain eight eigenvalues;

[0068] S14: using the decision tree to determine whether the characteristic value data is extreme data, if it is extreme data, it means there is no abnormality, and the extreme data is data with low characteristic values;

[0069] S15: Read a continuous section of feature value data before the current feature value data, and input it into the LSTM classifier. The LSTM classifier outputs an abnormality degree. If the abnormal values ​​output by the LSTM classifier appear continuously and the abnormality degree continues to increase to a predetermined amount, the battery is determined to be abnormal.

[0070] In a specific embodiment of the present invention, the long short-term memory network LSTM classifier used is used to classify the abnormality by the distance value from the feature point to the surface of the hypersphere. Of course, other methods can also be used as the abnormality for classification, such as using the distance value from the feature point to the hyperplane as the abnormality for classification; or, other time series model classifiers can be used, such as a recurrent neural network RNN ​​classifier.

[0071] In this embodiment, the LSTM classifier outputs the difference between the distance from the feature point to the center of the sphere and the radius. The difference indicates the degree of abnormality, and the larger the difference, the greater the degree of abnormality. If the abnormal points output by the LSTM classifier appear continuously and the number of abnormal points whose degree of abnormality continues to increase reaches a predetermined value (for example, 20), the battery is determined to be abnormal.

[0072] like Figure 2 As shown, another new energy vehicle battery spontaneous combustion warning method based on deep learning is provided in an embodiment of the present invention, and the steps are as follows:

[0073] S20a: Obtain historical operating data of the target new energy vehicle and establish and train a decision tree model based on the historical operating data, and obtain optimal model parameters through cross-validation;

[0074] S20b: According to the operation data of the predetermined period, establish and train the LSTM classifier, and obtain the optimal model parameters through cross-validation;

[0075] S20c: According to the abnormality change graph obtained by the LSTM classifier, a CNN (Convolutional Neural Networks) model is established and trained, and the optimal model parameters are obtained through cross-validation;

[0076] S21: Obtain the operating data of the target new energy vehicle at the current moment, and extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0077] S22: Determine whether the data of the four characteristic values ​​are valid, if they are valid data, proceed to the next step;

[0078] S23: Calculate the change of each eigenvalue relative to the mean of the training data and use them as new eigenvalues ​​to obtain eight eigenvalues;

[0079] S24: using a decision tree to determine whether the characteristic value data is extreme data, if it is extreme data, it means there is no abnormality, and the extreme data is data with low characteristic values;

[0080] S25: read a section of continuous data before the current data, input it into the LSTM classifier, calculate the abnormality and construct an abnormality change graph;

[0081] S26: Use CNN to determine whether the battery is abnormal based on the abnormality change graph. If it is abnormal, it is determined that the vehicle is at risk of battery spontaneous combustion at the current moment and a warning signal is issued.

[0082] In order to make the principles, characteristics and advantages of the present invention more clear, they are described in detail below in conjunction with specific implementation schemes.

[0083] S100: Based on the historical operation data at a certain moment, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation. Based on the historical operation data over a period of time, an LSTM classifier is established and trained, and the optimal model parameters are obtained through cross-validation. Based on the abnormality change graph obtained from the LSTM classifier results, a CNN model is established and trained, and the optimal model parameters are obtained through cross-validation.

[0084] In this embodiment:

[0085] Establish and train a decision tree model, and obtain the optimal model parameters through cross-validation. Specifically, the decision tree model is established and trained using the extreme gradient boosting algorithm XGBoost, mainly by calling the native Python API interface of the XGBoost library to establish the model, where the objective parameter is set to: binary:logistic, for binary classification logistic regression tasks. Label the acquired data, use each piece of data and its label value as input, and obtain its classification result through the decision tree model. The decision tree model is trained with the gap between the true value and the predicted value as the training objective function, and the training data is used to perform cross-validation optimization. The optimal parameters of the decision tree model are determined on the basis of satisfying the optimal evaluation index of the prediction model, and the optimized model is re-established on this basis to obtain the decision tree model. Among them, the method of labeling the acquired data is mainly determined according to the change law and numerical distribution of each eigenvalue. Specifically, for the eight eigenvalues ​​obtained, the data that simultaneously meets the conditions that the total voltage is less than 550V, the maximum battery value is less than 3.2V, the maximum temperature value is less than 25°C and the minimum temperature value is less than 20°C will be marked as 0, indicating that there is no abnormality; the others will be marked as 1, indicating that the danger cannot be judged and the next step of detection is required.

[0086] Establish and train the LSTM classifier, and obtain the optimal model parameters through cross-validation. Specifically, the following steps are used: call the keras library to build the LSTM neural network, initialize the center and radius to form a hypersphere feature space, and establish an LSTM classifier that maps the input data from the original space to the hypersphere feature space. For the operating data of vehicles with no battery anomalies, the 60 consecutive data [x1, x2, ..., x60] before the current data are used as input, and the feature points of the current operating data are obtained through the LSTM classifier. The objective function of the training is to minimize the hypersphere that can include all feature points, train the LSTM classifier, and use the training data to perform cross-validation optimization on it. The optimal parameters of the LSTM classifier are determined on the basis of meeting the optimal evaluation index of the prediction model, and the optimized model is re-established on this basis to obtain the LSTM classifier. The distance value from the obtained feature point to the surface of the hypersphere is used as the abnormality, and a point is added to the abnormality change graph.

[0087] Establish and train the CNN model, and obtain the optimal model parameters through cross-validation. Specifically, the following steps are used: label a series of abnormality change graphs obtained by the LSTM classifier, and generate a list of image paths and corresponding labels. The convolutional neural network is built by calling Tensorflow, and the MBGD gradient descent algorithm and the cross entropy loss function are used as optimizers to establish the CNN model. The abnormality change graph and its label value are used as inputs, and convolution operations are performed. The CNN model is trained with the objective function of minimizing the gap between the true value and the predicted value. The training data is used to perform cross-validation optimization on it, and the optimal parameters of the CNN model are determined on the basis of satisfying the optimal evaluation index of the prediction model. Based on this, the optimized model is re-established to obtain the CNN model.

[0088] Among them, the method of labeling the abnormality change graph is: the abnormality change graph obtained from the operating data of the battery abnormal vehicle is represented by 1, and the graph obtained from the battery normal vehicle is represented by 0.

[0089] S101: Obtain the operating data of the target new energy vehicle at the current moment, and extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value.

[0090] S102: Determine whether the data is valid. If it is invalid data, remove the data and re-execute the previous step to obtain new data; if it is valid data, proceed to the next step.

[0091] In this embodiment, in the acquired operating data of the new energy vehicle, invalid data that exceeds the valid range due to collection abnormalities appears. In order to avoid its impact on the detection results, it is necessary to determine whether the data is valid before detection. The specific method is: for the extracted data, there are always four characteristics: the total voltage exceeds the valid range of 0V~1000V and is invalid data; the maximum battery value exceeds the valid value range of 0V~15V and is invalid data; the maximum temperature value and the minimum temperature value exceed the valid range of -40℃~210℃ and are invalid data.

[0092] S103: Calculate the change of each feature relative to the mean of the training data and use them as new features to obtain eight feature values.

[0093] In this embodiment, the mean of the training data refers to the mean of the four features calculated based on all the training data. The calculation method of the new features is: taking the mean obtained from the training data as the standard, respectively calculating the difference between the four features of the data and the corresponding mean of each feature, and the four differences obtained are the new features, which are used to represent the changes of each feature.

[0094] S104: Use a decision tree to determine whether the current data is extreme data. If it is extreme data, it means there is no abnormality, and it is determined that the battery has no risk of spontaneous combustion and no warning is given; otherwise, proceed to the next step of detection.

[0095] In this embodiment, the data obtained through the above steps is input for detection, and the feature space is divided to determine whether it is extreme data, such as low temperature data. If it is extreme data, it means there is no abnormality, and it is determined that the battery has no risk of spontaneous combustion, and no warning is given; otherwise, the next step of detection is carried out.

[0096] S105: Read 60 consecutive data before the current data, input them into the LSTM classifier for judgment, calculate the abnormality and add a point in the abnormality change graph.

[0097] In this embodiment, since there is little driving data on vehicles with spontaneous battery combustion and it is generally difficult to obtain, there is often a large difference in the amount of data between abnormal data and non-abnormal data. To solve this problem, an LSTM classifier is used, which can directly build a model and train it through the operating data of vehicles with normal batteries, and has a good detection effect for the situation of imbalanced positive and negative samples.

[0098] S105a: Read 60 consecutive data before the current data to obtain a time series with 8 features and a length of 60;

[0099] S105b: Input the read data into the LSTM classifier for judgment, calculate the abnormality of the data, and add a point to the abnormality change graph.

[0100] S106: Use CNN to determine whether the battery is abnormal based on the abnormality change graph. If there is no abnormality, it is determined that the battery has no risk of spontaneous combustion and no warning is given. If there is an abnormality, it is determined that the vehicle currently has a risk of spontaneous combustion and the driver is immediately warned.

[0101] Based on the training and detection methods of S100 to S106, the risk of battery spontaneous combustion of new energy vehicles at the current moment can be detected. The detection method can extract features from real-time new energy vehicle operation data, and then identify whether the vehicle has the risk of battery spontaneous combustion at the current moment after deep learning model analysis. Therefore, based on this method, a new energy vehicle battery spontaneous combustion warning device based on deep learning can be further provided, and the basic modules of the device include a data processing device and a battery spontaneous combustion warning device.

[0102] The data processing device is mainly used to process the obtained vehicle operation data. The operation data includes vehicle number, data recording time, total voltage, maximum battery value, maximum temperature value, and minimum temperature value. The data processing device sends the collected data to the data processing device in real time or at a fixed time through remote transmission methods such as GPRS, WIFI, and Bluetooth.

[0103] The data processing device can be in the form of a server, cloud platform, etc. Its function is to implement the training and detection methods described in S0 to S6 above, process the massive data sent by different new energy vehicles, and identify vehicles with the risk of battery spontaneous combustion.

[0104] The battery spontaneous combustion warning device is used to warn vehicles with battery spontaneous combustion risks identified by the data processing device. It can be installed on the terminal of new energy vehicles to remind drivers whether their vehicle batteries have the risk of spontaneous combustion; it can also be installed on the remote monitoring platform for supervisors to view and then remind drivers. The battery spontaneous combustion warning device can be hardware that has an audible, visual, and tactile reminder effect on personnel, such as a buzzer, warning light, etc., or it can be a software module that can pop up a reminder on the control interface.

[0105] In order to enable those skilled in the art to better understand the specific implementation of the present invention, the implementation process is further described below through embodiments.

[0106] In this embodiment, data is collected from new energy vehicles that have been on the market and are equipped with data collection devices. The data is driving data for one month. The data is in days, and the amount of data per day is about several thousand; the time interval between each data is 10 seconds. Since most data items have a low degree of distinction between whether the battery is abnormal, four attributes are finally selected for judgment based on the statistics of the data. Among them, the data of vehicles with normal batteries are relatively stable, and the amplitude of change over time is small, while the amplitude of change of data of vehicles with abnormal batteries is often very large. Therefore, in addition to extracting these four attributes, the change of their relative mean is also calculated and used as a new feature for subsequent analysis. Experiments have proved that the extraction of these newly added features has a good effect on the detection results.

[0107] The specific detection method flow is described in detail below. The basic training and detection method of this embodiment is as described in S0 to S6 above, and the specific steps are not repeated here. The specific implementation and technical effects of each step are mainly described.

[0108] Step 1. Read in real new energy vehicle data

[0109] The operating data obtained from new energy vehicles with normal battery conditions is imported into the database to extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value.

[0110] Figure 4 The total voltage change of the vehicle with normal battery before removing invalid data in the embodiment of the present invention;

[0111] Figures 5 to 8 They are respectively the changes of the total voltage, the highest battery value, the highest temperature value, and the lowest temperature value of the vehicle with no battery abnormality in the embodiment of the present invention.

[0112] Step 2. Determine whether the data is valid and process it, calculate the feature changes to obtain eight feature values

[0113] Vehicles with a risk of battery spontaneous combustion refer to vehicles with characteristic values ​​that tend to increase and change dramatically during driving compared to vehicles with normal batteries. Therefore, when the data is generally high and exceeds the normal value, it is very likely that the vehicle battery is abnormal and there is a risk of battery spontaneous combustion. Therefore, the selection of characteristic data greatly affects the final test results. The main steps for data processing are as follows:

[0114] 2.1. Determine whether the acquired data is valid

[0115] Since the operation data of the new energy vehicle acquired in this embodiment has invalid data that exceeds the valid range due to abnormal collection, in order to avoid its impact on the test results, it is necessary to determine whether the data is valid before testing. The specific method is as follows: for the extracted data, among the four features, the total voltage exceeds the valid range of 0V~1000V and is invalid data, the highest battery value exceeds the valid value range of 0V~15V and is invalid data, and the highest temperature value and the lowest temperature value exceed the valid range of -40℃~210℃ and are invalid data. Figure 2 As shown, the invalid data is determined based on the total voltage change of a vehicle with normal battery.

[0116] If it is invalid data, remove the data and re-execute the previous step to obtain new data; if it is valid data, proceed to the next step.

[0117] 2.2. Feature extraction and addition

[0118] After the operation in the previous step, the interference of invalid data can be avoided. Since the amplitude of the four extracted eigenvalues ​​changes with time, there is a relatively obvious difference between vehicles with abnormal batteries and vehicles without abnormal batteries, that is, the data of vehicles with normal batteries has a small change amplitude and is relatively stable; while the eigenvalues ​​of vehicles with abnormal batteries will change greatly as the vehicle travels, so it is necessary to calculate the amplitude of the change as the newest feature for analysis, and use them as new features to obtain eight eigenvalues. The specific calculation method is: both the training data and the test data are based on the mean of each feature of the training data, and the difference between the four features of the data and the corresponding mean of each feature is calculated respectively. The four differences obtained are the new features, which are used to represent the changes of each feature. In this way, it can ensure that the calculation of changes in the two types of data has a unified standard.

[0119] Step 3. Label and divide the data

[0120] 3.1. Labeling Data

[0121] After the above steps, valid data with eight characteristic values ​​can be obtained, and each piece of data can be labeled accordingly. The specific method is: each piece of data whose characteristics simultaneously meet the conditions that the total voltage is less than 550V, the maximum battery value is less than 3.2V, the maximum temperature value is less than 25℃ and the minimum temperature value is less than 20℃ will be marked as 0, indicating that there is no abnormality; the others will be marked as 1, indicating that the danger cannot be judged and the next step of detection is required.

[0122] 3.2. Partitioning the data

[0123] The data obtained from new energy vehicles with no battery abnormalities are divided into training sets and cross-validation sets, of which the training set accounts for 70% and is used to train each model, and the cross-validation set accounts for 30% and is used to adjust the model parameters.

[0124] Step 4. Build a decision tree model and perform training and cross-validation

[0125] According to the historical operation data at a certain moment, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation. The specific operation is: using the extreme gradient boosting algorithm XGBoost, to establish and train the decision tree model, mainly by calling the native Python API interface of the XGBoost library, where the objective parameter is set to: binary:logistic, which is used for binary classification logistic regression tasks. For the decision tree model, each data and its label value are used as input, and the input data passes through 8 regression trees to obtain the final prediction result Where k refers to the number of trees, and in this model, k=8. The model adds L1 and L2 regularization terms to improve the generalization ability of the model, and takes minimizing the gap between the true value and the predicted value as the training objective function, so that the predicted value of the tree group is as close to the true value as possible, and finally obtains its classification result. The decision tree model is trained on the training set, and cross-validation is used in each round of boosting iteration. It is convenient to obtain the optimal number of boosting iterations, and finally determine the optimal parameters of the decision tree model on the basis of satisfying the optimal evaluation index of the prediction model. On this basis, the optimized model is re-established to obtain the decision tree model. The decision tree is used to determine whether a piece of data is valid. Specifically, the detected data is classified by dividing the feature space. If it is 0, it is represented as extreme data, and it can be determined that the battery has no risk of spontaneous combustion at this time, and no warning is given; if it is 1, it is represented as non-extreme data, and the next step of detection is required.

[0126] Step 5. Build an LSTM classifier and perform training and cross-validation

[0127] According to the historical operation data of a certain period of time, an LSTM classifier is established and trained, and the optimal model parameters are obtained through cross-validation. The specific operations are as follows: Build an LSTM neural network to extract the implicit time series features in the historical time series, and establish a hypersphere feature space in the output layer to map the obtained feature vector from the original space to the minimum volume hypersphere for classification of positive and negative samples. The LSTM classifier is as follows: Figure 3As shown. For the LSTM classifier, for the operation data of vehicles with normal battery, the 60 consecutive data [x1, x2, ..., x60] before the current data are used as input. After the LSTM network extracts the data features, the feature points of the current operation data are obtained. Among them, the model uses two layers of hidden layer neurons with 128 and the activation function is relu. A Dropout layer is added after each LSTM layer to prevent the network from overfitting. Finally, the fully connected layer Dense outputs 8 vectors to obtain the output set F∈R p For an input set X∈R d , the network mapping is recorded as For the classification part, the objective function of training is to minimize the smallest hypersphere that can encompass all feature points, and the learning weights W = {w 1 ,…,w 1}, l is the number of hidden layers, the center of the sphere C∈F, the radius R>0, and the objective function is:

[0128]

[0129] Among them, the first formula minimizes R to minimize the volume of the hypersphere, the second formula is the penalty factor for feature points outside the hypersphere, and the third formula is the regularization term of the network weight. The model maps the original data as close to the center c as possible by training the learning parameter W on the training set. It is cross-validated and optimized on the cross-validation set, and the optimal parameters of the LSTM classifier are determined on the basis of meeting the optimal evaluation index of the prediction model. The optimized model is re-established on this basis to obtain the LSTM classifier. The LSTM classifier is used to calculate the abnormality at a certain moment and add a point to the abnormality change graph. Specifically, the feature points of the current data are obtained by inputting 60 time-continuous data before the current data into the LSTM classifier, and the distance value from the feature point to the surface of the hypersphere is calculated as the abnormality, and a point is added to the abnormality change graph.

[0130] Step 6. Label and divide the abnormality change graph

[0131] 6.1. Label the abnormality change graph. Label the abnormality change graph obtained by LSTM classification detection with corresponding labels. Since the above data are all obtained from new energy vehicles with normal battery, the abnormality change graph obtained by LSTM classification detection is marked as 0, indicating that the battery is normal; according to the rules of the existing images, a part of the images are artificially generated and marked as 1, indicating that the battery is abnormal.

[0132] 6.2. Dividing image data

[0133] The labeled images are divided into a training set and a cross-validation set, where the training set accounts for 70% and is used to train the decision tree model, and the cross-validation set accounts for 30% and is used to adjust the model parameters.

[0134] Step 7. Build a CNN model and perform training and cross-validation

[0135] According to the image data, a CNN model is established and trained, and the optimal model parameters are obtained through cross-validation. The specific operation is: label a series of abnormality change graphs obtained by the LSTM classifier, and generate a list of image paths and corresponding labels (List). Convolutional neural network CNN is built, the input image is 224*224*3, starting from a convolution block with 64 filters, the kernel size is (3*3), the stride is 2, the relu activation layer is used, the convolution operation is performed, and then the filter is changed in the same way. Finally, 4 fully connected layers are added to the network, and the activation function used in the last layer is sigmoid to establish the CNN model. The model uses RMSprop with a learning rate of 0.001 as the optimizer, and optimizes the binary_cross_entropy cross entropy loss function of the binary classification. In this example, the abnormality change graph and its label value are used as input, and the convolution operation is performed to train the CNN model. It is cross-validated and optimized by using the training data, and the optimal parameters of the CNN model are determined on the basis of meeting the optimal evaluation index of the prediction model. On this basis, the optimized model is re-established to obtain the CNN model. The CNN model is used to determine whether the battery is abnormal based on the abnormality change graph. Specifically, the prediction value of the CNN model is used as the basis. If the result is 0, it means that the battery is normal, there is no risk of spontaneous combustion, and no warning is given. If the result is 1, it is determined that the vehicle has the risk of battery spontaneous combustion, and the driver is immediately warned to take preventive measures.

[0136] Step 8. Test the model

[0137] Import the operation data obtained from new energy vehicles with and without battery abnormalities into the database, and after processing in step 1 and step 2, obtain valid feature data that can be used for testing, input it into the determined models for testing, and the operation results are the results of new energy vehicle battery spontaneous combustion detection. Fig. 9 and Fig.10 The abnormality change diagram shown in the figure is Fig. 9 The detected result is that the battery is abnormal. Fig.10 The detected result is that there is no abnormality in the battery. Therefore, in general, the experimental results can prove that the new energy vehicle battery spontaneous combustion early warning method proposed by the present invention is relatively reasonable and effective.

[0138] Reference Fig.11, an embodiment of the present invention provides a new energy vehicle battery spontaneous combustion warning device based on deep learning, comprising:

[0139] The decision tree module 310 establishes and trains a decision tree model based on the historical operation data of the new energy vehicle, and obtains the optimal model parameters through cross-validation; and is used to determine whether the operation data is extreme data;

[0140] LSTM classifier 320, used to establish and train the LSTM classifier according to the historical operation data of the predetermined period, obtain the optimal model parameters through cross-validation; used to calculate the abnormality at a certain moment and construct an abnormality change graph;

[0141] The data acquisition module 330 is used to acquire the operating data of the target new energy vehicle at the current moment and extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0142] The data identification module 340 is used to determine whether the data of the four characteristic values ​​is valid. If it is invalid data, the data is removed and the previous step is re-executed to obtain new data; if it is valid data, the next step is continued;

[0143] A calculation module 350 is used to calculate the change of each eigenvalue relative to the mean of the training data, and use them as new features to obtain eight eigenvalues;

[0144] If the decision tree module determines that the eigenvalue data is extreme data, it means that there is no abnormality, and the extreme data is data with low eigenvalues;

[0145] A continuous segment of feature value data before the current feature value data is input into the LSTM classifier, and the LSTM classifier outputs an abnormality degree. If the abnormal values ​​output by the LSTM classifier appear continuously and the abnormality degree continues to rise to a predetermined amount, the battery is determined to be abnormal.

[0146] Reference Fig.12 Another new energy vehicle battery spontaneous combustion warning device based on deep learning provided by an embodiment of the present invention includes:

[0147] The decision tree module 410 establishes and trains a decision tree model based on the historical operation data of the new energy vehicle, and obtains the optimal model parameters through cross-validation; and is used to determine whether the operation data is extreme data;

[0148] LSTM classifier 420, based on the historical operation data of the predetermined period, establishes and trains the LSTM classifier, obtains the optimal model parameters through cross-validation; is used to calculate the abnormality at a certain moment and construct an abnormality change graph;

[0149] The CNN module 460 establishes and trains a CNN model based on the abnormality change graph obtained by the LSTM classifier, and obtains the optimal model parameters through cross-validation to determine whether the battery is abnormal;

[0150] The data acquisition module 430 acquires the operating data of the target new energy vehicle at the current moment and extracts the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value;

[0151] The data recognition module 440 determines whether the data of the four characteristic values ​​are valid. If the data is invalid, the data is removed and the previous step is re-executed to obtain new data. If the data is valid, the next step is continued.

[0152] A calculation module 450 calculates the change of each eigenvalue relative to the mean of the training data and uses them as new features to obtain eight eigenvalues;

[0153] If the decision tree determines that the eigenvalue data is extreme data, it means that there is no abnormality, and the extreme data is data with low eigenvalues;

[0154] Input a section of continuous data before the current data into the LSTM classifier, calculate the anomaly degree and construct an anomaly degree change graph;

[0155] The CNN module determines whether the battery is abnormal based on the abnormality change diagram. If it is abnormal, it determines that the vehicle is at risk of battery spontaneous combustion at the current moment and sends a warning signal.

[0156] The principle of the new energy vehicle battery spontaneous combustion warning device in the embodiment has been described in detail in the aforementioned warning method, and will not be repeated here.

[0157] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

[0158] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional unit.

[0159] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a function call device, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present application. Although the present invention has been described in detail with general descriptions and specific embodiments above, it is obvious to those skilled in the art that some modifications or improvements can be made to it on the basis of the present invention. Therefore, these modifications or improvements made on the basis of not departing from the spirit of the present invention all belong to the scope of protection claimed in the present invention.

Claims

1. A new energy vehicle battery spontaneous combustion warning method based on deep learning, It is characterized in that The steps include: S10: Based on the historical operation data of new energy vehicles, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation; based on the operation data of the scheduled period, an LSTM classifier is established and trained, and the optimal model parameters are obtained through cross-validation; S11: Obtain the current operating data of the target new energy vehicle and extract the required four characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value; S12: Determine whether the four characteristic value data are valid. If they are invalid data, remove them and re-acquire new data; if they are valid data, proceed to the next step; S13: Calculate the difference between each eigenvalue and the mean of the training data, and use them as new eigenvalues ​​to obtain eight eigenvalues; S14: using the decision tree to determine whether the characteristic value data is extreme data, if it is extreme data, it means that there is no abnormality and no warning is given, and the extreme data is data with low characteristic values; otherwise, the characteristic value data is non-extreme data, and the next step of detection is performed; S15: Read a continuous section of feature value data before the current feature value data, and input it into the LSTM classifier. The LSTM classifier outputs an abnormality degree. If the abnormal values ​​output by the LSTM classifier appear continuously and the abnormality degree continues to increase to a predetermined amount, the battery is determined to be abnormal.

2. A new energy vehicle battery spontaneous combustion warning method based on deep learning, It is characterized in that The steps include: S20a: Obtain historical operating data of the target new energy vehicle and establish and train a decision tree model based on the historical operating data, and obtain optimal model parameters through cross-validation; S20b: According to the operation data of the predetermined period, establish and train the LSTM classifier, and obtain the optimal model parameters through cross-validation; S20c: According to the abnormality change graph obtained by the LSTM classifier, a CNN model is established and trained, and the optimal model parameters are obtained through cross-validation; S21: Obtain the operating data of the target new energy vehicle at the current moment, and extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value; S22: Determine whether the data of the four characteristic values ​​are valid, if they are valid data, proceed to the next step; S23: Calculate the difference between each eigenvalue and the mean of the training data, and use them as new eigenvalues ​​to obtain eight eigenvalues; S24: using a decision tree to determine whether the characteristic value data is extreme data. If the characteristic value data is extreme data, it means that there is no abnormality and no warning is given. The extreme data is data with low characteristic values. Otherwise, the characteristic value data is non-extreme data and the next step of detection is performed. S25: read a section of continuous data before the current data, input it into the LSTM classifier, calculate the abnormality and construct an abnormality change graph; S26: Use CNN to determine whether the battery is abnormal based on the abnormality change graph. If it is abnormal, it is determined that the vehicle is at risk of battery spontaneous combustion at the current moment and a warning signal is issued.

3. The new energy vehicle battery spontaneous combustion warning method as claimed in claim 2, It is characterized in that The establishing and training of the decision tree model specifically includes: Use the extreme gradient boosting algorithm XGBoost to build and train the decision tree model. The model is built by calling the native Python API interface of the XGBoost library. The objective parameter is set to: binary:logistic, which is used for binary classification logistic regression tasks.

4. The new energy vehicle battery spontaneous combustion warning method as claimed in claim 2 or 3, It is characterized in that The establishing and training of the decision tree model specifically includes: Label the acquired data, take each piece of data and its label value as input, and obtain its classification result through the decision tree model. Minimize the gap between the true value and the predicted value as the training objective function, train the decision tree model, and use the training data to perform cross-validation optimization on it; The optimal parameters of the decision tree model are determined on the basis of satisfying the prediction model evaluation index, and the optimized model is re-established on this basis to obtain the decision tree model.

5. The new energy vehicle battery spontaneous combustion warning method as claimed in claim 2, It is characterized in that The establishment and training of the LSTM classifier specifically includes: Call the keras library to build an LSTM neural network, and use the hypersphere or hyperplane feature space to build an LSTM classifier that maps the input data from the original space to the hypersphere or hyperplane feature space.

6. The new energy vehicle battery spontaneous combustion warning method as claimed in claim 5, It is characterized in that The LSTM classifier for mapping input data from the original space to the hypersphere or hyperplane feature space is established, specifically comprising: For the operation data of vehicles with normal battery, N continuous data before the current data are used as input, and the feature points of the current operation data are obtained through the LSTM classifier. The objective function of training is to minimize the hypersphere or hyperplane that can include all the feature points, train the LSTM classifier, and use the training data to perform cross-validation optimization. The optimal parameters of the LSTM classifier are determined on the basis of satisfying the optimal evaluation index of the prediction model, and the optimized model is re-established on this basis to obtain the LSTM classifier, where N is a natural number with a predetermined value.

7. The new energy vehicle battery spontaneous combustion warning method as claimed in claim 2, It is characterized in that The training of the CNN model specifically involves building a convolutional neural network by calling Tensorflow, using the MBGD gradient descent algorithm and the cross entropy loss function as an optimizer to establish the CNN model.

8. A new energy vehicle battery spontaneous combustion warning device based on deep learning, It is characterized in that include: The decision tree module builds and trains a decision tree model based on the historical operation data of new energy vehicles, and obtains the optimal model parameters through cross-validation; Used to determine whether the operating data is extreme data; LSTM classifier, used to establish and train LSTM classifier according to historical operation data of a predetermined period, and obtain the optimal model parameters through cross-validation; used to calculate the abnormality at a certain moment and construct an abnormality change graph; The data acquisition module is used to obtain the operating data of the target new energy vehicle at the current moment and extract the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value; A data identification module is used to determine whether the data of the four characteristic values ​​is valid. If it is invalid data, the data is removed and the previous step is re-executed to obtain new data; if it is valid data, the next step is continued; A calculation module is used to calculate the difference between each eigenvalue and the mean of the training data, and use them as new features to obtain eight eigenvalues; If the decision tree module determines that the eigenvalue data is extreme data, if the eigenvalue data is extreme data, it means that there is no abnormality and no warning is given. The extreme data is data with low eigenvalues; otherwise, the eigenvalue data is non-extreme data and the next step of detection is performed; A continuous segment of feature value data before the current feature value data is input into the LSTM classifier, and the LSTM classifier outputs an abnormality degree. If the abnormal values ​​output by the LSTM classifier appear continuously and the abnormality degree continues to rise to a predetermined amount, the battery is determined to be abnormal.

9. A new energy vehicle battery spontaneous combustion warning device based on deep learning, It is characterized in that include: Decision tree module: Based on the historical operation data of new energy vehicles, a decision tree model is established and trained, and the optimal model parameters are obtained through cross-validation; it is used to determine whether the operation data is extreme data; LSTM classifier, based on the historical operation data of the predetermined period, the LSTM classifier is established and trained, and the optimal model parameters are obtained through cross-validation; Used to calculate the abnormality at a certain moment and construct an abnormality change graph; The CNN module builds and trains the CNN model based on the abnormality change graph obtained by the LSTM classifier, and obtains the optimal model parameters through cross-validation to determine whether the battery is abnormal; The data acquisition module obtains the operating data of the target new energy vehicle at the current moment and extracts the four required characteristic values: total voltage, maximum battery value, maximum temperature value, and minimum temperature value; The data identification module determines whether the data of the four characteristic values ​​is valid. If it is invalid data, the data is removed and the previous step is re-executed to obtain new data; if it is valid data, the next step is continued; The calculation module calculates the difference between each eigenvalue and the mean of the training data, and uses them as new features to obtain eight eigenvalues. If the decision tree determines that the characteristic value data is extreme data, it means that there is no abnormality and no warning is given. The extreme data is data with low characteristic values. Otherwise, the characteristic value data is non-extreme data and the next step of detection is performed. Input a section of continuous data before the current data into the LSTM classifier, calculate the anomaly degree and construct an anomaly degree change graph; The CNN module determines whether the battery is abnormal based on the abnormality change diagram. If it is abnormal, it determines that the vehicle is at risk of battery spontaneous combustion at the current moment and sends a warning signal.

Images