Fault prediction method and device, and training method of fault prediction model

Abstract

The application discloses a fault prediction method and device and a training method of a fault prediction model. The fault prediction model is trained by screening data with high contribution degree to historical fault events of a target vehicle model from a plurality of historical running data of the target vehicle model, and is used for predicting fault grades and probabilities of the vehicle, excluding factors with low correlation degree to faults of the target vehicle model in a vehicle system of the target vehicle model, reducing difficulty in establishing the fault prediction model, and improving accuracy of the fault prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle fault prediction, and in particular to a fault prediction method and device and a training method of a fault prediction model. BACKGROUND

[0002] When a vehicle fails, it needs to be repaired in time, otherwise the failure will escalate, causing poor vehicle safety. Except for a small number of sporadic failures, most vehicle failures have their formation and evolution process, which makes it possible to predict failures over time.

[0003] The prior art predicts vehicle failures by establishing a prediction model using data derived from a vehicle system as features. However, the vehicle system is a complex nonlinear system, and the formation of its failure is influenced by multiple factors. Redundant features that are not closely related to vehicle failures are involved in the training of the prediction model, making it difficult to train the prediction model. These redundant features also interfere with the accuracy of the fault prediction results output by the prediction model. SUMMARY

[0004] The present application provides a fault prediction method, device and training method of a fault prediction model to solve the technical problems of difficult training of existing vehicle fault prediction models and poor accuracy of fault prediction results output.

[0005] In view of the above problems, the present application is proposed to provide a fault prediction method, device and training method of a fault prediction model to overcome the above problems or at least partially solve the above problems.

[0006] In a first aspect, a fault prediction method is provided, comprising:

[0007] Obtaining real-time running data of a target vehicle, the target vehicle being a vehicle of a target vehicle model;

[0008] Determining a level and a probability of failure of the target vehicle within a set prediction time period according to the real-time running data of the target vehicle and a pre-trained fault prediction model, wherein the pre-trained fault prediction model is obtained based on historical running data whose contribution to historical fault events is above a set value.

[0009] Optionally, the pre-trained fault prediction model is constructed by the following steps:

[0010] For each historical running data of the plurality of historical running data, calculating the contribution of the historical running data to the historical fault events;

[0011] The fault prediction model is trained according to historical running data of the target vehicle model whose contribution to historical fault events is above a set value and grades and types of historical fault events of the target vehicle model.

[0012] Optionally, for each historical running data of the plurality of historical running data, the contribution of the historical running data to historical fault events is calculated.

[0013] For each historical running data of the plurality of historical running data, the contribution of the historical running data to historical fault events of the target vehicle model is calculated by using a Shapley value method.

[0014] Optionally, the fault prediction model is trained according to historical running data of the target vehicle model whose contribution to historical fault events is above a set value and grades and types of historical fault events of the target vehicle model.

[0015] The historical running data of the target vehicle model whose contribution to historical fault events is above a set value is feature extracted to obtain training samples.

[0016] Based on a plurality of preset sliding steps, the training samples are respectively input into corresponding initial neural network models for training to obtain a plurality of fault prediction models, and the plurality of fault prediction models have different prediction time lengths.

[0017] Optionally, the grades and probabilities of the target vehicle to have faults in a set prediction time length are determined according to real-time running data of the target vehicle and a pre-trained fault prediction model.

[0018] According to a set prediction time length, a combination of a minimum number of called fault prediction models from the plurality of fault prediction models is determined.

[0019] The real-time running data of the target vehicle is input into the combination of the minimum number of called fault prediction models to obtain the grades and probabilities of the target vehicle to have faults in the set prediction time length.

[0020] Optionally, the combination of the minimum number of called fault prediction models from the plurality of fault prediction models is determined according to a set prediction time length.

[0021] According to a set prediction time length, the combination of the minimum number of called fault prediction models from the plurality of fault prediction models includes N fault prediction models, the N fault prediction models have a set iteration order, the sum of prediction time lengths of the N fault prediction models is a set prediction time length, and N is a positive integer greater than 1.

[0022] The real-time running data of the target vehicle is input into the combination of the fewest number of the called fault prediction models, to obtain the level and probability of the target vehicle failing in the set prediction time length.

[0023] The first fault prediction model in the N fault prediction models receives the real-time running data of the target vehicle at a first time, sequentially performs a prediction operation according to the set iteration order, and outputs the level and probability of the target vehicle failing at a second time, the second time being a time at which the first time elapses by a set prediction time length.

[0024] Optionally, the plurality of fault prediction models comprises:

[0025] The prediction time lengths of the fault prediction models are 5 minutes, 10 minutes, 30 minutes, 60 minutes, and 120 minutes, respectively.

[0026] Optionally, the historical running data of the target vehicle model comprises historical vehicle speed information, historical mileage information, historical battery information, and historical temperature information of the target vehicle model.

[0027] In a second aspect, a fault prediction device is provided, comprising:

[0028] A contribution degree determination unit is configured to obtain real-time running data of a target vehicle, the target vehicle being a vehicle of a target vehicle model.

[0029] A fault prediction unit is configured to determine, according to the real-time running data of the target vehicle and a pre-trained fault prediction model, a level and probability of the target vehicle failing in a set prediction time length, wherein the pre-trained fault prediction model is obtained based on historical running data having a contribution degree to a historical fault event of the target vehicle model being above a set value.

[0030] In a third aspect, a training method of a fault prediction model is provided, comprising: training a fault prediction model based on historical running data having a contribution degree to a historical fault event of a target vehicle model being above a set value, the fault prediction model being used to determine, according to real-time running data of a target vehicle, a level and probability of the target vehicle failing in a set prediction time length, wherein the target vehicle is a vehicle of the target vehicle model.

[0031] In a fourth aspect, the application further provides a controller comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the server performs the method provided in the first aspect.

[0032] In a fifth aspect, the present application also provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor, so that the computer executes the method provided in the first aspect.

[0033] In a sixth aspect, the present application also provides a computer program product, which comprises a computer program. When the computer program is executed, the computer program causes a computer to execute the method provided in the first aspect.

[0034] The technical scheme provided by the present application has at least the following technical effects or advantages:

[0035] The fault prediction method, device and training method of the fault prediction model provided by the present application adopt the fault prediction model trained from the data with high contribution degree to the historical fault events of the target vehicle model, to predict the fault level and probability of the vehicle, exclude the factors with low correlation degree to the fault of the target vehicle model in the vehicle system of the target vehicle model, reduce the difficulty of establishing the fault prediction model, and improve the accuracy of the fault prediction model.

[0036] The above description is only a summary of the technical scheme of the present application. In order to enable the technical means of the present application to be more clearly understood, and to be implemented according to the content of the description, and in order to enable the above and other purposes, characteristics and advantages of the present application to be more apparent and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0037] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments and are not meant to limit the present application. Furthermore, the same reference numerals are used throughout the several drawings to designate the same or similar parts. In the drawings:

[0038] Figure 1 Fault prediction method flow for the embodiments of the present application Figure 1 ;

[0039] Figure 2 Fault prediction method flow for the embodiments of the present application Figure 2 ;

[0040] Figure 3 Fault prediction model establishment and application flow for the embodiments of the present application

[0041] Figure 4 System data flow process diagram for the embodiments of the present application

[0042] Figure 5 Historical running signal change example for the embodiments of the present application, which is backtracked when a fault occurs for a target vehicle model

[0043] Figure 6 Network structure diagram of GRU in the embodiment of the application;

[0044] Figure 7 Structure diagram of the fault prediction device in the embodiment of the application;

[0045] Figure 8 Schematic diagram of the controller in the embodiment of the application. DETAILED DESCRIPTION

[0046] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings.

[0047] Various structural schematic diagrams according to the embodiments of the application are shown in the accompanying drawings. These diagrams are not drawn to scale, in which certain details are exaggerated for the purpose of clarity and certain details may be omitted. The shapes of various regions, layers shown in the diagrams and their relative sizes and positional relationships may deviate in practice due to manufacturing tolerances or technical limitations, and regions / layers with different shapes, sizes and relative positions may be additionally designed by those skilled in the art according to actual needs.

[0048] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with specific embodiments, and it should be understood that the embodiments of the present disclosure and the specific features in the embodiments are detailed descriptions of the technical solutions of the present application, and not limitations of the technical solutions of the present application. In the case of no conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0049] When a vehicle fails, it needs to be repaired in time, otherwise the failure will escalate, causing poor vehicle safety. Except for a small number of occasional failures, most vehicle failures have their formation and evolution process, which makes it possible to predict failures over time.

[0050] Taking new energy vehicles as an example, the power batteries of current new energy vehicles face two major pain points of range anxiety and safety hazards. A high-safety, high-energy-density power battery system is an important technical approach to solving the pain points of power batteries. Solid-state batteries are a kind of batteries using solid electrodes and solid electrolytes, which can be matched with high specific energy materials, greatly reducing weight and improving energy density to 500 Wh / kg or even higher. In terms of safety, solid-state batteries have the advantages of high strength, high electrochemical stability and high ignition point. Therefore, as a new battery power system, studying the failure occurrence and evolution rules has a very important role in vehicle safety protection, prolonging battery life and promoting industrial development.

[0051] However, as the mainstream method for predicting classification problems, LSTM has the problem of gradient vanishing and gradient explosion caused by long-term dependence, but LSTM has three different gates and more parameters, which makes it more difficult to train.

[0052] New energy vehicles are the main development trend in the future, and their vehicle usage data are crucial for analyzing user usage habits, troubleshooting, and improving the next generation of products. Major companies have built corresponding platforms to collect relevant data on new energy vehicle travel and analyze and monitor the vehicle usage process. In actual practice, in addition to a small number of sporadic faults, most vehicle faults have their formation and evolution process, which makes it possible to predict faults over a certain period of time.

[0053] However, due to the complexity of the vehicle as a nonlinear system, its faults are affected by many factors. Currently, the mainstream research method is to select a small number of parameters based on a specific fault, which is different from the actual situation of vehicle operation. In addition, in the fault prediction process, the accuracy of short-term prediction is still controllable, but in actual application, a considerable amount of time is often needed to make the fault prediction have application value. The current mainstream approach is to substitute the predicted data into the model for multiple extrapolation calculations. The longer the prediction time, the more iterations, and the more difficult it is to ensure accuracy.

[0054] Therefore, the present application provides a fault prediction method, please refer to Figure 1 , Figure 1 The fault prediction method flowchart in the embodiments of the present application comprises:

[0055] S101, obtaining real-time running data of a target vehicle, the target vehicle being a vehicle of a target vehicle model.

[0056] S102, determining the grade and probability of the target vehicle occurring a fault within a set prediction time according to the real-time running data of the target vehicle and a pre-trained fault prediction model. The pre-trained fault prediction model is obtained based on historical running data whose contribution to historical fault events of the target vehicle model is above a set value.

[0057] The fault prediction method provided in the embodiments of the present application selects data with high contribution to historical fault events of the target vehicle model from a plurality of historical running data of the target vehicle model to participate in the training of the fault prediction model, excludes factors with low correlation to faults of the target vehicle model in the vehicle system of the target vehicle model, reduces the difficulty of establishing the fault prediction model, and improves the accuracy of the fault prediction model.

[0058] At present, all large enterprises have built corresponding platforms to collect relevant data of new energy vehicles running, therefore, the historical running data of the target vehicle model can be collected and sent by the vehicle-mounted data terminal of the target vehicle model in real time, the platform receives the data and performs preliminary analysis, cleaning and classification, and further pushes the data to the flink streaming processing program for full backup of the running data, so as to filter and analyze the modeling trend in the later stage.

[0059] It can be understood that the historical running data of the target vehicle model is remote transmission data, due to the uncertainty of equipment and network, missing values and abnormal values often occur, therefore, unified processing is needed, for missing values, mean filling processing can be used, and for abnormal values, IQR method (Interquartile Range, a method for measuring variability) can be used for processing.

[0060] The following will be introduced Figure 2 How to obtain the fault prediction model based on the historical running data with the contribution degree to the historical fault event of the target vehicle model above the set value in operation S102.

[0061] S201, for each historical running data of the plurality of historical running data, calculate the contribution degree of the historical running data to the historical fault event.

[0062] The vehicle system is a complex nonlinear system, and its fault formation is affected by many factors. Therefore, as many data fields as possible should be selected to participate in the pre-training of the fault prediction model, the contribution degree of the features is obtained through the model interpretation method, and the related redundant features are excluded, which is the process of feature construction.

[0063] In some optional embodiments, for each historical running data of the plurality of historical running data, the contribution degree of the historical running data to the historical fault event of the target vehicle model is calculated by using the Shapley value method.

[0064] Exemplarily, the shapley value is used to explain the contribution degree of each field, and the calculation formula is:

[0065] ;

[0066] In the formula, is the contribution degree of a certain feature to the result, i.e. the contribution degree of a certain historical running data to the historical fault event of the target vehicle model, is the set of all features, each feature represents a historical running data, is a subset of the set which does not contain the feature , and is the total number of input features. The model prediction value of the features in the set The model prediction value of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set The greater the value, the greater the impact of the features in the set

[0067] S202, according to the historical running data of the target vehicle model whose contribution to the historical fault event is above a set value, and the level and type of the historical fault event of the target vehicle model, a fault prediction model is trained.

[0068] According to historical experience, the fault types of the target vehicle model can be divided into: temperature fault, voltage fault, soc fault, single cell serious inconsistency fault, battery management system fault, other electrical fault, and current fault. Among them, SOC refers to State of Charge, which is used to reflect the remaining capacity of the battery, indicating the ability of the battery to continue working. In the case of complete discharge, the SOC is 0, and in the case of complete charging, the SOC is 1, which is generally represented by 0-100%.

[0069] Taking the current fault as an example, for each historical running data in the plurality of historical running data, the contribution of the historical running data to the historical current fault event of the target vehicle model is calculated by using the Shapley value method, and if the contribution is above a set value, the historical running data is used for training the fault prediction model.

[0070] It can be understood that the above division of the fault types of the vehicle is for the fault prediction model to output the level and probability of a specific type of fault. The fault types of the vehicle are obtained according to the real vehicle fault of the target vehicle model, that is, the fault that can be measured according to the running data of the target vehicle model.

[0071] In some optional embodiments, the historical running data of the target vehicle model includes: historical speed information, historical mileage information, historical battery information, and historical temperature information of the target vehicle model.

[0072] Exemplarily, for the historical running data of a certain target vehicle model, the contribution degree of each historical running data obtained above to the historical failure event of the target vehicle model is adopted, and then the historical running data with the contribution degree above a set value is obtained as: vehicle speed, cumulative mileage, soc, total voltage, total current, highest voltage of battery monomer, lowest voltage of battery monomer, highest temperature, and lowest temperature. The historical running data with the contribution degree above the set value has a corresponding vehicle frame number and a corresponding data receiving time, that is, one vehicle frame number has full field data of vehicle speed, cumulative mileage, soc, total voltage, total current, highest voltage of battery monomer, lowest voltage of battery monomer, highest temperature, and lowest temperature.

[0073] For a vehicle, the practicability of failure prediction is closely related to the prediction time. Based on a preset sliding step, a failure prediction model has a fixed prediction time length. For example, the time interval for collecting historical running data of a target vehicle model is 10s once, and if the sliding step k=30, the failure prediction model trained based on the historical running data of the target vehicle model can only provide 5min of early warning.

[0074] In actual operation, in order to make a medium and long term prediction of failure, the prediction result is usually put into the failure prediction model multiple times to calculate the prediction conclusion of N times of the sliding step (iteration number N*sliding step 5min). When the iteration number is small, the accuracy can be ensured. When the required prediction time is long, the model will be distorted.

[0075] Therefore, in some optional embodiments, the operation of S202 above specifically includes:

[0076] Firstly, the historical running data with a contribution degree to the historical failure event of the target vehicle model above a set value is subjected to feature extraction to obtain training samples.

[0077] Secondly, based on multiple preset sliding steps, the training samples are respectively input into corresponding initial neural network models for training to obtain multiple failure prediction models. The multiple failure prediction models have different prediction time lengths. The initial neural network model can adopt a GRU (Gate Recurrent Unit, recurrent neural network) basic model.

[0078] Based on this, the operation S102 above specifically includes:

[0079] Firstly, according to a set prediction length, a combination with the least number of called failure prediction models from the multiple failure prediction models is determined.

[0080] The operation of the first step of the operation S102 specifically comprises: determining, according to the set prediction time length, a combination of the least number of fault prediction models to be called from the plurality of fault prediction models, the combination comprising N fault prediction models, the N fault prediction models having a set iteration order, the sum of the prediction time lengths of the N fault prediction models being the set prediction time length, N being a positive integer greater than 1.

[0081] The second step is to input the real-time running data of the target vehicle into the combination of the least number of fault prediction models to be called, to obtain the level and probability of the target vehicle failing in the set prediction time length.

[0082] The operation of the second step of the operation S102 specifically comprises: the first fault prediction model in the N fault prediction models receiving the real-time running data of the target vehicle at the first time, sequentially performing the prediction operation according to the set iteration order, and outputting the level and probability of the target vehicle failing at the second time, the second time being the time when the first time elapses by the set prediction time length.

[0083] Specifically, when the N fault prediction models sequentially perform the prediction operation according to the set iteration order, the output of the previous fault prediction model is taken as the input of the next fault prediction model, and each fault prediction model outputs mainly the level and probability of predicting the failure. For the case of multiple iterations, each fault prediction model also outputs the predicted features, including: vehicle speed, cumulative mileage, soc, total voltage, total current, highest voltage of battery monomer, lowest voltage of battery monomer, highest temperature, lowest temperature, etc.

[0084] Illustratively, after determining the combination of the least number of N fault prediction models to be called, the real-time running data of the target vehicle at the first time is input into the first fault prediction model in the combination, and the first prediction running data corresponding to the first fault prediction model for the first prediction time length and the level and probability of the vehicle failing in the first prediction time length are output, the first prediction running data being the predicted value of the real-time running data at the first time for the first prediction time length. Then, the first prediction running data is input into the second fault prediction model in the combination, and the second prediction running data for the time length of the sum of the first prediction time length and the second prediction time length and the level and probability of the vehicle failing in the time length of the sum of the first prediction time length and the second prediction time length are output, and so on, until the N-1 prediction running data output by the N-1th fault prediction model is input into the Nth fault prediction model, and the level and probability of the target vehicle failing at the second time can be output.

[0085] Thus, by determining the combination of the fewest number of fault prediction models to be called from the plurality of fault prediction models according to the set prediction length, the iteration number is suppressed, and the iteration number and the time dimension of the prediction of the model are adjusted through hierarchical time aggregation, so as to achieve the effect of fewer iteration numbers and improved accuracy, that is, the time is divided and aggregated according to different levels, and the change trend of the vehicle fault in different time dimensions is analyzed.

[0086] Exemplarily, taking the time interval of 10s of the historical running data of the target vehicle model as the premise, based on a plurality of preset sliding steps, five time length fault prediction models are trained, including:

[0087] Model one: based on the preset sliding step k of 30s, the fault prediction model with the prediction time length of 5min is trained.

[0088] Model two: based on the preset sliding step k of 60s, the fault prediction model with the prediction time length of 10min is trained.

[0089] Model three: based on the preset sliding step k of 180s, the fault prediction model with the prediction time length of 30min is trained.

[0090] Model four: based on the preset sliding step k of 360s, the fault prediction model with the prediction time length of 1h is trained.

[0091] Model five: based on the preset sliding step k of 720s, the fault prediction model with the prediction time length of 2h is trained.

[0092] Thus, according to the actual requirements of the target vehicle, short-term, medium-term and long-term prediction conclusions are provided, for example, 5min fault prediction results are required, that is, the set prediction length is 5min, the combination of the fewest number of fault prediction models is called from the five fault prediction models, that is, model one is directly called, and the real-time running data of the target vehicle is input into the called model one, so as to directly output the grade and probability of the fault of the target vehicle within 5min.

[0093] For example, when 1.5h of prediction results are required, there are a plurality of combination schemes for calling the model to iterate, including: calling model one and iterating 18 times, or calling model two and iterating 9 times, or calling model three and iterating 3 times, and so on, thus, according to the principle of suppressing the iteration number, model three is selected to iterate 3 times.

[0094] For example, when a 6h prediction result is required, there are multiple combination schemes for calling the models for iteration, including: calling model one for 70 iterations, calling model two for 35 iterations, calling model three for 12 iterations, calling model four for 6 iterations, and calling model five for 3 iterations. Selecting the calling model five for 3 iterations can meet the requirements.

[0095] Thus, by using the method provided in the embodiments of the present application, according to the set prediction time length, based on the principle of suppressing the number of called fault prediction models from the multiple fault prediction models, the combination of the least number of called fault prediction models from the multiple fault prediction models is selected, which can greatly reduce the iteration number, and the accuracy of fault prediction is not affected to a certain extent, and the fewer iteration numbers also make the fault prediction process more rapid and efficient, saving system resources.

[0096] It can be understood that the above-mentioned fault prediction method can be applied to a terminal device and / or a server. The terminal device and / or the server can be understood as a test device. The server can be, but is not limited to, a network server, a database server, a cloud server, etc. The terminal device can be, but is not limited to, an office computer.

[0097] Based on the same inventive concept, the embodiments of the present application also provide a fault prediction model training method, including: training a fault prediction model based on historical running data with a contribution degree to a historical fault event of a target vehicle model above a set value; and the fault prediction model is used to determine a fault level and a probability of the target vehicle occurring within a set prediction time length, according to real-time running data of the target vehicle, wherein the target vehicle is a vehicle of the target vehicle model.

[0098] For example, the fault prediction model training method in the embodiments of the present application specifically includes: Figure 2

[0099] S201, for each historical running data of the multiple historical running data, calculating a contribution degree of the historical running data to a historical fault event.

[0100] S202, training a fault prediction model according to the historical running data with a contribution degree to a historical fault event of a target vehicle model above a set value and a fault level and a type of the historical fault event of the target vehicle model.

[0101] The following takes a solid-state battery passenger vehicle as an example, and introduces how to train a fault prediction model and use it for fault prediction of a target vehicle, in combination with examples and Figures 3 to 6

[0102] 01 Data acquisition: for example, the historical running data of the target vehicle model is obtained by collecting the historical running data of the target vehicle model from the vehicle data platform. Figure 3 ​​As shown, the receiving platform acquires the data sent by the vehicle data terminal, which includes historical fault data of the target vehicle model and real-time operation data of the target vehicle, wherein the historical fault data of the target vehicle model is derived from the historical actual operation data of the target vehicle model.

[0103] In the historical operation event of the target vehicle model, the relevant data signals of the target vehicle model are collected and sent in real time by the vehicle data terminal, the receiving platform performs preliminary analysis, cleaning and classification on the data, and further pushes the data to the flink stream processing program, performs full backup on the operation data, so as to perform filtering and related modeling trend analysis in the later period.

[0104] The specific system data flow is shown in Figure 4 The data platform receives the operation data of the solid-state battery passenger vehicle, backs up the original data to the hive database, the big data cluster acquires the data through the kafka consumption provided by the data platform, and finally stores the results in the MySQL cluster through flink data processing and spark offline analysis calculation, and synchronizes the dynamic data (left) and real-time data (right) between the big data cluster and the background server through the Redis cluster.

[0105] After the background server queries the result data and completes the related statistical aggregation calculation, it provides the web interface for display through the data API interface, or other business platforms call, the platform supports the access of vehicle dynamic addition and storage, and dynamic expansion of computing resources. The fault analysis and prediction model is integrated in the fault diagnosis and early warning module, and the prediction result is input by calling the real-time dynamic data of the vehicle.

[0106] It can be understood that the data platform refers to the receiving platform of data, the hive is a data warehouse tool that can store, load and analyze data, kafka is a high-throughput data transmission method, flink is a high-efficiency data processing framework, and Redis is a high-performance database supporting master-slave synchronization. The integration of flink and Redis can greatly improve the application efficiency of real-time data analysis. In the integration, flink is responsible for real-time data processing and calculation, and Redis is used as the storage of result data. After completing data processing, flink can efficiently write results into Redis. Due to the high-speed read-write capability of Redis, these result data can be quickly acquired by the backend service or application program, realizing fast real-time data service. Dynamic data (left) and real-time data (right) are both data calculation results.

[0107] The above operation aims to collect and store the historical operation data of the target vehicle model.

[0108] 02 Data processing: Due to the uncertainty of equipment and network, missing values and outliers often occur, which need to be uniformly processed before inputting the model. In the embodiment of the application, the mean filling method is used for missing values, and the IQR method is used for outliers.

[0109] 03 Feature construction: The solid-state battery vehicle system is a complex nonlinear system, and its fault formation is affected by many factors. Therefore, as many data fields as possible should be selected to participate in the pre-training model during feature construction, and the contribution of the features is obtained through model interpretation methods to exclude relevant redundant features.

[0110] The shapley value is used to explain the contribution of each field in the embodiment of the application, and the calculation formula is:

[0111]

[0112] In the formula, is the contribution of a certain feature to the result, that is, the contribution of a certain historical running data to the historical fault event of the target vehicle type, is the set of all features, and each feature represents a historical running data, is a subset of the set that does not contain the feature , is the total number of input features, represents the model prediction value of the combination of the features in the set and the feature , and the greater the value, the greater the influence of the combination of the features in the set and the feature on the model prediction, that is, the more important the combination of the features in the set and the feature , represents the model prediction value of the combination of the features in the set , and the greater the value, the greater the influence of the combination of the features in the set on the model prediction, that is, the more important the combination of the features in the set .

[0113] Exemplarily, in the embodiment of the application, the solid-state battery passenger vehicle fault types are divided, specifically: temperature fault, voltage fault, soc fault, single cell serious inconsistency fault, battery management system fault, other electrical fault, and current fault.

[0114] Through calculation, the fields of the historical operation data of which the contribution degree is above the set value to form the main features as the input of the failure prediction model are: vehicle speed, cumulative mileage, soc, total voltage, total current, battery cell maximum voltage, battery cell minimum voltage, maximum temperature, and minimum temperature. These features have corresponding vehicle frame numbers and corresponding data receiving times.

[0115] The fields of the data as the output of the failure prediction model are: failure type, failure level, and probability.

[0116] Exemplarily, the specific examples of the fields of the input and output features are shown in the following table:

[0117]

[0118] 04 Build training set and test set:

[0119] Data standardization: in order to remove the influence of the magnitude and unit of the historical operation data of the target vehicle model on the result, the data needs to be scaled in proportion and converted into dimensionless pure numerical values, so that different indicators can be compared and weighted.

[0120] The embodiment of the present application adopts Normalization operation, which uniformly maps the data to the interval [0, 1], the conversion formula is as follows:

[0121]

[0122] In the formula, is the maximum value of the sample data, is the minimum value of the sample data.

[0123] One-Hot encoding: for the classification variable of the failure type, it needs to be converted into a numerical form that the model can understand. Mainly, N-bit status registers are used to encode N states, each state has its own independent register bit, and only one bit is valid at any time. One-Hot encoding is a representation of a categorical variable as a binary vector. This first requires mapping the categorical value to an integer value. Then, each integer value is represented as a binary vector, except for the index of the integer, which is a zero value, and it is marked as 1. For example, the feature failure level to be processed has four classifications, which will correspond to one-hot encoding respectively: 0001, 0010, 0100, 1000.

[0124] Sliding window segmentation: The predicted label is the fault level at the time of fault occurrence, a total of 4 labels: [level 1, level 2, level 3, none], denoted as y. The vehicle signal data includes the following 9 feature quantities: [speed, total mileage, SOC, total voltage, total current, maximum battery voltage, minimum battery current, maximum battery temperature, minimum battery temperature]. First, locate the fault point. According to the fault level signal, each vehicle is segmented, and each segment belongs to one of the above fault levels: level 1, level 2, and level 3. Each segment can see the time series changes of its related 9 types of signal data, such as shown in FIG. 3, which represents the time series changes of the total voltage, SOC, and minimum temperature of a certain fault point corresponding segment. Figure 5

[0125] Fault sample data processing: The embodiment of the application selects the signals 200 time steps (timestep=200) (this is called window size s) before the fault occurs as x, and obtains 9 signals at each time point as a feature vector. Then for a certain fault level, x is a two-dimensional vector with a size of [200, 9].

[0126] Non-fault sample data processing: According to the fault time t, a window sliding step k is designed, which needs to ensure that k<=s, then the t+k time is selected as a non-fault sample. Similarly, the length of s=200 time steps before the x is selected as the non-fault sample, and 9 signal feature data at each time point is obtained to generate a two-dimensional vector with a dimension of [200, 9] as the sample x. According to the above "sliding window" logic, the positive samples (i.e. non-fault samples) are selected, and all fault labels and their corresponding 200 time steps of signals constitute the feature set X.

[0127] Dataset division: Here, a certain period of solid-state battery passenger vehicle operation data is selected as the feature sample for model input, each sample x is a 200*9 matrix, and an 8:2 ratio is used to divide the test set and the validation set.

[0128] 05 Model training: Fault prediction is actually a multi-classification problem, that is, given an input x, the model outputs a vector y. Each element of the vector y corresponds to the probability of the corresponding class, and we take the class corresponding to the element with the maximum probability as the prediction result at time t. In the embodiment of the application, the latest 200 pieces of 9 signal sequences of each vehicle are obtained and saved in real time, and then input to the trained model, and the output of the model is the fault level label that the vehicle may occur at the next time.

[0129] Vehicle fault has the characteristics of time series data, and the variant of recurrent neural network (RNN), gated recurrent unit (GRU), is the mainstream scheme in the industry, which can capture the time dependence in time series data. The embodiment of the application uses GRU to construct the network structure, as shown in FIG. 4. Figure 6 ​As shown, the name of the model input layer inputlayer is gru_input, and the shape of the input data is annotated in brackets behind it [(None, 100, 9)], which means that the size of the input layer is arbitrary, each sample has 100 time steps, and each time step contains 9 features. The gru and gru_1 layers are the recurrent layers of the model, the dropout and dropout_1 layers are used to prevent overfitting of the neural network, reduce the dependence between neurons, and enhance the generalization ability of the model, and the dense layer represents the fully connected layer of the model, which is used for feature space dimension conversion and output of the model.

[0130] The activation function of the GRU layer is tanh, and the activation function of the fully connected layer is softmax. The optimizer uses the Adam optimizer. The softmax is a normalized exponential function, and the sum of the final output of 4 prediction values is equal to 1. Specifically, the probability that the sample vector x belongs to the jth category is:

[0131] ;

[0132] In the formula, represents the input vector belongs to the category , is the jth element of , is the total number of categories, is a variable, and the variable traverses from 1 to . .

[0133] Finally, the model is trained using (X, y), the training loss is cross-entropy, and the accuracy is the evaluation standard. After a certain number of iterations, the final model is obtained. The formula of the loss function is as follows:

[0134] ;

[0135] In the formula, represents the true value, represents the predicted value.

[0136] The embodiments of the present application use the keras tool to realize the network structure, and the hyperparameters of the training process include:

[0137] 1) batch_size: the number of samples for each gradient update.

[0138] 2) epochs: the number of iterations of the training model.

[0139] 3) validation_split: the proportion of test data.

[0140] 4) hidden_size is the number of hidden layer neurons.

[0141] 5) n_features is the number of input features of the model.

[0142] The parameter settings are: n_features = 9, hidden_size = 80, validation_split = 0.2, epochs = 500, batch_size = 256. After multiple iterations and adjustments, a good prediction model is obtained.

[0143] 06 Model verification: the model effect is evaluated by the test set, and the accuracy (Accuracy) is used for result verification in the embodiments of the application, and the LSTM model is compared. The accuracy is defined as the ratio of the number of correctly classified samples to the total number of samples, and the higher the accuracy, the better the effect. In addition, the "model loss" index value is used for analysis as an important reference for model verification, and the lower the loss, the better the effect. Through 3-fold cross-validation, the test results shown in the following table are obtained:

[0144]

[0145] In summary, according to the historical running data whose contribution degree to the target vehicle model is above the set value and the level and type of the historical fault events of the target vehicle model, a fault prediction model is trained.

[0146] As shown in Figure 3 , the real-time running data of the target vehicle is input into the fault prediction model trained in the above embodiments, and the level and probability of the target vehicle fault occurring in the target prediction duration can be obtained.

[0147] The training method of the fault prediction model provided in the embodiments of the application realizes the prediction of future vehicle fault changes by monitoring and identifying the regularity of the historical signal feature changes of the target vehicle model, locates the vehicle fault in advance, performs targeted inspection and improvement in advance, avoids the escalation of the fault, changes the maintenance method of the maintenance personnel, and simplifies the cumbersome procedures of maintenance. The shapley value model explanation method is used to measure the contribution degree of each feature, which accurately constructs the features and increases the interpretability of the fault prediction model.

[0148] The fault prediction model trained by the method provided in the above embodiments can be used for the fault prediction method provided in the above embodiments, and through hierarchical time series aggregation, the change trend analysis of the fault in different time dimensions is realized to cope with different business demand scenarios, which can greatly reduce the number of algorithm iterations and reduce the error caused by multiple extrapolations of the algorithm.

[0149] As shown in Figure 3As shown, after the algorithm is deployed on the platform, the data collected by the real vehicle can be input into the model to perform short-term, medium-term and long-term fault level prediction. 0 represents a normal vehicle state, 1 represents a first-level fault, which does not affect the normal driving of the vehicle, 2 represents a second-level fault, which affects the performance of the vehicle and needs to remind the driver, and 3 represents a third-level fault, which affects the normal driving of the vehicle and needs to be checked. For medium and long term prediction, the system provides the fault level and its occurrence probability, reminding the relevant personnel to take targeted inspection and improvement in advance to avoid the escalation of the fault and improve the maintenance method.

[0150] Based on the same inventive concept, the embodiment of the present application also provides a fault prediction device, please refer to Figure 7 , Figure 7 The fault prediction device 700 in the embodiment of the present application is a schematic diagram of the fault prediction device, which comprises:

[0151] The contribution degree determination unit 701 is configured to obtain real-time running data of a target vehicle, wherein the target vehicle is a vehicle of a target vehicle model.

[0152] The fault prediction unit 702 is configured to determine the level and probability of a fault of the target vehicle occurring within a set prediction time according to the real-time running data of the target vehicle and a pre-trained fault prediction model, wherein the pre-trained fault prediction model is obtained based on historical running data whose contribution degree to historical fault events of the target vehicle is above a set value.

[0153] The fault prediction device provided by the embodiment of the present application can be used to implement the fault prediction method provided by the above-mentioned embodiment. The fault prediction model trained from the historical running data of the target vehicle model with a high contribution degree to the historical fault events of the target vehicle model is used to predict the fault level and probability of the vehicle, which excludes the factors in the vehicle system of the target vehicle model with a low correlation degree to the fault of the target vehicle model, reduces the difficulty of establishing the fault prediction model, and improves the accuracy of the fault prediction model.

[0154] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the computer execute the method provided by the above-mentioned embodiment.

[0155] The embodiment of the present application also provides a controller, as Figure 8 shown, the controller 800 comprises a memory 801, a processor 802, and a computer program 803 stored in the memory 801 and executable on the processor 802, and the processor 802 executes the computer program 803 to make the controller 800 execute the method provided by the above-mentioned embodiment.

[0156] The embodiment of the present application further provides a computer program product, comprising a computer program, which, when executed by a computer, causes the computer to perform the method provided by the above embodiment.

[0157] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by relevant hardware instructed by programs. The foregoing programs can be stored in a computer readable storage medium. When the programs are executed, the steps of the above-mentioned method embodiments are executed; and the foregoing storage medium includes various media capable of storing program codes, such as ROM, RAM, magnetic disks, or optical disks.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can still be modified, or some or all of the technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

[0159] In the specification provided herein, a large number of specific details are illustrated. However, it can be understood that the embodiments of the present application can be practiced without these specific details. In some examples, well-known methods, structures and techniques are not illustrated in detail in order not to obscure the understanding of the present specification.

Claims

1. A failure prediction method characterized by, The method comprises the following steps: obtaining real-time running data of a target vehicle, the target vehicle being a vehicle of a target vehicle model; determining, according to the real-time running data of the target vehicle and a pre-trained fault prediction model, a level and a probability of the target vehicle having a fault within a set prediction time, comprising: determining, according to the set prediction time, a combination of a minimum number of fault prediction models to be called from a plurality of fault prediction models, comprising: determining, according to the set prediction time, that the combination of the minimum number of fault prediction models to be called from the plurality of fault prediction models comprises N fault prediction models, the N fault prediction models having a set iteration order, the sum of the prediction time lengths of the N fault prediction models being the set prediction time, N being a positive integer greater than 1; inputting the real-time running data of the target vehicle into the combination of the minimum number of fault prediction models to be called to obtain the level and the probability of the target vehicle having a fault within the set prediction time, comprising: a first fault prediction model in the N fault prediction models receives real-time running data of the target vehicle at a first time, sequentially performs a prediction operation according to the set iteration order, and outputs a level and a probability of the target vehicle having a fault at a second time, the second time being a time at which the first time has elapsed for the set prediction time, the pre-trained fault prediction model being obtained based on historical running data having a contribution to historical fault events of the target vehicle model above a set value.

2. The failure prediction method according to claim 1, wherein The pre-trained fault prediction model is constructed by the following steps: for each historical running data of a plurality of historical running data of the target vehicle model, calculating a contribution of the historical running data to historical fault events of the target vehicle model; training the fault prediction model according to historical running data having a contribution to historical fault events of the target vehicle model above a set value and levels and types of historical fault events of the target vehicle model.

3. The failure prediction method according to Claim 2, wherein The method comprises the following steps: for each historical running data of the plurality of historical running data, calculating a contribution of the historical running data to historical fault events of the target vehicle model using a Shapley value method.

4. The failure prediction method of claim 2, wherein, The method comprises the following steps: extracting features from historical running data having a contribution to historical fault events of the target vehicle model above a set value to obtain training samples; based on a plurality of preset sliding steps, inputting the training samples into corresponding initial neural network models respectively for training to obtain a plurality of fault prediction models, the plurality of fault prediction models having different prediction time lengths.

5. The failure prediction method of claim 1, wherein, The plurality of fault prediction models comprise: fault prediction models having prediction time lengths of 5 minutes, 10 minutes, 30 minutes, 60 minutes and 120 minutes respectively.

6. The failure prediction method of claim 1, wherein, The historical running data of the target vehicle model includes historical vehicle speed information, historical mileage information, historical battery information, and historical temperature information of the target vehicle model.

7. A failure prediction device characterized by comprising: Comprise: The contribution degree determination unit is configured to obtain real-time running data of a target vehicle, wherein the target vehicle is a vehicle of a target vehicle model. The fault prediction unit is configured to determine a level and a probability of the target vehicle having a fault within a set prediction time period according to the real-time running data of the target vehicle and a pre-trained fault prediction model, and the fault prediction unit comprises: According to the set prediction time period, determine the combination of the least number of fault prediction models called from the plurality of fault prediction models, comprising: According to the set prediction time period, determining the combination of the least number of fault prediction models called from the plurality of fault prediction models includes N fault prediction models, the N fault prediction models have a set iteration order, the sum of the prediction time lengths of the N fault prediction models is the set prediction time period, and N is a positive integer greater than 1. Input the real-time running data of the target vehicle into the combination of the least number of fault prediction models called, to obtain the level and the probability of the target vehicle having a fault within the set prediction time period, comprising: The first fault prediction model in the N fault prediction models receives the real-time running data of the target vehicle at a first time, performs a prediction operation in turn according to the set iteration order, and outputs the level and the probability of the target vehicle having a fault at a second time, the second time is a time period after the first time by the set prediction time period, and the pre-trained fault prediction model is obtained based on historical running data of the target vehicle model having a contribution degree above a set value.

Images