A method for fault modeling and diagnosis of hybrid electric vehicles
By combining hybrid electric vehicle modeling and machine learning with swarm intelligence optimization algorithms, the problems of accuracy and efficiency in hybrid electric vehicle fault diagnosis have been solved, enabling efficient remote fault management and diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2023-05-29
- Publication Date
- 2026-06-30
AI Technical Summary
In the current technology, fault diagnosis of hybrid vehicles relies on human experience and instrument diagnosis, which is not very accurate and the management of vehicle fault tracing is lagging behind, making it impossible to achieve efficient and accurate whole vehicle fault diagnosis.
By using hybrid electric vehicle whole vehicle modeling, combined with machine learning and swarm intelligence optimization algorithms, and using convolutional neural networks for fault data diagnosis, a client and server system is established to achieve remote fault management.
It improves the accuracy and efficiency of fault diagnosis, enables users to query vehicle status information and meet enterprise management needs independently, and reduces diagnostic time and reliance on human experience.
Smart Images
Figure CN116400674B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of automotive fault diagnosis technology, and more specifically, to a method for modeling and diagnosing faults in a hybrid electric vehicle. Background Technology
[0002] Currently, vigorously developing new energy vehicles has become an important direction for major automakers. In the future automotive market, hybrid vehicles, as an important technological direction of new energy vehicles, will occupy a large proportion of the market.
[0003] The rapid development of hybrid vehicles has driven the widespread application of fault diagnosis technology. Early fault diagnosis relied heavily on manual experience, placing high demands on the technicians' skills and thus having limitations. Later developments utilized diagnostic instruments such as multimeters, oscilloscopes, engine diagnostic tools, and decoders. While instrument-based diagnosis improved the efficiency of technicians, mastering these instruments requires a high level of theoretical knowledge to fully utilize their functions, thus still presenting limitations. Furthermore, both manual and instrument-based diagnosis suffer from delays in subsequent vehicle fault tracing and management.
[0004] A search revealed Chinese invention patent CN113657442A, which discloses a fault diagnosis method, device, and storage medium for electric vehicle charging equipment. This application acquires key data from the charging equipment, extracts fault data features based on a convolutional neural network, and then inputs the multi-dimensional features into a backpropagation (BP) neural network to achieve fault diagnosis of the charging equipment. However, this application only diagnoses the charging equipment and does not address fault diagnosis of the entire vehicle, thus having certain limitations.
[0005] Chinese invention patent CN114167846A discloses a remote fault diagnosis system and method for new energy vehicles. This application collects real-time vehicle status information through a controller, transmits the information to a forwarding controller, and then to a TBOX controller. The controller uploads the information to a backend server for remote vehicle fault diagnosis. However, this application still relies on instrument-based diagnostics, which is time-consuming and its accuracy is difficult to guarantee. Summary of the Invention
[0006] 1. The technical problem that the invention aims to solve
[0007] The purpose of this invention is to overcome the problems of low accuracy and lagging vehicle fault tracing management in existing technologies that rely on human experience and diagnostic instruments for fault diagnosis. This invention provides a method for modeling and diagnosing faults in hybrid electric vehicles. This invention can collect corresponding fault data based on the fault phenomena input in vehicle modeling, and output diagnostic results from the collected fault data through machine learning combined with swarm intelligence optimization algorithms. Furthermore, it combines this data with front-end and back-end computer technologies to achieve remote management of vehicle fault phenomena, thus addressing the practical needs of users and enterprises.
[0008] 2. Technical Solution
[0009] To achieve the above objectives, the technical solution provided by the present invention is as follows:
[0010] The present invention provides a method for fault modeling and diagnosis of a hybrid electric vehicle, comprising the following steps:
[0011] Step 1: Model the hybrid vehicle as a whole to collect data on key components under standard input conditions;
[0012] Step 2: Conduct feasibility verification on the completed hybrid vehicle model to ensure that the model can accurately simulate faults;
[0013] Step 3: Run the model and collect raw fault data;
[0014] Step 4: Perform dimensionality reduction analysis on the original fault data. Through linear projection, map the high-dimensional data to a low-dimensional data space, select the feature information with high correlation, and remove redundant features.
[0015] Step 5: Establish a convolutional neural network, establish a one-to-one correspondence between the fault type and the output layer variables of the convolutional neural network, and use the Seagull algorithm to iteratively optimize the number of neurons and the learning rate of the fully connected layer.
[0016] Step 6: Use the trained neural network to perform diagnosis and output the fault classification results.
[0017] Furthermore, the model established in step one includes: a driver model, an engine model, a generator model, a battery model, a transmission model, and a driving dynamics model. Among them, the driver model is established using the PID principle, the engine model and the generator model are established using experimental data, and the Rint equivalent circuit model is used as the battery model.
[0018] Furthermore, the input standard operating condition CLTC-P was adopted as the test condition for the model and subsequent research.
[0019] Furthermore, during the model execution in step three, normal data is first collected over a period of time. Then, fault injection of the accelerator pedal is performed to obtain faulty sensor voltage data. Next, fault injection of the motor is performed to obtain abnormal motor speed data. Finally, the two faults are simulated simultaneously to obtain abnormal fault data.
[0020] Furthermore, suppose that M fault data samples were collected in step three. Each sample has N-dimensional features. Each feature also has its own feature value; the fourth step, dimensionality reduction analysis, is as follows:
[0021] (1) Decentralize all features, i.e., calculate the average value;
[0022] (2) Find the covariance matrix C of the M samples under N-dimensional features;
[0023] (3) Calculate the eigenvalues of the covariance matrix and its corresponding eigenvectors, and then convert the eigenvalues into... λ Sort by size and select the first n features;
[0024] (4) Project the original features onto the selected feature vectors to obtain the new dimensionality-reduced features. k Dimensional features.
[0025] Furthermore, the convolutional neural network established in step five includes convolutional layers, max pooling layers, and fully connected layers. After inputting fault sample data, it enters the first convolutional layer, then the max pooling layer, then the second convolutional layer, then the max pooling layer, and finally the fully connected layer and the output layer.
[0026] Furthermore, the first convolutional layer uses 8 convolutional kernels with a stride of 1; the second convolutional layer uses 16 convolutional kernels with a stride of 1; and the convolutional layers use the ReLU activation function.
[0027] Furthermore, the max pooling layer uses 2 A filter with a step size of 2 is used for pooling operations. The dropout method is used to deactivate neurons in each layer of the fully connected layer with a probability of 0.5. The Tanh activation function is used in the fully connected layer.
[0028] Furthermore, the Seagull Algorithm is introduced during the model training phase to find the optimal linear combination of hyperparameters of the neural network through iterative optimization.
[0029] The present invention provides a hybrid electric vehicle fault modeling and diagnosis system, comprising a client and a server, wherein the client includes:
[0030] The user information module creates and stores a user information table, recording the user's basic information.
[0031] The vehicle registration module creates and stores a vehicle information table, recording the vehicle's basic information.
[0032] The technician management module records the vehicle's maintenance personnel, making it easier for users to contact them later.
[0033] The fault collection module records all information about vehicle maintenance, making it easy for users to understand the overall condition of the vehicle.
[0034] The server-side includes:
[0035] Fault data and diagnostic results are stored in tables in MySQL, which allows administrators to quickly view and modify them, and facilitates data sharing on browser platforms;
[0036] The fault diagnosis model results are communicated with the system platform server via the HTTP protocol to upload data and send diagnostic results. Static web pages are supported by Apache servers, while dynamic Java applications are supported by Tomcat servers.
[0037] 3. Beneficial effects
[0038] Compared with existing known technologies, the technical solution provided by this invention has the following significant advantages:
[0039] (1) This invention establishes a vehicle fault model using MATLAB / Simulink, collects corresponding fault data based on the fault phenomena input in the vehicle modeling, and outputs diagnostic results on the collected fault data through machine learning combined with swarm intelligence optimization algorithm. It also combines it with computer front-end and back-end technologies to realize remote management of vehicle fault phenomena, which is convenient for users to query and understand vehicle status information on their own, and can also meet the actual needs of enterprise fault diagnosis and management.
[0040] (2) The present invention performs dimensionality reduction processing on fault data, inputs the collected fault data into a convolutional neural network for training, and uses the SOA algorithm to optimize the hyperparameters of the neural network. Compared with the use of genetic algorithm to optimize the BP neural network, the accuracy of fault diagnosis is improved. Attached Figure Description
[0041] Figure 1(a) and Figure 1(b) are the distribution diagrams of the engine operating points and the motor operating points, respectively.
[0042] Figure 2 A graph showing the proportion of feature information after dimensionality reduction processing of fault data.
[0043] Figure 3 This is a diagram of the convolutional neural network structure.
[0044] Figure 4 A flowchart for optimizing the neural network for the Seagull algorithm.
[0045] Figure 5 The graph shows the fitness function of the PCA-SOA-CNN model.
[0046] Figures 6(a) and 6(b) show the diagnostic classification diagrams of the PCA-SOA-CNN model.
[0047] Figures 7(a) and 7(b) are confusion matrices of the diagnostic results of the PCA-SOA-CNN model.
[0048] Figure 8 This is an architecture diagram of a fault diagnosis system based on a B / S architecture.
[0049] Figure 9 Flowchart for fault modeling and diagnosis. Detailed Implementation
[0050] Combination Figure 9 The present invention provides a method for fault modeling and diagnosis of a hybrid electric vehicle, comprising the following steps:
[0051] Step 1: Perform overall modeling of the hybrid vehicle in MATLAB / Simulink to collect data on key components under standard input conditions.
[0052] The overall structure of a hybrid electric vehicle mainly consists of a driver model, engine, generator, battery, transmission, and driveshaft. Since the driver's behavior changes with weather, road conditions, and other factors, it cannot be described by a precise mathematical model. Therefore, this embodiment uses the PID principle to establish the driver model.
[0053] During driving, the driver will adjust the vehicle speed according to the real-time needs of the vehicle. v req With the target required vehicle speed v aim deviation e ( t The state of the accelerator and brake pedals can be determined by the following formula:
[0054]
[0055] in, T req For the required torque, k p , k i , k d Let be the proportional, integral, and differential constants.
[0056] The engine is one of the important components that provides power output to a car. In this embodiment, engine experimental data is used for modeling. That is, some key performance parameters, such as torque, speed and fuel consumption, are obtained through engine experimental tests and represented in the engine characteristic diagram by a two-dimensional lookup table to obtain the engine model.
[0057] Based on the engine's torque and speed, the output power and fuel consumption can be calculated:
[0058]
[0059] In the formula, P e Engine power, in kW; m Engine fuel consumption, in grams; T e Engine torque, N·m; b e Engine fuel consumption rate, g / kW·h; ω e Engine speed, r·min -1 .
[0060] As a crucial component of hybrid vehicles, the electric motor powers the vehicle's electrical systems and charges the battery when its charge is low. This embodiment also uses experimental data for modeling. The electric motor efficiency is a function of its operating speed and torque, and can be obtained through a two-dimensional lookup table using the motor's numerical model diagram. The output power of the electric motor can be expressed as:
[0061]
[0062] In the formula, P m The output power of the electric motor is expressed in kW. ω m The motor speed is expressed in r·min. -1 ; T m For motor torque, N·m; when T m When >0, it indicates that the motor is in motor operating mode. T m When <0, it indicates that the motor is in generator operating mode; This refers to the operating efficiency of the motor.
[0063] The battery is a key component for maintaining a vehicle's range. It can be charged by connecting to the grid or obtain electrical energy through regenerative braking to power the vehicle's alternator. The Rint equivalent circuit model can effectively describe the battery's characteristics, reflecting its input and output relationship. The battery's State of Charge (SOC) value can be expressed by the following formula:
[0064]
[0065] In the formula, SOC 0 is the initial value of SOC; η in , η out These are the battery's charge and discharge efficiencies, respectively. C Indicates battery capacity, Ah; I b The current is the battery current, in A.
[0066] The transmission is also an important component of a car's powertrain, capable of changing the gear ratio and expanding the range of speed and torque variation for the drive wheels. The torque and speed input to the transmission can be expressed by the following formula:
[0067]
[0068]
[0069] In the formula, T in The torque of the gearbox input shaft is N·m; ω in The rotational speed of the gearbox input shaft, in rad / s; ω The rotational speed at the wheel is rad / s; T The torque at the wheel is N·m; i 0 represents the overall gear ratio of the vehicle; η This refers to the mechanical transmission efficiency from the torque coupling device to the wheel.
[0070] Finally, a driving dynamics model for hybrid vehicles needs to be established. Considering objective conditions such as windless weather and normal roads, the vehicle's driving equation formula can be established:
[0071]
[0072] In the formula, F t The driving force of a vehicle; F f The rolling resistance experienced by the vehicle; F w The air resistance experienced by the vehicle; F i The gradient resistance experienced by the vehicle; F j The acceleration resistance experienced by the vehicle; T tq Engine torque, expressed in N·m; i g Indicates the gear ratio of the transmission;i 0 indicates the gear ratio of the main reducer; f Expressed as the rolling resistance coefficient; C D This refers to the air drag coefficient; A For the windward area, m 2 ; u a The speed of the car, m·s -1 ; δ This represents the rotational mass conversion factor.
[0073] Step 2: Conduct feasibility verification on the completed hybrid vehicle model, verifying the reliability of components such as the engine, motor, and battery, to ensure that the model can be used to accurately simulate faults in the future.
[0074] This embodiment uses CLTC-P as the model and the test condition for subsequent research. The total cycle time is 1800s, and the cumulative mileage is 14.48km. The standard CLTC-P condition is input, and first, normal data is collected for a period of time. Then, a failure simulation of the accelerator pedal sensor is performed, and corresponding failure data is collected. Next, a failure simulation of the motor speed sensor is performed, and corresponding failure data is collected. Finally, a simultaneous occurrence of both failures is simulated, and corresponding failure data is collected.
[0075] Step 3: Run the model and collect raw fault data.
[0076] When the model is running, it operates normally from 0 to 4 seconds, during which the sensors collect various data during normal operation. From 6 to 8 seconds, a fault is injected into the accelerator pedal to obtain faulty sensor voltage data. From 7 to 9 seconds, a fault is injected into the motor to obtain abnormal motor speed data. From 9 to 11 seconds, two faults are simulated simultaneously to obtain abnormal fault data.
[0077] Step 4: Perform dimensionality reduction analysis on the original fault data. While ensuring as much information as possible, use linear projection to map the high-dimensional data to a low-dimensional data space, select features with high correlation, and remove redundant features. In this way, more characteristics of the original data are retained while using fewer data features.
[0078] Suppose that M fault data samples were collected in step three. Each sample has N-dimensional features. Each feature also has its own feature value.
[0079] (1) All features need to be decentralized, as shown in the following formula:
[0080]
[0081] (2) Find the covariance matrix C of the M samples under N-dimensional features. The solution formula is as follows:
[0082]
[0083] (3) Calculate the eigenvalues and corresponding eigenvectors of the covariance matrix. According to matrix theory, each eigenvalue has an eigenvector, as shown in the following formula:
[0084]
[0085] eigenvalues λ Sort by size and select the largest feature. In this embodiment, the top 3 are selected because their features account for more than 90% of the information.
[0086] (4) Project the original features onto the selected feature vectors to obtain the new dimensionality-reduced features. k Dimensional features.
[0087] Each sample X i The original characteristics are The new features obtained by projection are The formula for calculating the new features is as follows:
[0088]
[0089] Through the above dimensionality reduction calculation process, the sample X i From the original It has become what it is now. This achieves the goal of dimensionality reduction. From the obtained fault sample data, 400 samples are randomly selected and divided into a test set and a validation set in a 7:3 ratio to facilitate subsequent training of the neural network.
[0090] Step 5: Establish a convolutional neural network, and establish a one-to-one correspondence between the fault type and the output layer variables of the convolutional neural network. Each initial individual represents a local optimum of the problem.
[0091] Combination Figure 3 The convolutional neural network includes a convolutional layer, a max pooling layer, and a fully connected layer. After inputting fault sample data, it enters the first convolutional layer, then the max pooling layer, then the second convolutional layer, then the max pooling layer, and then the fully connected layer and the output layer in sequence.
[0092] Convolutional layers extract features using values within a sliding window. The first convolutional layer has 8 kernels, and the second has 16 kernels. Common activation functions include Sigmoid, Tanh, and ReLU. However, Sigmoid is prone to saturation and can cause gradient propagation to terminate. Therefore, in this embodiment, the convolutional layers use the ReLU function, which has the advantages of simple gradient calculation and fast convergence. The ReLU activation function expression is:
[0093]
[0094] After the convolutional layers obtain features, max pooling layers are used for feature compression, and statistical aggregation is performed on features at different locations. The max pooling layers use 2... A filter with a step size of 2 is used for pooling operations. To prevent overfitting of the neural network, the dropout method is used to deactivate neurons in each of the fully connected layers with a probability of 0.5, increasing the orthogonality of features in each layer.
[0095] Furthermore, this embodiment utilizes the seagull optimization algorithm (SOA) to iteratively optimize the number of neurons and learning rate in the fully connected layer, improving the accuracy of fault diagnosis and reducing training time. The SOA algorithm consists of two behavioral modes: migration (global search) and attack (local search). During migration, the algorithm simulates how a flock of seagulls moves from one location to another. During migration, individual seagulls need to meet three conditions:
[0096] 1) To avoid collisions, a variable A is introduced into the algorithm to calculate the updated position of the seagull, as shown in the following formula:
[0097]
[0098] In the formula, C s ( t This indicates a new location where the individual will not collide with other seagulls; P s ( t ) represents the seagull's current position; t represents the current iteration number; A This represents the movement behavior of an individual within the search space; f c This represents the control factor, and its value is typically 2. T This represents the maximum number of iterations.
[0099] 2) After avoiding collisions with other seagulls, they will move towards the optimal position, as expressed by the formula:
[0100]
[0101] In the formula, M s ( t () indicates the relative direction of the best position within the seagull population; B Represents a random number responsible for global balancing and local search; r It is a random number in the range [0,1].
[0102] 3) In the process of avoiding conflict, individual seagulls will gradually move towards the optimal position and reach a new location. The formula is as follows:
[0103]
[0104] After spotting its prey, the seagull spirals downwards in the air to strike, and its trajectory can be represented by the formula:
[0105]
[0106] In the formula, r Indicates the helix radius; θ Represents a random value within the range [0, 2π]. u and v It is a constant with a spiral shape; the attack position of the seagull can be represented as:
[0107]
[0108] The SOA algorithm is initialized with parameters, and the hyperparameters of the neural network, including the number of hidden layer nodes, the maximum number of iterations, and the learning rate, are optimized. The fault diagnosis neural network is trained by inputting the dimensionality-reduced fault dataset, and the root mean square error is selected as the fitness function to calculate the fitness value of individuals in the population to minimize the root mean square error. If the maximum number of iterations is reached, the iteration terminates; otherwise, the above steps are repeated. The linear combination of the output hyperparameters is taken as the optimal solution.
[0109] Step 6: Use the trained neural network to perform diagnosis and output the fault classification results.
[0110] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments.
[0111] Example 1
[0112] By setting the parameters of the hybrid electric vehicle as shown in Table 1 below, the driver model, engine, motor, battery and other components are modeled in MATLAB / Simulink, thus forming the whole vehicle model of the hybrid electric vehicle.
[0113] Table 1 Key Modeling Parameters
[0114]
[0115] In this embodiment, the model is designed to simulate accelerator pedal pressure sensor failure and motor speed sensor failure. A total of 400 sets of data are collected, including accelerator pedal voltage signal, motor speed, engine speed, generator speed, and vehicle speed. Some of the fault data collected by the sensors are shown in Table 2.
[0116] Table 2 Partial Fault Data
[0117]
[0118] X 1~ X 5 represents the accelerator pedal voltage signal, motor speed, engine speed, generator speed, and vehicle speed, respectively, serving as inputs to the diagnostic algorithm; Labels 1-4 are outputs, representing four types of faults: normal, accelerator pedal pressure sensor fault, motor speed sensor fault, and both faults occurring simultaneously, with 100 sets for each fault type. Some fault data are shown in Table 2. The accelerator pedal signal is the measured voltage value, with a voltage range between 0 and 2.5V; the generator speed can be positive or negative, indicating that the battery can be charged.
[0119] Figure 1(a) and Figure 1(b) are the engine operating characteristic diagram and the motor operating characteristic diagram obtained from the existing experimental data, respectively. It can be seen that the engine efficiency and motor efficiency are both within the normal iso-efficiency curve, indicating the effectiveness of the whole vehicle model.
[0120] Further dimensionality reduction processing of the collected fault data can yield... Figure 2 The information proportion diagram shown indicates that features 1, 2, and 3 contain more than 90% of the information in the overall data. Therefore, the high-dimensional original fault data is projected into low-dimensional data through dimensionality reduction, as shown in Table 3 below.
[0121] Table 3. Fault data after partial dimensionality reduction
[0122]
[0123] like Figure 3 The diagram shows the convolutional neural network structure proposed in this invention. The structural parameters are shown in Table 4 below. It includes an input layer, hidden layers, and an output layer. The hidden layers further include convolutional layers, max-pooling layers, and fully connected layers. Data enters the neural network, undergoes convolution using the sliding window method, is activated by the ReLU function, enters a max-pooling layer, then enters another convolutional layer, is activated again, enters another max-pooling layer, and finally, after a Dropout layer, the diagnostic results are classified and output.
[0124] In CNNs, the convolutional layer is a very important layer. The design of the convolutional kernel is closely related to the extraction of data features. The convolutional kernel is essentially a weighted matrix. It uses the sliding window method to perform a weighted sum on the input data and outputs the matrix to the next layer through the activation function.
[0125] To achieve optimal results, this invention employs two convolutional layers. Considering the features of the fault data collected in the model, the first convolutional layer uses 8 kernels with a stride of 1; the second convolutional layer uses 16 kernels with a stride of 1. The fault data undergoes convolutional processing through these two layers to collect effective features from the samples, which are then combined for output. Each convolutional layer is followed by a max-pooling layer to extract data. The two max-pooling layers in this invention have a dimension of 2. 2, step size is 2.
[0126] After passing through two convolutional and pooling layers, the fault data enters the fully connected layer, where the loss function of the neural network is calculated, and the diagnostic results are classified and output. The Tanh function is used in the fully connected layer, which can solve the problem of zero-mean output of the Sigmoid function, enabling the diagnostic classification of fault data. Furthermore, to avoid overfitting in the convolutional neural network, the Dropout method is added before the fully connected layer. The weights of neurons in the fully connected layer are set to 0 with a probability of 0.5, which is equivalent to randomly deleting the number of hidden layer neurons in the neural network. This reduces the network's sensitivity to small data changes and increases the orthogonality of features in each layer.
[0127] Table 4 Convolutional Neural Network Structure Parameters
[0128]
[0129] The Tanh activation function used in the fully connected layer is expressed as follows:
[0130]
[0131] In addition, the training of neural networks involves adjusting hyperparameters. Manual adjustment relies too much on expert experience, which can lead to poor model training results and prolong diagnostic time. Therefore, the Seagull Algorithm is introduced during the model training phase to find the optimal linear combination of hyperparameters through intelligent algorithm iteration.
[0132] like Figure 4 The diagram shown is a flowchart of the training process for the SOA-CNN of this invention. The specific steps for training the fault diagnosis model based on the Seagull Algorithm to optimize the convolutional neural network are as follows:
[0133] 1) Determine the hyperparameters and ranges that need to be optimized for the convolutional neural network, and initialize the parameters of the Seagull algorithm;
[0134] 2) Input the training set and train the neural network. In this invention, the mean square error between the predicted and actual results of fault diagnosis is used as the fitness function.
[0135] 3) Perform migration and attack operations on individual seagulls in the population, continuously update the location, step size, and distance of individual seagulls, and generate the latest seagull positions;
[0136] 4) Calculate the fitness function value of each individual seagull in the current population, and select the minimum fitness function value as the optimal seagull position;
[0137] 5) Check if the maximum number of iterations has been reached; otherwise, proceed to steps 3 and 4 to continue running.
[0138] 6) Output the optimal linear combination, substitute the optimized value into the convolutional neural network for training, and input the test set to obtain the diagnostic results.
[0139] The algorithm optimizes the hyperparameters of the neural network to obtain a kernel size of 4 for the first convolutional layer. 4. The kernel size of the second convolutional layer is 5. 5. The maximum number of iterations is 343, and the learning rate is 0.001.
[0140] To verify the diagnostic effectiveness of the SOA-CNN of this invention, the experiment was conducted on a computer with 16GB of memory, an AMD R7-5800H processor, and 3.2GHz RAM. The model was trained on MATLAB R2019b.
[0141] like Figure 5The figure shows the number of iterations of the SOA-CNN algorithm in this invention. It can be seen that the SOA-CNN algorithm escaped the local optimum and obtained the global optimum, and reached the minimum root mean square error in the 11th iteration, outputting the diagnostic classification result. This invention conducted four sets of comparative experiments to verify the diagnostic effects of the BP, GA-BP, CNN, and SOA-CNN models respectively. The experiments compared the accuracy and confusion matrix of the four models. The dimensionality-reduced fault dataset was divided into test set and validation set in a 7:3 ratio. The diagnostic classification results of SOA-CNN are shown in Figure 6. Figure 6(a) shows the comparison of the classification results of the training set with an accuracy of 99.29%, and Figure 6(b) shows the comparison of the classification results of the test set with an accuracy of 97.5%. The confusion matrix of the diagnostic classification results is shown in Figure 7. Figure 7(a) is the confusion matrix of the training set, showing that 73 classes that actually belong to class 1 were predicted as class 1, 1 class that actually belonged to class 3 was predicted as class 2, 1 class that actually belonged to class 2 was predicted as class 3, and 63 classes that actually belonged to class 4 were predicted as class 4. Figure 7(b) is the confusion matrix of the test set, showing that 25 classes that actually belonged to class 1 were predicted as class 1, 26 classes that actually belonged to class 2 were predicted as class 2, 1 class that actually belonged to class 2 was predicted as class 3, and 2 classes that actually belonged to class 1 were predicted as class 4.
[0142] Table 5 Comparison of diagnostic results from three neural networks
[0143]
[0144] As can be seen from Table 5, in the four control experiments, the method used in this invention has the highest diagnostic accuracy, achieved good diagnostic results, and has considerable reliability.
[0145] Example 2
[0146] like Figure 8 As shown, this embodiment also provides a fault diagnosis system. This system adopts a B / S architecture and is built on the Windows 10 platform. The development software is Eclipse, the front-end is written in HTML and CSS, and the back-end technology is a Java servlet service program, which is used for interactive browsing and data generation, generating static and dynamic page content. The system framework adopts SpringMVC, and the running server is Apache Tomcat 8.0. User login information and vehicle information are established in MySQL, and then the diagnostic results are stored in the vehicle fault diagnosis system, which can be easily queried and modified by users or administrators.
[0147] The system includes: a server and a client, wherein...
[0148] The server-side includes:
[0149] Fault data and diagnostic results are stored in tables in MySQL, which allows administrators to quickly view and modify them, and facilitates data sharing on browser platforms;
[0150] The fault diagnosis model results are communicated with the system platform server via the HTTP protocol to upload data and send diagnostic results. Static web pages are supported by Apache servers, while dynamic Java applications are supported by Tomcat servers.
[0151] The system platform servers, Apache and Tomcat, handle front-end and back-end HTTP requests, complete the main content design of the pages, and perform CRUD operations on the back-end database. They also synchronously upload the diagnostic results of the fault model to the servers for easy management and retrieval later. Clients include:
[0152] The user information module establishes a user information table in the database, including key information such as username and password, and is divided into ordinary users and system administrators;
[0153] The vehicle registration module creates a vehicle information table in the database, recording basic vehicle information such as owner, license plate number, purchase date, and vehicle model. This allows maintenance personnel to quickly understand the basic information of the corresponding vehicle model, facilitating subsequent maintenance management.
[0154] The technician management module records the vehicle's repair personnel in the fault diagnosis system, making it easier for users to contact the repair personnel later and protecting the user's rights;
[0155] The fault collection module records all repair information for the vehicle, allowing users to understand its overall condition and facilitating subsequent used car price assessments and insurance claims.
[0156] Through the analysis of the overall architecture, the user information module, fault collection module, technician management module and database of the fault phenomenon management system were designed and completed. The local computer was used as the server of the fault phenomenon collection and management system, and the designed management system was ported to the local computer.
[0157] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.
Claims
1. A method for modeling and diagnosing faults in a hybrid electric vehicle, characterized in that, Includes the following steps: Step 1: Perform overall modeling of the hybrid vehicle to collect data on key components under standard input conditions. The model includes: driver model, engine model, generator model, battery model, transmission model, and driving dynamics model, among which: A driver model is built using the PID principle. During driving, the driver will adjust the vehicle speed according to the real-time demand of the vehicle. v req With the target required vehicle speed v aim deviation e ( t The states of the accelerator and brake pedals are determined by the following formula: in, T req For the required torque, k p , k i , k d These are proportional, integral, and differential constants; Key performance parameters, including torque, speed, and fuel consumption, are obtained through engine experimental testing and represented in the engine characteristic diagram using a two-dimensional lookup table to obtain the engine model. Based on the engine's torque and speed, the output power and fuel consumption can be calculated: In the formula, P e Engine power; m This refers to engine fuel consumption. T e This refers to engine torque; b e Engine fuel consumption rate; ω e Engine speed; The generator model is also modeled using experimental data, and the generator's output power is expressed as: In the formula, P m This refers to the output power of the electric motor; ω m This refers to the motor speed; T m For motor torque; when T m When >0, it indicates that the motor is in motor operating mode. T m When <0, it indicates that the motor is in generator operating mode; For the working efficiency of the motor; The battery model describes the battery's characteristics using a Rint equivalent circuit model, reflecting the battery's input and output relationship. The battery's SOC value is expressed by the following formula: In the formula, SOC 0 is the initial value of SOC; η in , η out These are the battery's charge and discharge efficiencies, respectively. C Indicates battery capacity; I b This refers to the battery current. The torque and speed input to the gearbox are expressed by the following formula: In the formula, T in This refers to the torque of the gearbox input shaft; ω in This refers to the rotational speed of the gearbox input shaft. ω The rotational speed at the wheel; T The torque at the wheel; i 0 represents the overall gear ratio of the vehicle; η The mechanical transmission efficiency from the torque coupling device to the wheel; The driving dynamics model, considering objective conditions such as windless weather and normal roads, establishes the vehicle's driving equation formula: In the formula, F t The driving force of a vehicle; F f The rolling resistance experienced by the vehicle; F w The air resistance experienced by the vehicle; F i The gradient resistance experienced by the vehicle; F j The acceleration resistance experienced by the vehicle; T tq Indicates engine torque; i g Indicates the gear ratio of the transmission; i 0 indicates the gear ratio of the main reducer; f Expressed as the rolling resistance coefficient; C D This refers to the air drag coefficient; A For windward area; u a The speed at which the car is traveling; δ Indicates the rotational mass conversion factor; Step 2: Conduct feasibility verification on the completed hybrid vehicle model to ensure that the model can accurately simulate faults; Step 3: Run the model and collect raw fault data; Step 4: Perform dimensionality reduction analysis on the original fault data. Through linear projection, map the high-dimensional data to a low-dimensional data space, select features with high correlation, and remove redundant features. Step 5: Establish a convolutional neural network, establish a one-to-one correspondence between the fault type and the output layer variables of the convolutional neural network, and use the Seagull algorithm to iteratively optimize the number of neurons and the learning rate of the fully connected layer. Step 6: Use the trained neural network to perform diagnosis and output the fault classification results.
2. The method for fault modeling and diagnosis of a hybrid electric vehicle according to claim 1, characterized in that: The input standard operating condition CLTC-P was used as the test condition for the model and subsequent research.
3. The method for modeling and diagnosing faults in a hybrid electric vehicle according to claim 2, characterized in that: In step three, when the model is running, first, normal data is collected for a period of time. Then, fault injection of the accelerator pedal is performed to obtain faulty sensor voltage data. Next, fault injection of the motor is performed to obtain abnormal motor speed data. Then, the two faults are simulated simultaneously to obtain abnormal fault data.
4. The method for fault modeling and diagnosis of a hybrid electric vehicle according to claim 3, characterized in that: Suppose that M fault data samples were collected in step three. Each sample has N-dimensional features. Each feature also has its own feature value; the fourth step, dimensionality reduction analysis, is as follows: (1) Decentralize all features, i.e., calculate the average value; (2) Calculate the covariance matrix C of the M samples under N-dimensional features; (3) Calculate the eigenvalues of the covariance matrix and its corresponding eigenvectors, and then convert the eigenvalues into... λ Sort by size and select the first n features; (4) Project the original features onto the selected feature vector to obtain the new dimensionality-reduced feature vector. k Dimensional features.
5. The method for modeling and diagnosing faults in a hybrid electric vehicle according to claim 4, characterized in that: The convolutional neural network established in step five includes convolutional layers, max pooling layers, and fully connected layers. After inputting fault sample data, it enters the first convolutional layer, then the max pooling layer, then the second convolutional layer, then the max pooling layer, and finally the fully connected layer and the output layer.
6. The method for modeling and diagnosing faults in a hybrid electric vehicle according to claim 5, characterized in that: The first convolutional layer uses 8 kernels with a stride of 1; the second convolutional layer uses 16 kernels with a stride of 1; the convolutional layers use the ReLU activation function.
7. The method for fault modeling and diagnosis of a hybrid electric vehicle according to claim 6, characterized in that: Max pooling layer uses 2 A filter with a step size of 2 is used for pooling operations. The dropout method is used to deactivate neurons in each layer of the fully connected layer with a probability of 0.
5. The Tanh activation function is used in the fully connected layer.
8. The method for fault modeling and diagnosis of a hybrid electric vehicle according to claim 7, characterized in that: The Seagull Algorithm is introduced during the training phase of the model to find the optimal linear combination of hyperparameters of the neural network through iterative optimization.