Method for remotely diagnosing power system fault of electric automobile

Abstract

The invention provides a method for remotely diagnosing a power system fault of an electric vehicle. The method comprises the following steps: collecting real-time data of a power system of the electric vehicle; transmitting the real-time data to a cloud platform; performing feature importance degree sorting on the real-time data by using a random forest model; adopting a simulated annealing algorithm to optimize hyper-parameters of the random forest model; deploying the optimized random forest model on a cloud platform for real-time fault classification and diagnosis; and a real-time fault alarm and diagnosis report of the power system of the electric vehicle is provided by using the cloud platform.

Description

Technical Field

[0001] The present invention relates to automatic fault diagnosis of a vehicle, and more particularly to fault detection of an electric vehicle, and in particular to a method for remotely diagnosing a power system fault of an electric vehicle. Background Art

[0002] Remote diagnosis of electric vehicle faults is a technology that uses communication technology and big data analysis to monitor and diagnose faults in electric vehicles in real time. This technology uses on-board sensors, data transmission modules, cloud servers and user interfaces. On-board sensors are responsible for collecting operating data of the electric vehicle power system, such as battery status, motor performance and control system parameters. This data is sent to the cloud server in real time through data transmission modules, such as wireless networks or cellular networks. In the cloud, specialized analysis software processes and analyzes the data, using machine learning and artificial intelligence algorithms to identify potential faults and signs of performance degradation. User interfaces, such as smartphone applications or web dashboards, allow car owners or maintenance technicians to remotely view vehicle status and receive fault alerts.

[0003] Due to their complex structure and harsh operating environment, electric vehicles are prone to various faults such as motor failure, DC / DC converter failure, battery failure, etc. Therefore, effective and real-time fault diagnosis is crucial to ensure the safety and reliability of vehicles. Summary of the invention

[0004] The first aspect of the present disclosure provides a method for remotely diagnosing a power system fault of an electric vehicle. In some embodiments, the method may include: collecting real-time data of the power system of the electric vehicle; transmitting the real-time data to a cloud platform; using a random forest model to sort the feature importance of the real-time data; using a simulated annealing algorithm to optimize the hyperparameters of the random forest model; deploying the optimized random forest model on a cloud platform for real-time fault classification and diagnosis; and using the cloud platform to provide real-time fault alarms and diagnostic reports of the power system of the electric vehicle.

[0005] In some embodiments, the random forest model can perform real-time fault classification and diagnosis based on the training set, and the importance of feature j in the training set is I j Calculated by formula 1:

[0006]

[0007] Among them, ΔGini(t,j) represents the reduction of the Gini index caused by feature j in decision tree t, and T is the total number of decision trees.

[0008] In some embodiments, the simulated annealing algorithm can optimize the hyperparameters of the random forest model through a cooling mechanism. The cooling mechanism can use a logarithmic cooling schedule.

[0009] In some embodiments, the logarithmic cooling schedule may be calculated by Equation 2:

[0010]

[0011] Among them, T n is the temperature after iteration, T i is the initial temperature, k is the logarithmic growth constant, γ is the adjustment factor, δ is the power factor, and iter is the current number of iterations.

[0012] In some embodiments, the real-time data may include data of a DC / DC converter, a motor, a battery of the electric vehicle. The real-time data may include voltage, current, temperature, pressure, or any combination thereof.

[0013] In some embodiments, the initial temperature may be 100 to 200, and the γ adjustment factor may be 0.3 to 0.7. Multiple iterations may be performed until the hyperparameters converge to an optimal solution.

[0014] In some embodiments, the AWS platform may include an AWS SageMaker Endpoint module and / or an AWS CloudWatch module.

[0015] It should be understood that the present disclosure is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The novel features of the invention are set forth with particularity in the appended claims.A better understanding of the features and advantages of the invention will be obtained by referring to the following detailed description and accompanying drawings which set forth illustrative embodiments utilizing the principles of the invention.

[0017] Figure 1 It is a flowchart of a remote fault diagnosis method according to an embodiment of the present application.

[0018] Figure 2 It is a schematic diagram of the process of optimizing the random forest model by simulated annealing according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] Traditional vehicle fault diagnosis methods mainly rely on standard fault codes generated by the vehicle computer (ECU) after a fault occurs. By analyzing these fault codes and related log information, technicians can determine the cause of the fault and repair it. However, this diagnostic method that relies on fault codes is limited in the types of faults it covers.

[0020] The disclosed method for remotely diagnosing electric vehicle faults can be based on real-time data from onboard sensors of electric vehicles, using machine learning algorithms to identify electric vehicle faults in real time, thereby timely pushing fault information to operators and / or maintenance personnel to improve driving safety. The disclosed method is particularly suitable for diagnosing faults in the power system of electric vehicles.

[0021] like Figure 1 As shown, the present disclosure provides a method for remotely diagnosing a power system fault of an electric vehicle. In some embodiments, the method may include steps S101-S106. In step S101, real-time data of the power system of the electric vehicle is collected from the on-board sensor. These real-time sensor data provide detailed information about the current operating state of the power system of the electric vehicle. The real-time data may include the voltage, current, temperature, state of charge and health state of the battery. The real-time data may include information of the battery, including the voltage, current, temperature, state of charge (SOC, State of Charge) and health state (SOH, State of Health) of the battery. The real-time data may also include information of the motor, including the temperature, speed and torque of the motor. In step S102, the real-time data of the power system of the electric vehicle collected is transmitted to the cloud platform. The real-time data may be transmitted in real time through the on-board diagnostic interface (OBD-II) or a dedicated telemetry device and an on-board communication module. The cloud platform may provide real-time data processing and deployment of machine learning models. Examples of cloud platforms include Amazon's AWS, Microsoft's Azure, Google Cloud, IBM Cloud, Oracle Cloud, Alibaba Cloud, Tencent Cloud and Huawei iCloud. For example, Amazon's AWS provides a wide range of machine learning services and tools, such as SageMaker. Microsoft's Azure platform provides Azure Machine Learning services to support developers in creating and training machine learning models. Google Cloud Platform provides AIPlatform, an integrated end-to-end machine learning platform for building and deploying models.

[0022] In step S103, the real-time data is sorted by feature importance using a machine learning algorithm on the cloud platform. In some instances, a random forest model can be used to sort the feature importance of the real-time data. This sorting can be achieved using a machine learning algorithm. Examples of machine learning algorithms include random forests, gradient boosting machines, and neural networks. Feature importance sorting can not only improve the prediction accuracy of the model, but also reduce the consumption of computing resources during the feature selection process, thereby improving the overall computing efficiency.

[0023] In some instances, a random forest model can be used to rank feature importance on real-time data. The random forest model improves the accuracy and stability of predictions by building multiple decision trees and combining their predictions. When processing real-time data, random forests can evaluate the influence of each feature on the model's prediction results, a process called feature importance evaluation. In this way, the model can not only identify the features that have the greatest impact on the prediction results, but also exclude those unimportant or noisy features, thereby optimizing the performance of the model. Feature importance ranking is particularly valuable in real-time systems because it can instantly reflect the dynamic changes in the data stream and support fast decision-making.

[0024] In some instances, the random forest model of the present disclosure performs fault classification and diagnosis based on a training set. In real-time fault classification and diagnosis applications, the random forest model first requires a training set, which contains historical data, including both normal operation data and data when faults occur. Each decision tree is trained on a random subset of the training set, which ensures the diversity of the model. During the classification process, a new data point will be judged by all decision trees separately, and then the final classification result will be determined by voting or averaging. In some instances, the importance of feature j in the training set of the present disclosure is I j Calculated by formula 1:

[0025]

[0026] Here, ΔGini(t,j) represents the reduction in the Gini index caused by feature j in decision tree t, and T is the total number of decision trees. That is, ΔGini(t,j) represents the change in Gini impurity when feature j is used for splitting in the tth tree. The importance of feature j is measured by averaging the change in Gini impurity in all trees. Gini impurity is a metric used to evaluate the purity of a node. The higher the purity of a node, the more the data at this node tends to belong to the same category.

[0027] In step S104, the hyperparameters of the random forest model are optimized. The hyperparameters may include the number of decision trees, the maximum depth of the decision tree, the minimum number of samples required for splitting a node, etc. By adjusting these parameters, the complexity and degree of fit of the model can be controlled, thereby preventing overfitting or underfitting. For example, increasing the number of decision trees can enhance the generalization ability of the model, but it may also increase the computational cost and the risk of overfitting. Adjusting the minimum number of split samples and the minimum number of leaf node samples can control the growth of the tree, avoid an overly complex model structure, thereby reducing overfitting and improving the model's adaptability to new data.

[0028] In some instances, simulated annealing is used to optimize the hyperparameters of a random forest model. When optimizing the hyperparameters of a random forest model, the simulated annealing algorithm randomly selects an initial solution for the hyperparameters, then makes a small random perturbation to the current solution at each iteration, and calculates the objective function value of the new solution. The simulated annealing algorithm can effectively explore in the global search space, and may jump out of the local optimal solution and eventually find a hyperparameter configuration close to the global optimal. Using the simulated annealing algorithm to optimize the hyperparameters of a random forest model (the number of decision trees, the depth of the decision trees, and the method of feature selection) can significantly improve the prediction accuracy and generalization ability of the random forest model on a specific data set.

[0029] In some instances, the simulated annealing algorithm optimizes the hyperparameters of the random forest model through a cooling mechanism. The cooling mechanism can use a logarithmic cooling schedule. Logarithmic cooling scheduling is a cooling mechanism that reduces the temperature according to a logarithmic function. This method can quickly explore a wide range of parameter spaces in the early stage, and search more carefully in the later stage to avoid falling into the local optimal solution too early. This scheduling method helps the algorithm to strike a balance between global search and local search, thereby finding a suitable compromise between computational efficiency and optimization ability. In some instances, the logarithmic cooling schedule is calculated by Equation 2:

[0030]

[0031] Here, T n is the temperature after iteration, T i is the initial temperature, k is the logarithmic growth constant, γ is the adjustment factor, δ is the power factor, and iter is the current number of iterations. γ and δ are used to further control the speed of temperature drop. In this way, the temperature will gradually decrease as the number of iterations increases, thereby increasing the probability of the algorithm finding the global optimal solution. T i It is the temperature at the beginning of the algorithm. It is usually set to a higher value to have a larger search space in the early stage to avoid falling into the local optimal solution. The lower temperature in the later stage is helpful for fine-tuning and finding the global optimal solution. As the number of iterations (iter) increases, the value of the logarithmic function will gradually increase, so that the temperature Tn Gradually decrease. This approach ensures a larger search space in the early stage and gradually converges in the later stage. γ and δ are used to further fine-tune the speed of temperature drop. By adjusting the values ​​of γ and δ, the temperature drop curve can be controlled to better meet the needs of specific problems.

[0032] In step S105, the optimized random forest model is deployed on the cloud platform for real-time fault classification and diagnosis. The optimized random forest model can process large amounts of data more efficiently, reduce the risk of overfitting, and improve the accuracy of fault detection. Deploying the optimized model on the cloud platform can utilize the powerful computing power and scalability of cloud computing to achieve real-time data processing and analysis. Through the cloud platform, the model can continuously receive data and update its predictions in real time, thereby achieving continuous monitoring of the power system status of the electric vehicle and instant fault diagnosis.

[0033] In step S106, the cloud platform with the optimized random forest model deployed is used to provide real-time fault alarms and diagnostic reports of the electric vehicle's power system. When the cloud platform detects an abnormality in the electric vehicle's power system, it immediately sends a fault alarm to the owner and provides a detailed diagnostic report. These reports not only point out the problem, but also provide possible solutions to help the owner or maintenance technician quickly locate and solve the problem.

[0034] The following non-limiting embodiments of the present disclosure provide examples of methods for remotely diagnosing power system faults of electric vehicles using machine learning algorithms. In these embodiments, different logarithmic cooling scheduling parameters are used in the cooling mechanism of the simulated annealing algorithm for optimizing the random forest model, and the optimization effects of these parameters on the simulated annealing algorithm are compared.

[0035] Example 1

[0036] Reference Figure 1 The method disclosed in the present invention for remotely diagnosing power system faults of electric vehicles is described. In step S101, real-time data of the power system of the electric vehicle is collected from the on-board data acquisition device (such as sensors, buses, etc.), including the DC / DC output voltage (DCHV), the total battery pack voltage (PackVolt) and the motor current (MotorC). The DC / DC output voltage (DCHV) is the output voltage of the DC-to-DC converter. The total battery pack voltage (PackVolt) reflects the energy state of the battery pack. The motor current (MotorC) is used to evaluate the working efficiency and power consumption of the motor. In exemplary embodiment 1, the real-time data collected includes: DC / DC output voltage (DCHV): 330V; total battery pack voltage (PackVolt): 700V; motor current (MotorC): 80A.

[0037] In step S102, the collected real-time data of the power system of the electric vehicle is transmitted to a cloud platform, such as an AWS cloud platform, via the Internet or a communication network. For example, the real-time data of the power system of the electric vehicle can be transmitted to the AWS IoT Core of the AWS cloud platform via the MQTT and HTTPS protocols, and the data is stored in AWS S3 for storage. Since the collected initial data usually contains noise, the data quality can be improved by data preprocessing (e.g., filtering and denoising). The preprocessing of the initial data can be performed on the vehicle side or in the cloud. For example, AWS IoT Core can implement real-time analysis and processing of data, such as filtering and denoising algorithms. Once the data is cleaned and preprocessed, it can be integrated into a data set that can be used for machine learning and data analysis, such as the feature importance ranking described in the present disclosure.

[0038] In step S103, the real-time data is sorted by feature importance using a random forest model. Random forest is an ensemble learning method that makes predictions by constructing multiple decision trees.

[0039] Methods for building a random forest model can include:

[0040] Prepare the input dataset D: The input dataset D contains multiple samples, each of which consists of a feature vector xi and a corresponding label yi. That is, the original dataset D = {(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x n ,y n )}, where x i =(x i1 ,x i2 ,...,x im ) is the feature vector of the i-th sample, y i is the corresponding label.

[0041] Random sampling: Multiple samples are randomly selected from the original data set with replacement to construct training subsets. Each training subset is used to build a decision tree.

[0042] Feature subset selection: For each decision tree, a subset of features (feature subsets) is randomly selected from the feature set to perform node splitting. Usually, features, where m is the total number of features. This reduces the risk of overfitting in each tree while maintaining the predictive power of the entire model.

[0043] Decision tree construction: Build a decision tree using each training subset. When building a decision tree, each node split is based on the selected feature subset. These decision trees will be built independently and will not affect each other. The decision tree construction process can use the Gini Index to select the optimal splitting feature. The formula for the Gini Index is:

[0044]

[0045] Here, q i represents the proportion of category i in the current node, and C is the number of categories. The goal is to select features that can minimize the Gini index to split the node.

[0046] Iterative process: Repeat the above steps k times to generate k decision trees.

[0047] In the embodiments of the present disclosure, the initial random forest model is trained by a training data set before deployment. The training data set can be derived from historical fault data of electric vehicles, or obtained from experimental tests and simulation data of electric vehicles. These data are used to train the model to identify various fault types.

[0048] In Embodiment 1 of the present disclosure, preparing the input data set D may include:

[0049] Determine features and labels: In the random forest model, the feature vector x of each sample i Represents an observed parameter or measurement of the system. In Example 1, the features may include: [x i1 :DC / DC output voltage (DCHV) = 330V]; [x i2 : Total battery pack voltage (PackVolt) = 700V]; [x i3 :Motor current (MotorC) = 80A]. Label y i is the same as the eigenvector x i The corresponding target value or classification result. In fault diagnosis, the label y i It can be a binary value indicating whether a fault exists (e.g., 0 for no fault, 1 for fault), or a categorical value indicating different fault types.

[0050] Construct the input data set: Each sample in the data set includes a feature vector x i and a label y i If there are multiple samples from different monitoring data, each sample forms a row in the dataset. Example of a single sample: D = [x 1 =(330,700,80)(feature vector of a sample); y 1=0 (assuming the label of this sample is “no fault”). Example of multiple samples: D = {((330,700,80),0),((340,710,85),1),...,((320,680,75),0)}.

[0051] Subsequently, the random forest model is trained using the input dataset D. Each feature vector x i The decision tree will be used in the random forest to predict the label y i . Tag y i Pertaining to fault detection or classification.

[0052] Once the random forest model is trained, it can receive new monitoring data as input (e.g., new values ​​of DCHV, PackVolt, MotorC) and predict the fault state (label y i ).

[0053] In the decision tree algorithm of the present disclosure, the evaluation of feature importance is performed by observing the role of the feature in the splitting of the tree node. If a feature can significantly reduce the impurity at the split point, such as Gini impurity or information entropy, then this feature is important for the prediction of the model. In some examples, the importance of feature j in the training set of the present disclosure is I j Calculated by formula 1:

[0054]

[0055] Here, ΔGini(t,j) represents the reduction in the Gini index caused by feature j in decision tree t, and T is the total number of decision trees. That is, ΔGini(t,j) represents the change in Gini impurity when feature j is used for splitting in the tth tree. The importance of feature j is measured by averaging the change in Gini impurity in all trees.

[0056] In step S104, the hyperparameters of the random forest model are optimized. In some embodiments, the performance of the random forest model can be optimized by adjusting the following parameters: the number of trees in the forest (the number of decision trees); the minimum number of samples required for a node to split (the minimum number of split samples); and the minimum number of samples required for a leaf node (the minimum number of leaf node samples). These parameters directly affect the complexity, accuracy, and sensitivity to different features of the model. By reasonably adjusting these parameters, the performance of the random forest model can be improved and a more accurate ranking of feature importance can be obtained. In some embodiments, the number of decision trees is 100, the minimum number of split samples is 5, and the minimum number of leaf node samples is 2.

[0057] In some instances, a simulated annealing algorithm is used to optimize the hyperparameters of the random forest model. In some instances, the simulated annealing algorithm optimizes the hyperparameters of the random forest model through a cooling mechanism. The cooling mechanism can use a logarithmic cooling schedule. In some instances, the logarithmic cooling schedule is calculated by Formula 2:

[0058]

[0059] Here, T n is the temperature after iteration, T i is the initial temperature, k is the logarithmic growth constant, γ is the adjustment factor, δ is the power factor, and iter is the current iteration number. γ and δ are used to further control the speed of temperature drop. In some examples, the initial temperature T i =100, logarithmic growth constant k=0.1, adjustment factor γ=0.5, power factor δ=0.2. After 10 rounds of iterations, the final optimization results of the above related parameters in the random forest model are: the number of decision trees is 150, the minimum number of split samples is 4, and the minimum number of leaf node samples is 3. Therefore, the fault classification accuracy of the optimized random forest model is about 97.5%.

[0060] Figure 2 A method for optimizing the parameters of a random forest model using a logarithmic cooling schedule is shown. The method may include steps S201-S211. In step S201, the model parameters are initialized. For example, parameters such as the initial temperature, logarithmic growth constant, adjustment factor, power factor, and number of iterations may be initialized. In step S202, the random forest model parameters, including the number of decision trees, the minimum number of split samples, and the minimum number of leaf node samples, are initialized. In step S203, the random forest model parameters are initialized using the logarithmic cooling schedule for each round of iteration.

[0061] Formula 2 calculates the current temperature T n :

[0062]

[0063] Here, iter is the current iteration round.

[0064] In step S204, the initial parameter settings of the random forest model are generated, for example, the number of decision trees, the minimum number of split samples, and the minimum number of leaf node samples. In step S205, the model performance of the current parameter settings is evaluated. Cross-validation or other performance evaluation methods can be used to evaluate the model performance of the current parameter settings. In the cross-validation method, the data set is divided into several parts, and then one part is used for model training, and the rest is used for testing to evaluate the generalization ability of the model.

[0065] In step S206, a neighborhood search or random perturbation is performed based on the current parameter settings to generate new parameter settings. Neighborhood search is a small-scale exploration around the current parameters to find parameter settings that may improve model performance. Random perturbation is to add random factors on the basis of the current parameter settings to jump out of the local optimal solution and increase the possibility of finding the global optimal solution. In step S207, the model performance of the new parameter settings is evaluated. Evaluating model performance can include using a validation data set to test the predictive ability of the model to ensure that the new parameter settings can improve the accuracy, efficiency and generalization ability of the model. The evaluation process can include calculating performance indicators of various models, such as accuracy, recall, F1 score, etc.

[0066] In step S208, a decision is made whether to accept the new parameter settings. This decision can be made based on whether the improvement in model performance meets the predetermined goals and whether the new parameters can provide stable performance improvements without overfitting. If the new parameters fail to bring significant performance improvements or cause the model to overfit, the new parameters are rejected. On the contrary, if the new parameter settings can improve performance while maintaining model complexity, the new parameters are accepted. In some instances, an acceptance probability function can be used to decide whether to accept the new parameter settings. In some instances, the acceptance probability function can be used based on the current temperature T n The acceptance probability function is calculated by the difference between the performance of the model under the new parameter settings and the performance of the model under the current parameter settings. The acceptance probability function can be formula 3:

[0067]

[0068] Here, P is the probability of accepting a new parameter setting, ΔE is the difference between the performance of the model under the new parameter setting and the performance of the model under the current parameter setting, K is the Boltzmann constant, and T n is the current temperature. If the performance of the model under the new parameter setting is better, the new parameter setting can be directly accepted. If the performance of the model under the new parameter setting is poor, it can be decided whether to accept the new parameters based on the acceptance probability.

[0069] In step S209, if the new parameter settings are accepted, the current parameter settings are updated to the new parameter settings; otherwise, the current parameter settings are kept unchanged. In step S210, steps S204 to S209 are repeated until all set iterations are completed. In step S211, the predetermined number of iterations is completed to obtain the optimal random forest parameter settings, and these parameters are used to train the final model.

[0070] The simulated annealing algorithm disclosed in the present invention is not limited to logarithmic cooling scheduling, but can also be numerical cooling scheduling, linear cooling scheduling, power function cooling scheduling, logarithmic exponential scheduling, etc.

[0071] In some embodiments, the random forest model can be trained and optimized using AWS SageMaker.

[0072] In step S210, the optimized random forest model is deployed on a cloud platform (e.g., AWS) for real-time fault classification and diagnosis. In step S211, the cloud platform on which the optimized random forest model is deployed is used to provide real-time fault alarms and diagnostic reports for the power system of the electric vehicle.

[0073] Example 2

[0074] The method for remotely diagnosing power system faults of electric vehicles in exemplary embodiment 2 of the present disclosure is basically the same as the method in embodiment 1. In exemplary embodiment 2, the real-time data collected includes: DC / DC output voltage (DCHV): 310V; battery pack total voltage (PackVolt): 690V; motor current (MotorC): 85A. In the parameters of the random forest model, the number of decision trees is 120, the minimum number of split samples is 6, and the minimum number of leaf node samples is 2. In the logarithmic cooling scheduling parameters of formula 2, the initial temperature T i =150, logarithmic growth constant k=0.2, adjustment factor γ=0.3, power factor δ=0.1. After 15 rounds of iterations, the final optimization results of the above related parameters in the random forest model are: the number of decision trees is 180, the minimum number of split samples is 3, and the minimum number of leaf node samples is 2. Therefore, the fault classification accuracy of the optimized random forest model is about 98.1%.

[0075] Example 3

[0076] The method for remotely diagnosing power system faults of electric vehicles in exemplary embodiment 3 of the present disclosure is basically the same as the methods in embodiments 1 and 2. In exemplary embodiment 2, the real-time data collected includes: DC / DC output voltage (DCHV): 320V; battery pack total voltage (PackVolt): 680V; motor current (MotorC): 78A. In the parameters of the random forest model, the number of decision trees is 110, the minimum number of split samples is 5, and the minimum number of leaf node samples is 3. In the logarithmic cooling scheduling parameters of formula 2, the initial temperature T i =200, logarithmic growth constant k=0.05, adjustment factor γ=0.7, power factor δ=0.3. After 12 rounds of iterations, the final optimization results of the above related parameters in the random forest model are: the number of decision trees is 160, the minimum number of split samples is 4, and the minimum number of leaf node samples is 2. Therefore, the fault classification accuracy of the optimized random forest model is about 97.8%.

[0077] Table 1. Real-time data collected in Examples 1-3

[0078]

[0079] Table 2. Parameters of the random forest model and logarithmic cooling schedule parameters in Examples 1-3

[0080]

[0081] Table 3. Parameters of the optimized random forest model in Examples 1-3 and the fault classification accuracy of the model

[0082]

[0083]

[0084] In the random forest model, each parameter has an impact on the fault classification accuracy of the model. The data of Examples 1-3 show that increasing the number of decision trees improves the fault classification accuracy of the model, indicating that more decision trees can better capture data features and patterns. Reducing the minimum number of split samples can improve the fault classification accuracy of the model. A smaller minimum number of split samples allows the model to perform more detailed splits, thereby better learning the subtle features in the data. Adjusting the minimum number of leaf node samples has a relatively small effect on improving classification accuracy, but it still helps to balance the complexity and generalization ability of the model.

[0085] In the process of optimizing the hyperparameters of the random forest model using the simulated annealing algorithm, the parameters of the simulated annealing algorithm have an impact on the fault classification accuracy of the final model. i (such as 200) may make the optimization process more exploratory at the beginning, which can avoid falling into the local optimal solution, but may require more iterations to converge. Moderate initial temperature (such as 100 or 150) usually provides a balance between convergence speed and accuracy. A larger logarithmic growth constant k (such as 0.2) can speed up the temperature reduction rate, making the search process converge earlier, thereby improving the classification accuracy to a certain extent. A smaller constant (such as 0.05) provides a slower temperature reduction rate, which helps to refine the search, but the classification accuracy is relatively low. A larger adjustment factor γ (such as 0.7) will introduce more temperature changes in each round of iteration, making the search process more flexible. A smaller factor (such as 0.3) makes the temperature change more stable, which helps to maintain the stability of the search. A larger power factor δ (such as 0.3) will make the temperature decrease faster, which is suitable for a shorter optimization process. A smaller factor (such as 0.1 or 0.2) makes the temperature decrease slower, which helps to search for a long time and find a better solution.

[0086] The present disclosure also provides a system including one or more computer processors and a computer readable memory. The computer readable memory may include machine executable code, which, when executed by the one or more computer processors, implements the method for remotely diagnosing a power system fault of an electric vehicle of the present disclosure.

[0087] Although preferred embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that these embodiments are provided by way of example only. Other variations, changes and substitutions may be conceived by those skilled in the art without departing from the present invention.

Claims

1. A method for remotely diagnosing a power system fault of an electric vehicle, the method comprising: Collecting real-time data of the power system of the electric vehicle; Transmitting the real-time data to a cloud platform; Using a random forest model to sort the feature importance of the real-time data; Using simulated annealing algorithm to optimize the hyperparameters of the random forest model; Deploying the optimized random forest model on the cloud platform for real-time fault classification and diagnosis; The cloud platform is used to provide real-time fault alarms and diagnostic reports of the power system of the electric vehicle.

2. The method according to claim 1, wherein the random forest model performs the real-time fault classification and diagnosis based on a training set, and the importance of feature j in the training set is j Calculated by formula 1: in, ΔGini(t,j) represents the reduction of the Gini index caused by feature j in decision tree t, and T is the total number of decision trees.

3. The method of claim 2, wherein the simulated annealing algorithm optimizes the hyperparameters of the random forest model via a cooling mechanism that uses a logarithmic cooling schedule.

4. The method of claim 3, wherein the logarithmic cooling schedule is calculated by Equation 2: in, T n is the temperature after iteration, T i is the initial temperature, k is the logarithmic growth constant, γ is the adjustment factor, δ is the power factor, and iter is the current number of iterations.

5. The method of claim 4, wherein the real-time data includes data of a DC / DC converter, a motor, and a battery of the electric vehicle.

6. The method of claim 5, wherein the real-time data comprises voltage, current, temperature, pressure or any combination thereof.

7. The method of claim 4, wherein the initial temperature is 100 to 200, the γ adjustment factor is 0.3 to 0.7, and a plurality of iterations are performed until the hyperparameters converge to an optimal solution.

8. The method of claim 1, wherein the cloud platform is an AWS cloud platform.

9. The method of claim 8, wherein the AWS platform includes an AWS SageMaker Endpoint module.

10. The method of claim 8, wherein the AWS platform includes an AWS CloudWatch module.

Images