New energy automobile motor fault detection method

Abstract

The invention discloses a new energy automobile motor fault detection method, which comprises the steps of constructing an AlexNet network model which comprises six convolution stacking layers and three full connection layers which are connected in sequence; a plurality of motor stator current signals with known faults and a plurality of motor stator current signals without faults are collected, the collected motor stator current signals are converted into images respectively, corresponding labels are given to the images, and a sample set is obtained; a corresponding fault type label is given to the image corresponding to the faulty motor stator current signal, and a fault-free label is given to the image corresponding to the fault-free motor stator current signal; the AlexNet network model is trained, and a new energy automobile motor fault detection model is obtained; obtaining a motor stator current signal, converting the motor stator current into an image, and inputting the image into a new energy automobile motor fault detection model; the new energy automobile motor fault detection model outputs a detection result; wherein the detection result is no fault or fault type.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy vehicle fault detection, and in particular relates to a new energy vehicle motor fault detection method. Background Art

[0002] As an important power source in the drive system of new energy (electric) vehicles, the safety of the motor will affect the performance of the vehicle, driving experience, driving safety, and even cause life threats. Therefore, in order to improve the driving experience of electric vehicles, reduce the maintenance cost and driving risks of electric vehicles, it is imperative to develop fault diagnosis and monitoring technology for electric vehicle motors.

[0003] Permanent magnet synchronous motor is a type of motor widely used in new energy vehicles. It has many advantages such as high power density ratio, high torque, good dynamic performance, simple structure and easy maintenance.

[0004] At present, there are two methods for detecting permanent magnet synchronous motor faults: (1) Time-frequency domain analysis method based on MSCA. The frequency domain fault features proposed by this method have overlapping frequency components, making it difficult to distinguish the frequency domain features of various faults. In addition, closed-loop control will weaken the fault frequency domain features, resulting in many frequency domain features not being obvious under motor closed-loop control, limiting the scope of application of this technology.

[0005] (2) Diagnostic method based on signal processing. This method requires additional sensors to detect non-stator currents, such as air gap flux, motor vibration signals, etc. Under the condition of multiple vibrations such as electric vehicles, it is difficult to ensure the reliability of the diagnostic method. In addition, the installation position of the sensor can easily affect the diagnostic effect of the motor, and it is also easy to fall off due to shaking during vehicle driving. Summary of the invention

[0006] The purpose of the present invention is to provide a method for detecting motor faults of new energy vehicles, which can simultaneously detect different fault types and improve the accuracy of the detection results.

[0007] The technical solution provided by the present invention is: A method for detecting motor faults in a new energy vehicle, comprising: Construct the AlexNet network model, which includes 6 convolutional stacked layers and 3 fully connected layers connected in sequence; Among them, the first convolution stacked layer, the second convolution stacked layer, the fourth convolution stacked layer and the sixth convolution stacked layer respectively include a convolution layer, an LRN layer and a pooling layer connected in sequence; the third convolution stacked layer and the fifth convolution stacked layer respectively include a convolution layer and an LRN layer connected in sequence; Collecting multiple motor stator current signals with known faults and multiple motor stator current signals without faults, converting the collected motor stator current signals into images and assigning corresponding labels to obtain a sample set; dividing the sample set into a training set, a verification set and a test set; Among them, the image corresponding to the faulty motor stator current signal is assigned a corresponding fault type label, and the image corresponding to the fault-free motor stator current signal is assigned a fault-free label; The AlexNet network model is trained using the training set to obtain a new energy vehicle motor fault detection model; Acquire the motor stator current signal according to the sampling period, convert the motor stator current into an image, and input the image into the new energy vehicle motor fault detection model; the new energy vehicle motor fault detection model outputs the detection result; The detection result is no fault or a fault type.

[0008] Preferably, the motor stator current signal is converted into an image using an image conversion method based on an autocorrelation matrix.

[0009] Preferably, when constructing the sample set, the process further includes expanding the image converted from the motor stator current, and using the expanded image set as the sample set; The method for expanding the image is as follows: randomly selecting a portion of the image, making an image copy, and rotating the image copy; mixing the rotated image with the original image to obtain an expanded image set.

[0010] Preferably, when performing image expansion, the image rotation angle is 90 degrees to 270 degrees.

[0011] Preferably, the new energy vehicle motor fault detection method further includes using a genetic algorithm to determine the size of the convolution kernel of each convolution layer, including the following steps: Step 1: Randomly initialize the convolution kernel values ​​of the six convolution layers in the AlexNet network model to obtain multiple vectors consisting of the convolution kernel values ​​of the six convolution layers as the initial population; Among them, each vector corresponds to an AlexNet network model; Step 2: Use the training set to train each AlexNet network model, calculate the loss function of each AlexNet network model, and use the inverse of the loss function as the fitness of the corresponding vector; Step 3: Select the vectors with the largest fitness in the population and enter the next generation directly according to the set ratio; perform crossover and mutation on the remaining individuals and enter the next generation; Step 4: Repeat steps 2 and 3 until the maximum number of iterations is reached; select the convolution kernel value corresponding to the vector with the largest fitness as the convolution kernel value of the six convolution layers in the AlexNet network model.

[0012] Preferably, during initialization, the upper limit of the convolution kernel value is set to 19×19 and the lower limit is set to 1×1.

[0013] Preferably, the pooling layers in the first convolution stacked layer, the second convolution stacked layer, the fourth convolution stacked layer and the sixth convolution stacked layer all adopt maximum pooling layers, and the pooling kernel size is 3×3.

[0014] Preferably, the first two front-connected layers of the three fully-connected layers are respectively provided with Dropout, and the last fully-connected layer is provided with a SoftMax classifier, and the detection result is outputted through the SoftMax classifier.

[0015] The beneficial effects of the present invention are: The new energy vehicle motor fault detection method provided by the present invention can simultaneously detect different fault types and improve the accuracy of the detection results. DETAILED DESCRIPTION

[0016] The present invention is described in further detail below so that those skilled in the art can implement it according to the description.

[0017] 1. Build AlexNet network model The AlexNet network model includes 6 convolutional stacking layers and 3 fully connected layers connected in sequence; the ReLU function is used as the activation function. Among them, the first convolutional stacking layer, the second convolutional stacking layer, the fourth convolutional stacking layer and the sixth convolutional stacking layer respectively include sequentially connected convolutional layers, LRN layers and pooling layers; the third convolutional stacking layer and the fifth convolutional stacking layer respectively include sequentially connected convolutional layers and LRN layers. The pooling layers in the first convolutional stacking module, the second convolutional stacking module, the fourth convolutional stacking module and the sixth convolutional stacking module all use the maximum pooling layer, and the pooling kernel size is 3×3. Dropout is respectively set in the first two fully connected layers of the three fully connected layers, and the last fully connected layer is set with a SoftMax classifier, and the detection results are output by the SoftMax classifier. Among them, the number of neurons in the fully connected layer is 2048, and the final output SoftMax is 3, which corresponds to 3 detection results, and the three detection results are two different fault types and no fault. In this embodiment, the three detection results are demagnetization fault, eccentricity fault and no fault.

[0018] As a preferred embodiment, a genetic algorithm is used to determine the size of the convolution kernel of each convolution layer. The specific process is as follows.

[0019] S1. Randomly initialize the convolution kernel values ​​of the six convolution layers in the AlexNet network model to obtain multiple vectors consisting of the convolution kernel values ​​of the six convolution layers to form an initial population, where each vector corresponds to an AlexNet network model.

[0020] To avoid excessive redundant data, the upper limit of the convolution kernel value is set to 19×19 and the lower limit is set to 1×1 during initialization. The convolution kernel value of the first layer (input layer) is set to the largest, and the convolution kernel value of the last layer is set to the smallest; and from the input to the output direction, the convolution kernel value of the next layer is not greater than the convolution kernel value of the previous layer.

[0021] S2. Use the training set to train each AlexNet network model. After the training, calculate the loss function of each AlexNet network model, and use the inverse of the loss function as the fitness of the corresponding vector.

[0022] S3, select the vectors with large fitness in the population and enter the next generation directly according to the set ratio (quantity); the remaining individuals enter the next generation after crossover and mutation; S4. Repeat S2-S3 until the maximum number of iterations is reached; select the convolution kernel value corresponding to the vector with the largest fitness as the convolution kernel value of the 6 convolution layers in the AlexNet network model.

[0023] In one embodiment, cross entropy is used as the loss function. Cross entropy is a loss function widely used in classification problems and is defined as follows: ; in, and Represent the target and output probability distributions respectively, represents the vector dimension, Table vector value in the dimension.

[0024] The test set was used to test the AlexNet network model that used the genetic algorithm to optimize the convolution kernel value. The test results showed that the optimized AlexNet network model converged faster during training and had a relatively stable detection accuracy compared to the AlexNet network model set according to empirical values.

[0025] 2. Model Training More samples are used to train the optimized AlexNet network model so that the model can achieve more accurate detection results.

[0026] The sample set is obtained by expanding the collected samples. Multiple motor stator current signals with known faults and multiple motor stator current signals without faults are collected, and the collected motor stator current signals are respectively converted into images and assigned corresponding labels to obtain a sample set; the sample set is divided into a training set, a validation set, and a test set. Among them, the image corresponding to the motor stator current signal with faults is assigned a corresponding fault type label, and the image corresponding to the motor stator current signal without faults is assigned a fault-free label.

[0027] In this embodiment, the collected samples include: motor signals with different degrees of demagnetization faults, with a demagnetization degree of 10% to 50%; and motor signals with different degrees of eccentricity faults, with an eccentricity degree of 10% to 20%; and these signals are collected under different operating states of the motor, including various operating states within the range of 1000 rpm to 4000 rpm.

[0028] As a preferred method, the motor stator current signal is converted into an image using an image conversion method based on an autocorrelation matrix. As long as the data sample size is large enough, the data will obey the Gaussian distribution. Therefore, the mapping function of the data scattered into the gray value range is set to a one-dimensional Gaussian distribution function, and the conversion formula is as follows: ; In the formula, Indicates the autocorrelation matrix Row, No. Elements of a column; is the mean of the original data, is the variance of the original data, is the standard deviation of the original data. Compared with the image conversion method based on permutation, the image conversion method based on the signal autocorrelation matrix requires less original data and is more timely; compared with other image conversion methods, the image conversion method based on the autocorrelation matrix makes a complete mapping relationship between the original data and the image pixel value, and the logic is more rigorous.

[0029] The training of the network requires a large number of samples. Through the learning of a large number of samples, deep and specific features can be obtained, and the accuracy of detection can be improved. In order to obtain a large number of sample sets and prevent overfitting during network training, the image samples of the motor stator current conversion collected are expanded to improve the robustness of the convolutional neural network.

[0030] The method for expanding the image is as follows: randomly selecting a portion of the image, making a copy of the image, rotating the copy of the image, and mixing the rotated image with the original image to obtain an expanded image set. When expanding the image, the image rotation angle is 90 degrees to 270 degrees.

[0031] In one embodiment, the expanded sample set contains 30,000 images, including 10,000 transformed images of fault-free signals, 10,000 transformed images of demagnetization fault signals, and 10,000 transformed images of eccentricity fault signals.

[0032] The expanded sample set is divided into training set, validation set and test set in a ratio of 8:1:1. The batch size used in the training process is 32, the optimizer is Adam, the learning rate is set to 0.002, and the dropout regularization parameter is set to 50%.

[0033] When the training iteration number of the improved AlexNet network model is 100, the accuracy and loss function values ​​basically reach a stable level. Therefore, the iteration coefficient is set to 100 as the upper limit of the iteration number.

[0034] The trained model was tested by classification using the validation set. The detection accuracy of the model is the ratio between the model's judgment result on the sample fault category and the sample's true fault category. It has been verified that the model's recognition accuracy for each type of fault is above 98.7%, and the detection and recognition effect is relatively ideal.

[0035] The specific test results are shown in Table 1: Table 1 Model detection accuracy

[0036] Afterwards, the model was tested using the test set, and both the precision and recall rates were able to reach over 98%.

[0037] 3. Use the trained model (new energy vehicle motor fault detection model) to detect motor faults During the driving process of the vehicle, the motor stator current signal is obtained according to the sampling period, and the motor stator current is converted into an image and input into the new energy vehicle motor fault detection model; the new energy vehicle motor fault detection model outputs a detection result; wherein the detection result is no fault or a fault type.

[0038] When the motor stator current is converted into an image for detection, the collected motor stator current is converted into an image as the input of the convolution layer in the first convolution stack layer. The convolution layer and the pooling layer respectively extract features and reduce dimensions of the sample image, and integrate the information into the next convolution stack layer. Finally, the feature information is summarized in the fully connected layer, and the sample image is classified through the SoftMax function in the fully connected layer to obtain the detection result.

[0039] Test example The new energy vehicle motor fault detection model provided by the present invention is compared with the AlexNet6 network model that has not been optimized by the genetic algorithm in terms of detection accuracy and recall rate. 1000 groups of samples are randomly selected from the sample set as the input of each network model. It is verified that the accuracy and recall rate of the unoptimized AlexNet6 network model are significantly lower than those of the new energy vehicle motor fault detection model provided by the present invention.

[0040] The detection accuracy and recall results of the two models are shown in Table 2.

[0041] Table 2 Detection accuracy and recall of the two models

[0042] The present invention improves the AlexNet network model by adopting a genetic algorithm, which can reduce the training loss of the network and thus improve the accuracy of fault detection.

[0043] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the embodiments shown and described herein.

Claims

1. A method for detecting motor faults in new energy vehicles, characterized in that: The steps include: Construct the AlexNet network model, which includes 6 convolutional stacked layers and 3 fully connected layers connected in sequence; Among them, the first convolution stacking layer, the second convolution stacking layer, the fourth convolution stacking layer and the sixth convolution stacking layer respectively include a convolution layer, an LRN layer and a pooling layer connected in sequence; the third convolution stacking layer and the fifth convolution stacking layer respectively include a convolution layer and an LRN layer connected in sequence; Collecting multiple motor stator current signals with known faults and multiple motor stator current signals without faults, converting the collected motor stator current signals into images and assigning corresponding labels to obtain a sample set; dividing the sample set into a training set, a verification set and a test set; Among them, the image corresponding to the faulty motor stator current signal is assigned a corresponding fault type label, and the image corresponding to the fault-free motor stator current signal is assigned a fault-free label; The AlexNet network model is trained using the training set to obtain a new energy vehicle motor fault detection model; Acquire the motor stator current signal according to the sampling period, convert the motor stator current into an image, and input the image into the new energy vehicle motor fault detection model; the new energy vehicle motor fault detection model outputs the detection result; Wherein, the detection result is no fault or a fault type; It also includes using a genetic algorithm to determine the size of the convolution kernel of each convolution layer, including the following steps: Step 1: Randomly initialize the convolution kernel values ​​of the six convolution layers in the AlexNet network model to obtain multiple vectors consisting of the convolution kernel values ​​of the six convolution layers as the initial population; Among them, each vector corresponds to an AlexNet network model; Step 2: Use the training set to train each AlexNet network model, calculate the loss function of each AlexNet network model, and use the inverse of the loss function as the fitness of the corresponding vector; Step 3: Select the vectors with the largest fitness in the population and enter the next generation directly according to the set ratio; perform crossover and mutation on the remaining individuals and enter the next generation; Step 4: Repeat steps 2 and 3 until the maximum number of iterations is reached; select the convolution kernel value corresponding to the vector with the largest fitness as the convolution kernel value of the six convolution layers in the AlexNet network model.

2. The new energy vehicle motor fault detection method according to claim 1 is characterized in that: The motor stator current signal is converted into an image using an image conversion method based on the autocorrelation matrix.

3. The new energy vehicle motor fault detection method according to claim 2 is characterized in that: When constructing the sample set, the image converted from the motor stator current is expanded, and the expanded image set is used as the sample set; The method for expanding the image is as follows: randomly selecting a portion of the image, making an image copy, and rotating the image copy; mixing the rotated image with the original image to obtain an expanded image set.

4. The new energy vehicle motor fault detection method according to claim 3 is characterized in that: When expanding the image, the image rotation angle is 90 degrees to 270 degrees.

5. The new energy vehicle motor fault detection method according to claim 1 is characterized in that: During initialization, the upper limit of the convolution kernel value is set to 19×19 and the lower limit is set to 1×1.

6. The new energy vehicle motor fault detection method according to claim 5 is characterized in that: The pooling layers in the first convolution stacking layer, the second convolution stacking layer, the fourth convolution stacking layer and the sixth convolution stacking layer all adopt maximum pooling layers, and the pooling kernel size is 3×3.

7. The method for detecting motor faults in new energy vehicles according to claim 6, characterized in that: The first two front connection layers of the three fully connected layers are respectively provided with Dropout, and the last fully connected layer is provided with a SoftMax classifier, and the detection result is outputted through the SoftMax classifier.

Images