Embedded motor fault detection method based on LDS-CNN

Abstract

The invention relates to the technical field of fault detection, in particular to an embedded motor fault detection method based on an LDS-CNN, and the method comprises the steps: constructing a sensorless vector control estimation motor operation model; three-phase current signals corresponding to the rotating speeds of the motor in different states are collected, and signal preprocessing is carried out; according to the method, an LDS-CNN network is constructed, and the LDS-CNN network comprises a first DS convolution layer, a first maximum pooling layer, a second DS convolution layer, a second maximum pooling layer, Dropout, a GlobalAvg Pool, a Pooling1D and a Softmax classifier. The problems that an existing model cannot give consideration to the lightweight degree, the real-time performance and the anti-interference performance at the same time, and application requirements of a microcontroller unit are difficult to meet are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to an embedded motor fault detection method based on LDS-CNN. Background Art

[0002] Motor systems are widely used in industry, transportation, energy and other fields. At present, motor fault diagnosis methods can be roughly divided into two categories: deep learning methods can extract features from raw signals, reduce dependence on artificial feature engineering, and improve the accuracy and robustness of fault diagnosis.

[0003] At present, the application of deep learning in fault diagnosis is mostly concentrated on PCs and embedded microprocessor units, while the deployment research on microcontroller units (MCUs) is still in the early stages. Due to the advantages of MCUs such as small size, high integration, and low power consumption, deploying deep learning models on MCUs helps to achieve close integration of monitoring and diagnosis processes with equipment. However, existing deep learning models, such as traditional CNNs, often face limitations in computing resources and storage space when applied on embedded systems.

[0004] Lightweight CNN variant models are deployed into embedded systems to achieve real-time diagnosis of motor faults, such as deploying the SEDS-CNN model into an embedded system consisting of a JetsonNano and a signal acquisition module, and deploying the ECNN model into an STM32 development board and a signal acquisition processing circuit for online detection of motor faults. However, these existing models still have much room for improvement in key performance indicators such as lightweight, real-time, and anti-interference, and are difficult to meet the application requirements of microcontroller units (MCUs) with more limited computing and storage resources. Summary of the invention

[0005] In view of the shortcomings of the existing methods, the present invention solves the problem that the existing models cannot take into account lightweight, real-time and anti-interference performance at the same time and are difficult to meet the application requirements of microcontroller units.

[0006] The technical solution adopted by the present invention is: an embedded motor fault detection method based on LDS-CNN comprises the following steps:

[0007] Step 1: Construct a sensorless vector control estimation motor operation model;

[0008] As a preferred embodiment of the present invention, the formula for estimating the motor operation model by sensorless vector control is:

[0009]

[0010] Where V a , V b , V cis the three-phase stator voltage; I a , I b , I c is the three-phase stator current; R is the stator winding resistance; L is the induced electromotive force of the rotor in the three-phase winding; E a , E b , E c It is three-way reverse electromotive force.

[0011] Step 2: Collect the three-phase current signals corresponding to the motor speed in different states, and perform current signal preprocessing;

[0012] As a preferred embodiment of the present invention, different states include: normal load and overload.

[0013] As a preferred embodiment of the present invention, a microcontroller is used to collect three-phase current signals.

[0014] As a preferred implementation of the present invention, the preprocessing includes: overlapping sampling, signal normalization and Fourier transform processing.

[0015] Step 3: Construct the LDS-CNN network. The LDS-CNN network includes: the first DS convolution layer, the first maximum pooling layer, the second DS convolution layer, the second maximum pooling layer, Dropout, GlobalAvg Pool and Softmax classifier;

[0016] As a preferred implementation of the present invention, the convolution kernel size of the first DS convolution layer is 3x3, the convolution kernel size of the second DS convolution layer is 5x5, and the step size is 1.

[0017] As a preferred embodiment of the present invention, the LDS-CNN network is evaluated using accuracy, model size and parameter quantity.

[0018] As a preferred implementation of the present invention, the LDS-CNN network uses the Adam optimizer to update weights and sets the learning rate, batch and maximum number of iterations.

[0019] As a preferred embodiment of the present invention, an embedded motor fault detection system based on LDS-CNN includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement an embedded motor fault detection method based on LDS-CNN.

[0020] As a preferred embodiment of the present invention, a computer readable medium stores a computer program code, and the computer program code implements an embedded motor fault detection method based on LDS-CNN when executed by a processor.

[0021] Beneficial effects of the present invention:

[0022] 1. By constructing a lightweight deep separable convolutional neural network (LDS-CNN) and optimizing the computational structure of the convolution layer, the network model can process input signals with less computational overhead. Specifically, LDS-CNN reduces the amount of computation and improves the processing speed by combining deep convolution and point-by-point convolution, while maintaining a high fault classification accuracy.

[0023] 2. Efficiency: The use of a deep separable convolutional network greatly reduces the computational workload and storage requirements of the model, making it suitable for running on embedded systems with limited computing resources and capable of completing fault diagnosis tasks in real time;

[0024] 3. High precision: Automatic feature extraction through deep learning models can more accurately identify and classify motor faults, reducing human intervention and misdiagnosis common in traditional methods;

[0025] 4. Strong real-time performance: The optimized lightweight network can run in real time on resource-constrained embedded devices, has the ability of rapid response and efficient diagnosis, and adapts to the real-time requirements in industrial applications;

[0026] 5. The pre-trained LDS-CNN model is deployed on an embedded system based on the Renesas RA6T1 chip, making full use of the resources of a single control board to achieve real-time fault detection while driving the motor without adding additional sensors, effectively reducing system cost and complexity, and providing an innovative and efficient solution for the application of motor fault diagnosis in resource-constrained environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the LDS-CNN network structure diagram of the present invention;

[0028] Figure 2 It is based on the Renesas RA6T1 embedded platform diagram;

[0029] Figure 3 is a block diagram of a motor fault detection system of the present invention;

[0030] Figure 4 It is a detection accuracy diagram of the model under different signal-to-noise ratios of the present invention;

[0031] Figure 5 It is a time-frequency domain comparison diagram of normal and abnormal three-phase current;

[0032] Figure 6 is the accuracy of different models at different motor speeds. DETAILED DESCRIPTION

[0033] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0034] Traditional CNN has certain applications in motor fault diagnosis, but it faces many challenges. Taking the common CNN architecture for motor fault diagnosis as an example, it usually contains multiple convolutional layers, pooling layers and fully connected layers. In terms of computing resource usage, as the number of convolutional layers increases, the amount of computation increases exponentially. For example, a model with 4 traditional convolutional layers (convolution kernel size is 3x3) may increase its computational complexity (measured in FLOPs) several times when processing motor fault data compared to a model with only 2 convolutional layers. This is because each convolutional layer needs to perform a large number of convolution operations on the input data, and as the number of layers increases, although the size of the intermediate feature map is reduced under the action of the pooling layer, the overall computational burden is still heavy.

[0035] In terms of memory resource usage, the fully connected layer of traditional CNN accounts for the majority of memory consumption; neurons in the fully connected layer need to be connected to all feature maps in the previous layer, resulting in a large number of parameters; for example, in a CNN model for motor fault classification, if the last fully connected layer has 100 neurons and the feature map size of the previous layer is 10x10x50 (hypothetically), then the number of parameters in this layer will reach 500,000, which is a huge storage pressure for resource-constrained embedded systems.

[0036] In addition, for other related models, such as some variants based on recurrent neural networks (RNNs) when applied to motor fault diagnosis, although they have certain advantages in processing time series data, they also have problems in model size and computational efficiency; RNN and its variants (such as LSTM, GRU) need to save the hidden state of each time step during the calculation process due to their internal loop structure, which adds additional computing and storage overhead; in motor fault diagnosis tasks, when processing longer current signal time series, this overhead will be more obvious, which can easily lead to slower model training and inference speeds, and make deployment on embedded devices difficult.

[0037] like Figure 1 As shown, an embedded motor fault detection method based on LDS-CNN includes the following steps:

[0038] Step 1: Construct a sensorless vector control estimation motor operation model;

[0039]

[0040] Where V a , V b , Vc is the three-phase stator voltage; I a , I b , I c is the three-phase stator current; R is the stator winding resistance; L is the induced electromotive force of the rotor in the three-phase winding; E a , E b , E c It is three-way reverse electromotive force.

[0041] This model can be used to efficiently estimate the operating status of the motor, achieving precise control and fault monitoring.

[0042] Step 2: Use the microcontroller to collect the three-phase current signals corresponding to the motor speed in different states.

[0043] And perform signal preprocessing;

[0044] The states include: normal load and overload;

[0045] like Figure 2 , using an embedded platform based on Renesas RA6T1, including inverter board, MCU

[0046] (R7FA6T1AD3CFP) and CPU card board; the platform is based on 32-bit ARM Cortex-M4 architecture, powered by 24V DC power supply, equipped with 12-bit high-precision analog-to-digital converter (ADC), operating frequency of 120MHz, integrated with 512kB Flash memory and 64kB SRAM, meeting the performance requirements of real-time motor fault detection;

[0047] The microcontroller MCU adopts the built-in channel sampling and holding function in the Renesas RA6T1 embedded device. The microcontroller collects the three-phase signal of the motor while using sensorless vector control to drive the motor. The three-phase signal is transmitted to the processing unit of the microcontroller in real time for motor control feedback and motor fault diagnosis and analysis.

[0048] like Figure 3 As shown in the figure, in order to realize the accurate acquisition and model training of BLDC three-phase current, an embedded software system based on Renesas RA6T1 chip is designed. The system integrates motor control, signal acquisition and fault detection functions; the system realizes real-time fault detection of sensorless vector motor control based on RA6T1, without relying on additional hardware, and makes full use of the computing and processing capabilities of RA6T1 and its own channel sampling and holding function to perform real-time motor fault detection while driving the motor through sensorless vector control.

[0049] The system uses sensorless vector control technology to drive BLDC. In terms of signal acquisition, the system integrates a high-precision ADC to collect three-phase current data in real time. After MCU processing, it is used for motor control feedback and fault detection.

[0050] After the current signal is preprocessed, the pre-trained model is used for fault diagnosis. The detection results are displayed by the indicator light and transmitted to the PC in real time through the Universal Asynchronous Receiver-Transmitter (UART) bus. The BLDC parameters are shown in Table 1.

[0051] Table 1 Motor parameters

[0052]

[0053] Preprocessing includes: overlapping sampling, signal normalization and Fourier transform processing of the collected three-phase current data; using overlapping sampling method for 2kHz three-phase current signal to improve data utilization, capture instantaneous changes and reduce edge effects, the sampling window is as large as 512 points, and the overlap between adjacent windows is 64 points; collecting normal and abnormal three-phase currents of the motor running for 20 seconds;

[0054] Signal normalization ensures that the data is in a uniform scale range to avoid the influence of different signal dimensions on the diagnostic results; Fourier transform is used for time-frequency domain conversion to better extract the potential features in the three-phase current signal of the motor.

[0055] Step 3: Construct an LDS-CNN network, with the preprocessed current signal as input and whether it is abnormal as output; the LDS-CNN network includes: the first DS convolution layer, the first maximum pooling layer, the second DS convolution layer, the second maximum pooling layer, GlobalAvgPool and Softmax classifier;

[0056] Table 2 Model parameters

[0057]

[0058] The convolution kernel size of the first DS convolution layer is 3x3 and the stride is 1; the convolution kernel size of the second DS convolution layer is 5x5 and the stride is 1.

[0059] The Adam optimizer is used to update the weights in training, the learning rate is set to 0.0015, the batch is 3, the maximum number of iterations is set to 300, and the training set and test set are divided into 8:2.

[0060] Since the computing resources of MCU RA6T1 are only 512kb flash and 64RAM, and some other overheads are also included, the computing resource overhead for the network model is smaller; the present invention greatly reduces the computational complexity and memory requirements by constructing two layers of depth-separable convolutional layers (DS), two layers of maximum pooling and global average pooling operations; the network structure is more suitable for the operating environment of embedded devices and improves the real-time performance of the motor fault diagnosis system.

[0061] The characteristics of standard CNN are that the number of channels of each convolution kernel is equal to the number of channels of the input volume, the number of convolution kernels is consistent with the number of channels of the output volume, and the convolution operation is performed on the input channels one by one and then accumulated to complete the output; its parameter quantity and computational complexity are mainly affected by the number of input channels C. Input , convolution kernel size K H ×K W , Output channel number C Output and the output feature map size H Output ×W Output The influence of four groups of parameters; after ignoring the addition operation, the multiplication operation G and parameter quantity P are shown in formulas (2) and (3) respectively:

[0062] G= C Input ×K H ×K W ×C Output ×H Output ×W Output (2)

[0063] P= C Input ×K H ×K W ×C Output (3)

[0064] The DS convolution layer decomposes the standard convolution operation into two steps: depth convolution and point-by-point convolution. It performs convolution operations on each input channel independently and only extracts local features within the channel. This process effectively reduces the number of parameters, but ignores the correlation between different channels. To make up for this deficiency, the depth convolution is followed by point-by-point convolution (also called 1x1 convolution). Its purpose is to enhance the correlation between channels through cross-channel combination and transformation, integrate the feature information between channels, and restore the network's complete expression ability of signal features. Through this decomposition strategy, the DS convolution layer significantly reduces the computational complexity and parameter scale while maintaining the model's expression ability. Its multiplication operation G DS and parameter P DS As shown in formulas (4) and (5) respectively:

[0068] G DS =G depth +G point = C Input×K H ×K W ×H out ×W out +C Input ×C Output ×H Output ×W Output (4)

[0069] P DS =P depth +P point = C Input ×K H ×K W +C Input ×C Output (5)

[0070] The ratio of the number of parameters and the amount of computation of the DS convolution layer to the standard convolution layer is shown in formula (6) and formula (7):

[0071]

[0072] Taking the 3×3 convolution kernel as an example, the number of parameters and computation of the DS convolution layer is only about one-ninth of that of the standard CNN; DS-CNN effectively reduces model parameters and computation while still maintaining excellent feature extraction capabilities. This enables it to deploy high-precision motor fault detection models to resource-constrained embedded devices.

[0073] The main function of the maximum pooling layer is to divide the input feature map into several rectangular areas and select the maximum value of each sub-area as the output, thereby reducing the dimension sampling and parameter reduction of the feature map output by the convolution layer and reducing the calculation amount of the model. Its definition is shown in formula (8):

[0074]

[0075] in, The kth feature map of the input feature map in region R ij The maximum pooling output value of For region R ij The element at (p,q) in .

[0076] The Dropout layer randomly sets hidden neurons to 0 in proportion, which is equivalent to introducing noise in the convolution area to enhance the model's anti-interference ability. This layer is used to randomly hide some nodes during the training process so that they do not participate in calculations and updates in certain iterations, thereby preventing the model from overfitting.

[0077] After the feature map passes through two DS layers, it enters the global average pooling layer for feature compression; this layer performs global average pooling on each feature map to output a more concise feature representation, and to a certain extent improves the real-time and robustness of the model. At the same time, it greatly reduces the network parameters brought by model training, reduces the demand for model memory resources, and improves the lightweight degree of the model. The definition is shown in formula (9):

[0078]

[0079] Among them, y k is the global average pooling output value of the k-th feature map; R is the number of elements in the k-th feature map; is the element at position (p,q) in region R; compared with the fully connected layer, the global average pooling layer significantly reduces the training parameters, avoids overfitting and improves the generalization ability of the model.

[0080] Softmax normalizes the network output value to the (0,1) interval for classification. Its definition is shown in formula (10):

[0081]

[0082] Among them, Q(x j ) represents the output of the model, Num class Represents the number of categories, x j Represents the j-th output of the previous layer.

[0083] The LDS-CNN model is trained using a motor fault dataset. The network parameters are adjusted through an optimization algorithm to further improve the diagnostic accuracy and speed of the model. During the training process, the model will continuously adjust the weights to better adapt to the characteristics of the motor fault mode.

[0084] The trained LDS-CNN model can perform real-time fault diagnosis on the motor's operating data, identify motor faults and classify and output them. By extracting features from the input signal, the model can quickly determine whether the motor is faulty and output the corresponding fault type.

[0085] The trained LDS-CNN model is deployed to embedded devices (such as microcontrollers, embedded computing platforms, etc.) to achieve real-time fault detection and diagnosis. Through the deployed model, the system can run efficiently in the embedded environment and provide timely feedback on the fault status of the motor, providing a basis for subsequent maintenance or adjustment.

[0086] Experimental process:

[0087] The three-phase current signal generated by the Renesas RA6T1 embedded motor system is collected in real time at a sampling frequency of 2kHz, which is used as the experimental data set. This data set contains 8 types of sample data of the motor at different speeds and different load conditions; Figure 5 The left and right sides show the time-frequency domain comparison of normal and abnormal three-phase current signals respectively; in the normal state, the amplitude of the motor three-phase current is small, the waveform and spectrum are relatively simple, and the interference and harmonic content are low; in the abnormal state, the amplitude of the loaded motor current increases, and higher harmonics and random noise components are more likely to appear in the spectrum.

[0088] As shown in Table 3, when evaluating the model performance, the present invention uses the three-phase current data of the brushless DC motor to perform real-time motor fault classification; in order to evaluate the performance of LDS-CNN in a resource-constrained embedded environment, the model of the present invention is compared with several classic models (SEDS-CNN, ECNN, TICNN, CNN-LSTM); Table 3 summarizes the network structure parameters of each model. The results of multiple tests show that LDS-CNN performs well in detection accuracy, model size and parameter quantity.

[0089] Table 3 Comparison of test results of different models

[0090]

[0091] In terms of diagnostic accuracy, all four models showed excellent recognition performance, demonstrating the powerful feature learning ability of CNN; although ECNN and SEDS-CNN performed well in accuracy, LDS-CNN had significant advantages in model lightweighting, with a model size of only 31KB and significantly reduced parameters; the model size of LDS-CNN was reduced by approximately 7.5 times, 10.1 times, 14.2 times and 15.1 times compared to SEDS-CNN, ECNN, CNN-LSTM and TICNN, respectively; the number of parameters was reduced by approximately 62 times, 413 times, 149 times and 390 times, demonstrating its application potential in edge computing scenarios.

[0092] The RA6T1 embedded system not only needs to perform motor fault detection, but also must use sensorless vector control to drive the motor; therefore, a portion of computing and memory resources must be reserved for motor drive, while a portion of resources must be reserved to handle temporary emergencies; therefore, attention must be paid to lightweight design to optimize the use of computing and memory.

[0093] The performance evaluation of the model of the present invention with different numbers of convolutional layers is shown in Table 4, and its FLOPs value is calculated as follows: the FLOPs of the model with one-layer convolution structure is about 7634, which mainly comes from its small amount of convolution operations and simple connection operations; the FLOPs of the model with two-layer convolution structure is about 6506, which is lower than that of the one-layer model, but still at a low level; as the number of convolutional layers increases to three, the FLOPs increases to 11519, and reaches 12347 at four layers.

[0094] From the perspective of hardware performance, the RA6T1 embedded system of the present invention is based on a 32-bit ARM Cortex-M4 architecture with an operating frequency of 120MHz. Its computing power is limited when processing large-scale computing tasks. At the same time, the storage bandwidth of the system also restricts the loading and storage speed of model parameters. Under such hardware conditions, although the computational complexity of a one-layer convolution structure is low, the performance is poor, with an accuracy of only 80.95%, which cannot meet the actual fault detection requirements. Although the accuracy of a four-layer convolution structure reaches 100%, its high FLOPs value and memory usage (RAM usage is 20.82KB, ROM usage is 0.75KB) easily cause system delays or even freezes during operation, and there is a risk of overfitting, which affects the generalization ability of the model. The model with a two-layer convolutional structure achieves a good balance between performance and resource utilization; its FLOPs value is relatively moderate, and can ensure a fast reasoning speed within the computing power of the embedded system; at the same time, the memory usage (RAM usage is 9.17KB, ROM usage is 0.75KB) is also within an acceptable range, and will not cause excessive resource competition for other functions of the system (such as motor drive control); in terms of accuracy, it reaches 99.75%, the recall rate is 99.42%, and the F1 value is 99.37%, which can accurately identify motor fault states and effectively achieve the optimal resource utilization efficiency while ensuring performance. Therefore, the present invention finally selects a two-layer convolutional structure as the model architecture of the present invention.

[0095] Table 4 Model performance under different numbers of convolutional layers

[0096]

[0097] In order to evaluate the performance of the model, the four algorithms were tested at different speeds in an embedded system based on Renesas RA6T1. The results are shown in the figure. Figure 6 As shown in the figure, LDS-CNN improves the generalization ability of the model through global average pooling, making its detection accuracy at four speeds better than other algorithms, showing good generalization ability.

[0098] Multiple ablation experiments show that, as shown in Table 5, LDS-CNN outperforms other algorithms in terms of accuracy and lightweight, and is suitable for deployment in resource-limited embedded systems.

[0099] Table 5 Comparative experiments of different algorithms

[0100]

[0101] Standard evaluation indicators are used to evaluate the performance of the LDS-CNN model in different application environments, mainly including the evaluation of indicators such as accuracy, robustness, model size and network parameters; compared with several classic models such as SEDS-CNN, ECNN, CNN-LSTM and TICNN, the performance of the model is comprehensively evaluated; Figure 4 , motor fault diagnosis is performed in a noise environment simulating different signal-to-noise ratios, and the robustness of the LDS-CNN network under noise interference is evaluated; by adding Gaussian noise with different signal-to-noise ratios to the signal, the changes in the robustness of several different models during fault diagnosis are observed. The LDS-CNN network shows good robustness in noise environments with different signal-to-noise ratios.

[0102] The LDS-CNN network is run in a real-time embedded environment to evaluate the lightweight degree of the system. By testing the model size and parameter quantity of the model during the fault diagnosis of the motor during operation, it is proved that the LDS-CNN can maintain a high accuracy rate with a small parameter quantity and model size. The experimental results are shown in Table 6.

[0103] Table 6 Comparison between the proposed model and lightweight CNN

[0104]

[0105] Through the description of the above implementation methods, we can fully understand the fault diagnosis method based on LDS-CNN and its implementation, performance evaluation and effect comparison in practical applications; the application of this method in motor fault diagnosis has verified its high efficiency and robustness, and it has strong practical application value.

[0106] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. An embedded motor fault detection method based on LDS-CNN, characterized in that: The following steps are involved: Step 1: Construct a sensorless vector control estimation motor operation model; Step 2: Collect the three-phase current signals corresponding to the motor speed in different states, and perform current signal preprocessing; Step 3: Construct the LDS-CNN network, including: the first DS convolution layer, the first maximum pooling layer, the second DS convolution layer, the second maximum pooling layer, Dropout, GlobalAvg Pool and Softmax classifier.

2. The embedded motor fault detection method based on LDS-CNN according to claim 1 is characterized in that: The convolution kernel size of the first DS convolution layer is 3x3, and the convolution kernel size of the second DS convolution layer is 5x5, and the stride is 1.

3. The embedded motor fault detection method based on LDS-CNN according to claim 1 is characterized in that: The formula for estimating the motor operation model by sensorless vector control is: Where V a , V b , V c is the three-phase stator voltage; I a , I b , I c is the three-phase stator current; R is the stator winding resistance; L is the induced electromotive force of the rotor in the three-phase winding; E a , E b , E c It is three-way reverse electromotive force.

4. The embedded motor fault detection method based on LDS-CNN according to claim 1 is characterized in that: The different states include: normal load and overload.

5. The embedded motor fault detection method based on LDS-CNN according to claim 1 is characterized in that: The three-phase current signals are collected using a microcontroller.

6. The embedded motor fault detection method based on LDS-CNN according to claim 1 is characterized in that: Preprocessing includes: Overlap sampling, signal normalization and Fourier transform processing.

7. The embedded motor fault detection method based on LDS-CNN according to claim 1 is characterized in that: The LDS-CNN network is evaluated using accuracy, model size, and number of parameters.

8. The embedded motor fault detection method based on LDS-CNN according to claim 1 is characterized in that: The LDS-CNN network uses the Adam optimizer to update weights and set the learning rate, batch, and maximum number of iterations.

9. The embedded motor fault detection system based on LDS-CNN is characterized by: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the LDS-CNN-based embedded motor fault detection method as described in any one of claims 1 to 8.

10. A computer readable medium storing computer program code, characterized in that: When the computer program code is executed by a processor, the computer program code implements the embedded motor fault detection method based on LDS-CNN as described in any one of claims 1 to 8.

Images