An anomaly detection method, device, and storage medium for elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion

Abstract

This invention discloses a method, device, and storage medium for anomaly detection of permanent magnet synchronous motors (PMSMs) in elevators based on DWT-CNN multi-current fusion. It improves upon existing CNN network algorithms for extracting current features from PMSMs by combining a DWT layer to extract high- and low-frequency components of the current signal and a CNN layer to enhance noise resistance. To capture key information in the sequence, a hybrid attention mechanism consisting of a channel frequency attention layer (FAM) and a time attention layer (EAM) is introduced into the network to obtain fused feature signals rich in time-domain frequencies. A weight-updated LSTM anomaly detection classifier is designed to map the relationship between the fused features and the output, and to identify the correlation features before and after the motor anomaly. This invention improves the effect of fusing multiple current signals, effectively reduces noise and interference from three-phase electrical asymmetry, and comprehensively enhances diagnostic accuracy.

Description

Technical Field

[0001] This invention relates to the field of elevator permanent magnet synchronous motor technology, and in particular to an abnormality detection method, device and storage medium for elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion. Background Technology

[0002] Permanent magnet synchronous motors are the power source for maintaining the stable and efficient operation of elevators. Their reliability is crucial to the normal operation of public elevators, and timely detection of abnormalities in permanent magnet synchronous motors can effectively ensure public safety.

[0003] In actual elevator operation, permanent magnet synchronous motors are characterized by low speed, high torque, and frequent starts. They are susceptible to various factors during use, leading to faults that can be categorized as bearing faults, stator faults, rotor faults, and magnetic circuit faults. Currently, methods for detecting anomalies in permanent magnet synchronous motors primarily involve detecting vibration, current, and magnetic field signals. Vibration signals can effectively reflect internal faults, but they are significantly affected by environmental noise. Magnetic field signals are also greatly influenced by changes in the ambient magnetic field. Non-intrusive current signals, on the other hand, offer advantages such as ease of acquisition, minimal influence from other components, and high precision, making them highly suitable for the complex environment of elevator machine rooms.

[0004] There are various methods for detecting the current of elevator permanent magnet synchronous motors. Sometimes, to save costs, only two-phase currents are collected, and the third-phase current is derived using the formula that the sum of the three voltages at any given moment in a three-phase system is zero. Therefore, the collected original current signal has two possibilities: two-phase or three-phase, and there is a problem of incomplete symmetry between the three-phase currents. Existing current signal detection methods are divided into traditional signal processing methods and deep learning-based methods. Traditional methods generally use control algorithms based on physical models. These algorithms rely on the mathematical model of the system and usually require detailed system parameters and prior knowledge, making them unsuitable for elevator permanent magnet synchronous motor current detection with two possible results. Deep learning-based methods contain network structures with multiple hidden layers and multiple sensing layers, which can learn and represent features at different levels of abstraction. They have advantages such as strong feature extraction capabilities, good adaptability and generalization capabilities, and end-to-end learning, making them more applicable. Existing deep learning methods based on convolutional neural networks directly use the features of the last convolutional layer for fault identification, which fails to fully explore and utilize some useful hierarchical features extracted by intermediate layers, and lacks a fault signal learning mechanism, making them prone to overfitting. This invention provides a DWT-CNN-based method for detecting anomalies in elevator permanent magnet synchronous motors. This method can solve the problem of the number of current input phases and fully extract current feature information, thereby improving the efficiency and accuracy of anomaly detection in elevator permanent magnet synchronous motors. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to provide a method, device and storage medium for abnormal detection of elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion, addressing the problems of large noise interference, diverse current phases and incomplete symmetry of three-phase currents.

[0006] To achieve the above objectives, the technical solution of the present invention includes the following steps: A method for anomaly detection of elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion, the method comprising the following steps:

[0007] Step 1: Collect multiple (three-phase or two-phase) abnormal current signals of the elevator permanent magnet synchronous motor under different operating conditions.

[0008] Step 2: Label the collected current signals to construct a permanent magnet synchronous motor current dataset.

[0009] Step 3: Preprocess the collected current signals and divide them into training set, validation set, and test set.

[0010] Step 4: Construct the MWAL-CNN (Multilayer Wavelet Attention LSTM Convolutional Neural Network) model, which includes: a first discrete wavelet transform layer, a first convolutional neural network layer, a first hybrid attention mechanism layer, a second discrete wavelet transform layer, a second convolutional neural network layer, a second hybrid attention mechanism layer, a third discrete wavelet transform layer, a third convolutional neural network layer, a third hybrid attention mechanism layer, and an anomaly detection classifier, all connected in sequence.

[0011] Specifically, the first, second, and third discrete wavelet transform layers are collectively referred to as high- and low-frequency DWT (Discrete Wavelet Transformation) layers, each comprising: a channel unit formed by parallel connection of low-frequency and high-frequency component channels, a Cat (Concatnate) layer connected in series with the channel unit, and a Dropout (Dropout) function layer connected in series with the Cat layer; the low-frequency component channel comprises: a Conv1d (Conv1d, one-dimensional convolution) convolution layer, a normalization layer, and a ReLU (Rectified Linear Unit) layer connected in series; the high-frequency component channel comprises: a Conv1d convolution layer, a normalization layer, and a ReLU layer connected in series. The first, second, and third hybrid attention mechanism layers are collectively referred to as hybrid attention mechanism layers, each comprising: a cascaded FAM (Channel Frequency Attention Module) channel frequency attention layer and an EAM (Excitation Attention Module) temporal attention layer; the FAM channel frequency attention layer includes a cascaded 1*1 convolutional layer, an average pooling layer, a convolutional layer, a sigmoid activation layer, and a residual block; the EAM temporal attention layer includes a cascaded convolutional layer, a ReLU layer, and a cascaded convolutional layer, a ReLU layer, and a normalization layer. The anomaly detection classifier includes: a cascaded convolutional layer, a normalization layer, a pooling layer, a weight-updating LSTM (Long Short-Term Memory) network layer, and a fully connected layer.

[0012] Step 5: Use the DWT-CNN-based multi-current fusion model obtained in Step 4 to train and update the data in Step 3, and obtain the optimal detection model for abnormal current signals of elevator permanent magnet synchronous motor.

[0013] Step 6: Use the optimal model from Step 5 to detect the test dataset, obtain abnormal current feature information, and classify it.

[0014] Step 7: Output the results based on the abnormal information of the three-phase current of the permanent magnet synchronous motor.

[0015] Furthermore, in step 1, the acquisition of three-phase current abnormal signals refers to using Hall effect sensors to measure three-phase or two-phase current abnormal signals at specified locations on each phase current path, under different environmental noise conditions, and under different abnormal operating states of the permanent magnet synchronous motor.

[0016] Furthermore, in step 2, abnormal current data is labeled on the dataset signal, and the abnormality type is divided into bearing fault, stator fault, and rotor fault, and the abnormal current intensity and abnormal duration area are labeled.

[0017] Furthermore, in step 3, the preprocessing includes dividing the long signal data of each sensor into windows to increase the number of samples, and performing Butterworth filtering and FFT (Fast Fourier Transform) processing on each divided signal.

[0018] Further, in step 3, the abnormal current dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1 for model training.

[0019] Furthermore, in step 4, the MWAL-CNN model structure layer specifically includes a high- and low-frequency DWT layer, a standard CNN layer, a hybrid attention mechanism layer, and an anomaly detection classifier. The first stage consists of a high-frequency and low-frequency DWT layer, several standard CNN layers, and a hybrid attention mechanism layer. The high-frequency and low-frequency DWT layer includes channel units formed by parallel low-frequency and high-frequency component channels, which decompose and process the low-frequency and high-frequency components separately. The low-frequency component channel is constructed by two stacked Conv1d convolutional layers, a normalization layer, and a ReLU layer, while the high-frequency component channel is constructed by one Conv1d convolutional layer, a normalization layer, and a ReLU layer. Finally, it is processed through a Cat layer module and a Dropout function. The hybrid attention mechanism layer includes a FAM channel frequency attention layer and an EAM temporal attention layer. The FAM channel frequency attention layer is first constructed by a 1*1 convolutional layer and an AvgPool (AvgPool, average pooling layer), followed by two convolutional layers to map cross-channel feature information, then a Sigmoid activation layer to map feature weight vectors, and finally a residual block to optimize backpropagation. The EAM temporal attention layer includes two convolutional layers, one ReLU layer, and a normalization layer, and then connects the input features and weight features. The second and third phases have the same structure as the first phase. The fourth phase connects to an anomaly detection classifier, which includes a convolutional layer, a normalization layer, a pooling layer, a weight-updating LSTM network layer, and a fully connected layer.

[0020] Furthermore, in step 4, the high- and low-frequency DWT layer includes a channel unit formed by parallel low-frequency component channels and high-frequency component channels. The convolutional kernel of the high-frequency component channel is a large convolutional kernel (kernel size 5) relative to the low-frequency component, which can filter out most of the interference noise in the measured signal, making the signal clearer and more accurate. The low-frequency component channel processes the low-frequency component of the measured signal. The low-frequency component of the signal contains rich information. The concatenated convolutional layers form a simple encoding and decoding structure. The low-frequency component reconstructed by this structure can make the signal features more obvious and output a more explicit low-frequency feature map, thereby improving the final model performance.

[0021] Furthermore, in step 4, the hybrid attention layer comprising the FAM channel frequency attention layer and the EAM time attention layer is used. The FAM channel frequency attention layer can map the cross-channel information of the measured signal through convolutional layers, obtain different weight vectors with pooling, and extract the frequency and channel information of the measured signal laterally with residual blocks. In particular, it can amplify the less obvious frequency information, making the frequency information of the signal transformed by FFT in step 3 more obvious. The EAM time attention layer can extract the time information of the measured signal vertically. Combined with the LSTM in step 4, it can effectively extract the correlation information between the preceding and following signals, thus improving the final model performance.

[0022] Furthermore, in step 5, the training process is carried out within the PyTorch framework, using the abnormal current signals of the training set for training and the abnormal current signals of the validation set for model evaluation.

[0023] Furthermore, in step 6, the abnormal current information includes the abnormality type, the abnormal current value, the timestamp of the abnormal current occurrence, and the duration.

[0024] Furthermore, in step 7, the output result refers to uploading the detected abnormal current information to the cloud platform and storing and displaying it.

[0025] Secondly, the present invention also provides an abnormal detection device for elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion, including a current signal acquisition module, a PLC (Programmable Logic Controller) controller, a storage device, a communication module, and a PC (Personal Computer). When the PC processor executes the executable code, it implements the abnormal detection method for elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion.

[0026] Thirdly, the invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the aforementioned method for anomaly detection of elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion.

[0027] The beneficial effects of this invention are:

[0028] (1) To address the problem of unclear feature extraction from multiple current signals, a high-low frequency DWT-CNN feature fusion algorithm is designed to perform targeted feature processing on high- and low-frequency current signals and effectively fuse the features of multiple current signals.

[0029] (2) To address the problem of various environmental noise interferences, the multi-scale features of high and low frequency current signals are processed by encoding CNN and standard CNN layers to reduce the high frequency noise generated during elevator operation.

[0030] (3) To address the problem of unclear frequency and time characteristics of multiple current signals, a hybrid attention mechanism of FAM channel frequency attention layer and EAM time attention layer is constructed; the abnormal current signal of permanent magnet synchronous motor is processed by a network structure composed of DWT-CNN-hybrid attention mechanism to form a complete and rich fusion feature signal in the time domain frequency domain, and learn the information hidden in the time domain and frequency domain.

[0031] (4) To improve the accuracy of anomaly type diagnosis, a weight-updated LSTM anomaly detector is proposed, which includes a concatenated convolutional layer, a normalization layer, a pooling layer, a weight-updated LSTM layer and a fully connected layer. This not only enables the effective transmission and expression of information in long-term sequences, but also better reflects the correlation characteristics before and after the occurrence of anomalies. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the abnormal current detection principle of the permanent magnet synchronous motor of the present invention;

[0034] Figure 2 This is a global framework diagram of the improved DWT-CNN model;

[0035] Figure 3 Here is a block diagram of the hybrid attention mechanism in the improved DWT-CNN model;

[0036] Figure 4This is a structural diagram of an abnormal detection device for elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion according to the present invention. Detailed Implementation

[0037] To more clearly illustrate the technical problems, technical solutions, and beneficial effects of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0038] Reference Figure 1 An embodiment of the present invention, a method for detecting anomalies in elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion, includes the following steps:

[0039] Step 1: Signal Acquisition

[0040] Abnormal current signals were acquired using a laboratory-built elevator permanent magnet synchronous motor experimental platform and real-world elevator permanent magnet synchronous motors that experienced malfunctions. The acquisition method involved using Hall effect current sensors to measure abnormal three-phase current signals at the elevator input current terminal under different abnormal operating conditions of the permanent magnet synchronous motor.

[0041] Step 2: Data Preprocessing

[0042] The actual acquired signals are affected by various factors such as electromagnetic interference, mechanical vibration, and power supply stability. To extract key information and reduce errors, noise reduction processing is required. In this embodiment, the long signal data from each sensor is first divided into windows to increase the number of samples. Butterworth filtering and Fast Fourier Transform are then applied to each segmented signal. The window size is 2048.

[0043] Step 3: Dataset Labeling and Construction

[0044] The preprocessed data is labeled. The labeling refers to the labeling of abnormal intensity of the current signal, the labeling of abnormal duration region of the current signal, and the labeling of abnormal type.

[0045] The dataset construction refers to dividing the labeled current signal set into a training set, a validation set, and a test set in a ratio of 7:2:1.

[0046] Step 4: Construct the algorithm model

[0047] The overall structure of the improved DWT-CNN-based multi-current fusion model is as follows: Figure 2 As shown, the overall structure includes multiple DWT-CNN-hybrid attention mechanism layers and an anomaly detection classifier. Figure 2Input(Input) is the input of the preprocessed signal. It can input single-phase current, two-phase current and three-phase current. The input signal size for single-phase current is (64, 1, 2048), the input signal size for two-phase current is (64, 2, 2048), and the input signal size for three-phase current is (64, 3, 2048).

[0048] Phase 1: The current signal is preprocessed to form time series data, ready for training, and then passed through one standard DWT layer ( Figure 2 The DWT length is shortened, and the output is (64, 4, 1039).

[0049] The second stage: The DWT-CNN-hybrid attention mechanism layer extracts the frequency and time features of the multi-current signal. The DWT-CNN-hybrid attention mechanism layer includes a high- and low-frequency DWT layer, a standard CNN layer, and a hybrid attention layer.

[0050] In the DWT-CNN-hybrid attention mechanism layer, the DWT is a high- and low-frequency discrete wavelet transform, which first passes through a standard CNN layer with a kernel size of 3. Figure 2 sconv1) extracts feature information, outputting a size of (64, 4, 1037), and then passes it through a DWT layer ( Figure 2 DWT), output size (64, 12, 534), will be decomposed into low-frequency signal components and high-frequency signal components, and then processed separately. The low-frequency current component is formed by a simple encoding and decoding through two one-dimensional convolutions with smaller kernels, including two sets of cascaded Conv1d convolutional layers, normalization layers, and ReLU layers to form the encoding Conv1d, finally obtaining a clearer low-frequency feature map. Through the first encoding Conv1d ( Figure 2 After sconv2), the second dimension of the input is halved, resulting in an output size of (64, 6, 532). Conv1d is then decoded using a second kernel of size 3. Figure 2 The second dimension of the input is restored after sconv4, and the output size is (64, 12, 530); the high-frequency current component is reduced for noise by a one-dimensional convolution with a large kernel, and is formed by a set of Conv1d convolutional layers, normalization layers and ReLU layers. Figure 2 sconv3), output size (64, 6, 532). After processing the high and low frequency current components separately, they are recombined in the Cat layer, with an output size of (64, 24, 530), followed by the Dropout function. Figure 2 Dropout improves robustness while keeping the output size constant.

[0051] The hybrid attention mechanism layer in the DWT-CNN hybrid attention mechanism layer is divided into a FAM channel frequency attention layer and an EAM temporal attention layer. (FAM channel frequency attention layer) Figure 3The upper part first goes through a 1*1 convolutional layer ( Figure 3 Conv1 reduces input feature aliasing, then passes through AvgPool ( Figure 3 Avgpool compresses global information into channel representation vectors, followed by two 1x1 convolutional layers. Figure 3 Conv2 and Conv3) map cross-channel feature information, and then introduce a Sigmoid activation layer ( Figure 3 The Sigmoid algorithm maps feature vectors to the 0-1 interval to form feature weight vectors, and finally introduces residual blocks ( Figure 3 X*Input_x+Input_x) optimizes backpropagation, outputting the frequency characteristics of the current signal; the EAM time attention layer is... Figure 3 The lower half consists of a 1x1 convolutional layer connected in series. Figure 3 Conv1), a ReLU layer ( Figure 3 ReLU), then concatenated with a 1*1 convolutional layer ( Figure 3 Conv2), a normalization layer ( Figure 3 Bathnorm (Batch Normalization) and a Sigmoid activation layer ( Figure 3 Composed of Sigmoids, the concatenated structure can output different weight values, which are then multiplied by matrix multiplication. Figure 2 X*Y+Input_x) connects the input features and weight features, and finally outputs the time features of the current signal.

[0052] The hybrid attention mechanism layer extracts only subtle features, modifies the weights of some inconspicuous features, keeps the output size unchanged, and is then followed by Maxpool(maxpool, max pooling). Figure 2 maxpool prevents overfitting, output size (64, 24, 256).

[0053] The current signal passes through three layers of the DWT-CNN-hybrid attention mechanism, with the output size decreasing sequentially and the second dimension features increasing sequentially. The output sizes of the first layer are (64, 24, 256), the second layer is (64, 48, 69), and the third layer is (64, 96, 3). The time and frequency characteristics of the output current signal are more obvious.

[0054] Phase Three: As Figure 2 As shown, this stage is an anomaly detection classifier that performs a linear mapping between fused features and the output. The anomaly detection classifier contains one convolutional layer. Figure 2 sconv1), pooling layer ( Figure 2 avgpool), 1 LSTM layer with weight updates ( Figure 2LSTM and fully connected layers Figure 4 Before the current signal enters the anomaly detection classifier, it first passes through an AvgPool average pooling layer to compress the global information into a channel representation vector, with an output size of (64, 96, 1). Then, it undergoes matrix transformation to exchange the second and third dimension data, passes through an LSTM, and has an output size of (64, 1, 64). After the anomaly detection classifier, redundant dimension data is removed, and the signal passes through a Fuction (Fuction, fully connected function) layer to output the anomaly current signal detection classification result, with a result size of (64, 14).

[0055] Step 5: Model Training and Validation

[0056] The improved DWT-CNN-based multi-current fusion model obtained in step 4 was used to train and validate the training and validation sets from step 3. The experiment was conducted using the PyTorch framework. During model training, the Adam (Adaptive Moment Estimation) optimization algorithm was used to dynamically adjust the learning rate to adapt to gradient changes at different parameters and time steps. Simultaneously, first-norm regularization was employed to prevent overfitting of the training data. The initial learning rate was set to 0.0001, the batch size to 300, and the batch size to 64. The training set data was first input into the model for training, and then the validation set was used for model validation. The results were evaluated; if the performance was unsatisfactory, the model was adjusted and training continued until the expected results were achieved.

[0057] Step 6: Abnormal Current Signal Detection

[0058] Another set of abnormal current signals will be collected following the procedure in step 1 and preprocessed. The DWT-CNN-based multi-current fusion model from step 5 will be used to detect this data, obtaining abnormal current feature information and classifying it. Accuracy, recall, and F1 score will be used as evaluation criteria. If the detection performance is unsatisfactory, the model will be retrained and tested again to ensure a better detection result. The final detection accuracy is 99.91%.

[0059] Step 7: Output Results

[0060] The classified abnormal current signals are subjected to operations such as interception, amplification, time-frequency analysis, and peak detection. The obtained abnormal current intensity, duration, and timestamp information are uploaded to the cloud platform and can be viewed on the cloud platform interface.

[0061] The present invention also provides an embodiment of an abnormal detection device for elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion, the detection device corresponding to the above method.

[0062] like ​As shown, an embodiment of an elevator permanent magnet synchronous motor anomaly detection device based on DWT-CNN multi-current fusion includes a current signal acquisition module, a PLC controller, a storage device, a communication module, and a PC. The current signal acquisition module has an ADC (Analog-to-digital converter) for acquiring current signals and converting analog signals into digital signals. The PLC controller is used to monitor and control the detection task, realizing information exchange and integration between devices. The storage device stores the acquired data. The communication module transmits data to the PC using the TCP / IP (Transmission Control Protocol / Internet Protocol) protocol. The PC refers to any computer with data processing capabilities and display, used to implement the elevator permanent magnet synchronous motor anomaly detection method based on DWT-CNN multi-current fusion.

[0063] The present invention provides an embodiment of an abnormal detection device for elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion, which is implemented by combining software and hardware. The operation instructions can be transmitted from the PC to the PLC, or the PLC can directly run the internal processor instructions. The implementation process of each unit and module is detailed in the corresponding steps of the above method, and will not be repeated here.

[0064] The device embodiments described above can determine the spatial position of each unit according to actual needs, and include but are not limited to the above hardware. Those skilled in the art can understand and implement them without creative effort.

[0065] This invention also provides a computer-readable storage medium, which can be either a non-volatile or volatile computer-readable storage medium. The computer-readable storage medium stores instructions for an anomaly detection method of elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion, which can be executed by a computer. The specific working process of the system can refer to the corresponding steps in the foregoing method embodiments. The computer-readable storage medium can serve as an internal storage unit of the PC in the above device embodiments, or as any external storage device with data processing capabilities, including a portable hard drive, SD card (Secure Digital Card), optical disc, etc.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit it. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for anomaly detection in elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion, characterized in that, Includes the following steps: Step 1: Collect multiple abnormal current signals of the elevator permanent magnet synchronous motor under different operating conditions; Step 2: Label the collected multiple abnormal current signals to construct a permanent magnet synchronous motor current dataset; Step 3: Preprocess the data in the permanent magnet synchronous motor current dataset and divide it into training and validation sets; Step 4: Construct the MWAL-CNN model, which includes: a first discrete wavelet transform layer, a first convolutional neural network layer, a first hybrid attention mechanism layer, a second discrete wavelet transform layer, a second convolutional neural network layer, a second hybrid attention mechanism layer, a third discrete wavelet transform layer, a third convolutional neural network layer, a third hybrid attention mechanism layer, and an anomaly detection classifier, all connected in series. The first discrete wavelet transform layer, the second discrete wavelet transform layer, and the third discrete wavelet transform layer each include: a channel unit formed by connecting low-frequency component channels and high-frequency component channels in parallel, a Cat layer connected in series with the channel unit, and a Dropout function layer connected in series with the Cat layer; The low-frequency component channel includes: a Conv1d convolutional layer, a normalization layer, and a ReLU layer connected in series. The high-frequency component channel includes: a Conv1d convolutional layer, a normalization layer, and a ReLU layer connected in series. The first, second, and third hybrid attention mechanism layers all include: a cascaded FAM channel frequency attention layer and an EAM time attention layer; The FAM channel frequency attention layer is specifically as follows: First, a 1*1 convolutional layer is passed to reduce the input feature aliasing effect. Then, the global information is compressed into a channel representation vector through AvgPool. Next, two 1*1 convolutional layers are passed to map cross-channel feature information. Then, a Sigmoid activation layer is introduced to map the feature vector to the 0-1 interval to form a feature weight vector. Finally, a residual block is introduced to optimize backpropagation and output the frequency features of the current signal. The EAM temporal attention layer specifically includes two attention branches. The first branch consists of a 1*1 convolutional layer and a ReLU layer connected in series. The second branch consists of a 1*1 convolutional layer, a normalization layer, and a sigmoid activation layer connected in series. The output of the cascaded structure is different weight values. Then, the input features and weight features are connected by matrix multiplication to finally output the temporal features of the current signal X*Y+Input_x. Step 5: Use the MWAL-CNN model obtained in Step 4 to train and update the training set and validation set in Step 3 to obtain the optimal detection model for abnormal current signals of elevator permanent magnet synchronous motor. Step 6: Use the optimal elevator permanent magnet synchronous motor abnormal current signal detection model from Step 5 to detect the abnormal current signal of the elevator permanent magnet synchronous motor to be tested, obtain abnormal current feature information and classify it, and output the results.

2. The method for anomaly detection of elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion according to claim 1, characterized in that, In step 3, the preprocessing includes: window division, Butterworth filtering, and FFT.

3. The method for anomaly detection of elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion according to claim 1, characterized in that, In step 4, the anomaly detection classifier includes a convolutional layer, a normalization layer, a pooling layer, an LSTM network layer for weight updating, and a fully connected layer connected in series.

4. The method for anomaly detection of elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion according to claim 1, characterized in that, In step 4, the abnormal current characteristic information includes the abnormal type, abnormal current value, timestamp of the abnormal current occurrence, and duration.

5. An abnormal detection device for elevator permanent magnet synchronous motors based on DWT-CNN multi-current fusion, comprising a current signal acquisition module, a PLC controller, a communication module, and a PC, characterized in that, When the processor of the PC executes the executable code, it implements the abnormal detection method for elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion as described in any one of claims 1-4.

6. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the abnormal detection method for elevator permanent magnet synchronous motor based on DWT-CNN multi-current fusion as described in any one of claims 1-4.

Images