A multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution

Abstract

The present invention relates to a multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution, which mainly solves the technical problem that it is difficult for traditional convolution networks to accurately depict the multi-variable and non-linear relationships contained therein. The technical solution adopted by the present invention is as follows: The specific steps of the demagnetization monitoring method are as follows: S1: Collect demagnetization-sensitive characteristic parameters from existing data, and determine the types of required multi-channel sensors and the sensor arrangement forms; S2: Collect demagnetization characteristic signal data under typical working conditions of the permanent magnet synchronous traction machine through multi-channel sensors, and construct a representative demagnetization characteristic database; S3: According to the physical properties and sensitive manifestation forms of each demagnetization-sensitive characteristic, preprocess the characteristic signal data in the demagnetization characteristic signal database, and update the demagnetization characteristic database; S4: Build a lightweight network model; S5: According to the demagnetization degree interval division, determine the output type of the network; perform network configuration and network training.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator fault monitoring, and particularly relates to a multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution. Background Art

[0002] With the increase in high-rise buildings, elevators have become an essential part of people's daily lives. Permanent magnet synchronous traction machines have the advantages of large low-speed output torque, compact structure, small occupied space, and easy installation, and are commonly used as power devices for elevators. However, under the combined influence of driving force, magnetic field, and thermal stress, permanent magnet synchronous traction machines are prone to sudden changes in the stability of permanent magnet materials and demagnetization phenomena.

[0003] Since the permanent magnets of the traction machine are installed on the rotor, during the daily operation and maintenance of the elevator, it is necessary to disassemble the permanent magnet traction machine to determine the demagnetization fault. Both its installation and disassembly require special tools and the procedures are complicated, which brings great inconvenience to on-site demagnetization detection.

[0004] At present, the detection methods for the demagnetization of permanent magnet synchronous traction machines are relatively single, and most of the demagnetization detections are in the simulation stage, lacking detection methods for application. To solve this problem, some researchers have used convolutional neural networks to extract and correlate the time-frequency domain information and spatial information contained in vibration signals. However, when dealing with complex system problems of multi-sensor information fusion, in the face of various signals with different physical properties, and each signal has different fault characteristics, traditional convolutional networks are no longer sufficient to accurately depict the multi-variable and non-linear relationships contained therein, and the diagnostic effects shown are also unsatisfactory. Summary of the Invention

[0005] The purpose of the present invention is to solve the above technical problems and provide a multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] A multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution, the specific steps of which are as follows:

[0008] S1: Collect the demagnetization manifestation forms of permanent magnet synchronous traction machines from existing materials, screen out the demagnetization-sensitive characteristic parameters under the demagnetization manifestations, and combine with the elevator control system to determine the types and layout forms of the required multiple sensors;

[0009] S2: Set the determined multiple sensors on the permanent magnet synchronous traction machine according to the layout form, collect the demagnetization characteristic signal data of the permanent magnet synchronous traction machine under typical working conditions through the multiple sensors, and construct a representative demagnetization characteristic database;

[0010] S3: According to the physical properties and sensitive manifestation forms of each demagnetization-sensitive feature, preprocess the feature signal data in the demagnetization feature signal database, and update the demagnetization feature database;

[0011] S4: Build a lightweight network model, introduce the depthwise separable method to establish a multi-sensor information fusion framework, build a multi-scale feature extraction module for each sensor channel, use stacking to integrate multi-scale extracted features, and introduce a channel attention mechanism to guide the network to focus on feature information that is more beneficial to demagnetization classification;

[0012] S5: According to the demagnetization degree interval division, determine the output type of the network; perform network configuration and network training.

[0013] Furthermore, the demagnetization-sensitive feature parameters include the demagnetization fault characteristic phase current of the permanent magnet synchronous traction machine, the load end temperature rise change rate, the output speed pulsation of the traction wheel, the dynamic output torque of the traction wheel during normal operation, and the body vibration parameters;

[0014] The required multi-channel sensor types and sensor arrangement forms are specifically as follows: For a permanent magnet synchronous traction machine control system using id=0 control, set a speed monitoring sensor, a rotor position monitoring sensor, a current sensor on the three-phase power inverter, a housing temperature sensor on the housing, a torque sensor on the rotating shaft, and a body vibration acceleration sensor on the body in the permanent magnet synchronous traction machine control system.

[0015] Furthermore, the typical working conditions are: full-load start-up and rise, full-load constant-speed rise, no-load start-up and rise, no-load constant-speed rise, full-load rise braking, no-load rise braking, full-load start-down, full-load constant-speed down, no-load start-down, no-load constant-speed down, full-load down braking, and no-load down braking.

[0016] Furthermore, the demagnetization feature signal data is the phase current of the permanent magnet synchronous traction machine under each typical operating condition sampled by the current sensor, the load end temperature rise change of the permanent magnet synchronous traction machine under each typical operating condition sampled by the housing temperature sensor, the output speed pulsation of the traction wheel of the permanent magnet synchronous traction machine pre-sampled by the rotor position monitoring sensor, the dynamic output torque of the traction wheel of the permanent magnet synchronous traction machine pre-sampled by the torque sensor, and the body vibration of the permanent magnet synchronous traction machine pre-sampled by the inertial vibration measuring instrument.

[0017] Furthermore, the preprocessing of the feature signal data in the demagnetization feature database refers to processing according to the data physical characteristics and sensitive manifestation forms respectively, specifically as follows:

[0018] For the fuselage vibration signal data, a window function is used for processing, that is, a window with a length of L is overlapped and sampled on the characteristic signal data with a length of N at a sliding step of S, and normalization is performed;

[0019] For the dynamic output torque and rotational speed pulsation characteristic signal data, it is divided into start-stop process signals and stable operation pulsation signals. For the start-stop process signal data, a processing method of dividing by the maximum torque during stable operation without losing magnetism is adopted; for the stable operation pulsation characteristic signal data, the same processing method as that of the fuselage vibration is adopted, that is, a window with a length of L is overlapped and sampled on the characteristic signal data with a length of N at a sliding step of S, and normalization is performed;

[0020] For the phase current characteristic signal data, normalization processing is performed on the stable operation signal;

[0021] For the characteristic signal data of the temperature rise change at the load end, a processing method of dividing by the temperature rise in the stable operation state without losing magnetism is adopted.

[0022] Furthermore, the multi-scale feature extraction module uses convolution kernels covering large, medium, and small feature scales and is sensitive to information at each stage frequency;

[0023] The adoption of stacking and integrating multi-scale feature extraction means that the extracted features are processed in a stacking manner rather than an adding manner to maximize the preservation of the feature information extracted by each multi-scale module;

[0024] The introduction of the depthwise separable method to establish a multi-sensor information fusion framework is specifically as follows: feature fusion adopts the method of pointwise convolution, and after pointwise convolution, global pooling is adopted to reduce network parameters; in particular, before softmax classification, the method of connecting a fully connected layer after Flatten is cancelled, and instead, global average pooling is used to further reduce the number of parameters;

[0025] The channel attention mechanism is specifically as follows: after extracting feature parameters through the multi-scale module, the stacked features respectively pass through one-dimensional global average pooling and one-dimensional global max pooling to integrate global feature information; after adjusting the dimensions, the results of the two poolings are respectively subjected to one-dimensional convolution without bias terms to capture the mutual relationship between the features of each channel, and then the results of the two-way convolution are added and passed through an activation function to generate adaptive weights for each channel. The obtained adaptive weights are multiplied by the features of the stacked layer and enter the feature fusion link.

[0026] Furthermore, according to the division of the demagnetization degree interval, the output type of the network is determined. Specifically, the demagnetization degree interval is divided into three stages, namely, mild demagnetization, moderate demagnetization, and severe demagnetization stages. Accordingly, the output type of the network is determined as 4 diagnostic labels: no demagnetization, mild demagnetization, moderate demagnetization, and severe demagnetization. The processed characteristic signal data is assigned to the above 4 diagnostic labels, and the diagnostic labels are encoded in one-hot form.

[0027] Furthermore, the network configuration is specifically that the optimizer selects the Adaptive Moment Estimation (Adma) method, which can adaptively adjust the learning rate of each parameter. Combined with cross-entropy as the loss function, it can accelerate the convergence speed when the model effect is poor, making the training data distribution close to the real data distribution.

[0028] The network training means dividing the processed characteristic signal data into a training set and a test set according to a certain proportion, randomly sorting the training set data and the test set data respectively, feeding them into the network, and injecting noise during training to improve the network's anti-noise ability. To ensure the consistency of the training parameter dimension features, the same length of data is used for training in each channel, and all-zero padding is adopted. The length N of the characteristic signal data is taken to have the characteristic independent of the sampling initial phase, including at least one signal cycle, and is based on the lowest characteristic frequency in the fault space signal under the same sampling frequency.

[0029] Furthermore, considering that the elevator is a special equipment for carrying people, to ensure the safe and reliable operation of the system, a diagnostic method outside the system is adopted to extract the signal features and diagnose the demagnetization of the traction machine without interfering with the operation of the traction system.

[0030] The beneficial effects of the present invention are as follows:

[0031] Aiming at the non-linear nature among the parameters of the permanent magnet synchronous traction machine, the present invention utilizes multi-sensor information fusion and lightweight multi-scale neural network to consider the contribution weights of different demagnetization-sensitive features to the demagnetization of the permanent magnet synchronous traction machine. The demagnetization-sensitive features are extracted as the measurement targets of multi-sensors, and the sensors are selected and arranged in combination with the actual operation selection scheme. A database is built for typical working conditions to make the samples more representative. The present invention introduces a depthwise separable convolution framework. For different sensor messages, an optimized multi-scale network can be used to extract features before fusion, fully ensuring the flexibility of network construction and the adaptive feature extraction ability of the network. Using this framework can greatly reduce the network parameters to ensure the real-time performance of diagnosis. A multi-scale convolution feature extraction module is built to expand the system feature extraction interval, taking into account both detailed information and overall features, and improving the network robustness. The channel attention mechanism is introduced to focus on the representative feature information of each channel that is more significant for improving the demagnetization feature classification task. Description of the Drawings

[0032] Figure 1 Schematic diagram of the overall framework of the diagnostic process of the present invention;

[0033] Figure 2 Schematic diagram of multi-scale feature extraction of the present invention;

[0034] Figure 3 Schematic diagram of the overall structure of the network of the present invention;

[0035] Figure 4 Schematic diagram of the depthwise separable framework of the present invention;

[0036] Figure 5 Schematic diagram of the sensor arrangement form of the present invention;

[0037] In the figure: S1, S2, and S3 are the multi-scale convolution kernel sizes, F in is the network input feature size, C is the number of channels, i.e., the number of sensors, m2 is the number of pointwise convolution kernels of the convolution kernel, and is also the feature depth after pointwise convolution, F out is the feature size after pointwise convolution, F1, F2, and F3 are the features extracted by the multi-scale convolution kernels, A1 to A 12 are the feature information extracted by the network. Specific embodiments

[0038] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0039] As shown in the attached Figure 1 figures, a multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution in this embodiment specifically includes the following steps:<t

[0040] S1: Collect the demagnetization manifestation forms of permanent magnet synchronous traction machines from existing materials, screen out the demagnetization-sensitive characteristic parameters under the demagnetization manifestations, and combine with the elevator control system to determine the required types of multi-channel sensors and the sensor arrangement form;

[0041] The demagnetization-sensitive characteristic parameters include the demagnetization fault characteristic phase current of the permanent magnet synchronous traction machine, the load-end temperature rise change rate, the output speed pulsation of the traction wheel, the dynamic output torque of the traction wheel during normal operation, and the body vibration parameters;

[0042] The selection principle of the multi-channel sensor types is: independent of the normal operation of the elevator, and try to utilize the existing sensors during the operation of the permanent magnet synchronous traction machine without adding additional sensor arrangements. The sensor arrangement is as shown in Figure 5As shown in the figure. The required multi-channel sensor types and sensor arrangement forms are specifically as follows: For the permanent magnet synchronous traction machine control system controlled by id = 0, a rotational speed monitoring sensor is set in the permanent magnet synchronous traction machine control system, a rotor position monitoring sensor is set, a current sensor is set on the three-phase power inverter, a housing temperature sensor is set on the housing, a torque sensor is set on the rotating shaft, and a body vibration acceleration sensor is set on the body.

[0043] S2: Set the determined multi-channel sensors on the permanent magnet synchronous traction machine according to the arrangement form, collect the demagnetization characteristic signal data of the permanent magnet synchronous traction machine under typical working conditions through the multi-channel sensors, and construct a representative demagnetization characteristic database;

[0044] Preferably, the selected typical working conditions are: full-load starting and rising, full-load constant-speed rising, no-load starting and rising, no-load constant-speed rising, full-load rising braking, no-load rising braking, full-load starting and falling, full-load constant-speed falling, no-load starting and falling, no-load constant-speed falling, full-load falling braking, and no-load falling braking.

[0045] The demagnetization characteristic signal data are the phase currents of the permanent magnet synchronous traction machine under various typical operating conditions sampled by the current sensor, the load-end temperature rise changes of the permanent magnet synchronous traction machine under various typical operating conditions sampled by the housing temperature sensor, the traction wheel output speed pulsations of the permanent magnet synchronous traction machine under various typical operating conditions pre-sampled by the rotor position monitoring sensor, the dynamic output torque of the traction wheel of the permanent magnet synchronous traction machine under various typical operating conditions pre-sampled by the torque sensor, and the body vibration signal data of the permanent magnet synchronous traction machine under various typical operating conditions pre-sampled by the inertial vibration measuring instrument.

[0046] Use the CAN bus to realize the transmission of the above signal data to ensure the timeliness of information and the accuracy of parameters. Each sensor signal is introduced into the DSP for unified processing after passing through the amplifier circuit and filter circuit as required, and the processed signal is imported into the PC host computer to construct a demagnetization characteristic database for use in the next step of fault information collection and further analysis and diagnosis.

[0047] S3: According to the physical properties and sensitive manifestation forms of each demagnetization-sensitive characteristic, preprocess the characteristic signal data in the demagnetization characteristic signal database and update the demagnetization characteristic database;

[0048] The preprocessing of the characteristic signal data in the demagnetization characteristic database refers to processing according to the physical characteristics and sensitive manifestation forms of the data, specifically as follows:

[0049] For the body vibration signal data, use a window function for processing, that is, a window with a length of L is overlapped and sampled on the characteristic signal data with a length of N at a sliding step size of S, and normalization is performed;

[0050] For the dynamic output torque and rotational speed pulsation characteristic signal data, it is divided into start-stop process signals and stable operation pulsation signals. For the start-stop process signal data, the processing method of dividing by the maximum torque during stable operation without losing magnetism is adopted; for the stable operation pulsation characteristic signal data, the same processing method as the body vibration is adopted, that is, a window with a length of L is overlapped and sampled on the characteristic signal data with a length of N with a sliding step of S, and normalization is performed.

[0051] For the phase current characteristic signal data, normalization processing is performed on the stable operation signal.

[0052] For the load end temperature rise change characteristic signal data, the processing method of dividing by the temperature rise in the stable operation state without losing magnetism is adopted.

[0053] The above normalization method adopts the maximum and minimum value method, and the formula is:

[0054]

[0055] where: x i is the sequence signal, is the normalized sequence. To ensure the consistency of the training parameter dimension features, the same length of data is used for training in each channel.

[0056] The value of the length N of the characteristic signal data is taken to have the characteristic independent of the sampling initial phase, and at least includes one signal cycle. The signals collected by each sensor are truncated, and based on the lowest frequency in the fault space signal under the same sampling frequency. The value of N should not be too large, as it will increase the calculation amount and weaken the real-time performance of diagnosis.

[0057] S4: Build a lightweight network model, introduce the depthwise separable method to establish a multi-sensor information fusion framework, build a multi-scale feature extraction module for each sensor channel, stack and integrate the multi-scale extracted features, introduce a channel attention mechanism to guide the network to focus on the feature information more beneficial to the demagnetization classification, and the feature information is as Figure 3 shown as A1 to A 12 ; A1 to A 12 are the feature information extracted by the multi-scale convolution;

[0058] The above-mentioned lightweight network model is built using one-dimensional convolution, as shown in Appendix Figure 3 ; the network framework is a depthwise separable information fusion framework. A multi-scale convolution feature extraction module is adopted for a single channel under the depthwise separable framework, and an attention mechanism is additionally introduced before feature fusion to enhance the guiding role of favorable features.

[0059] Specifically, for the first one-dimensional convolution, a large-scale convolution kernel is used, with a convolution kernel size of 127×1 and a stride of 4, which is equivalent to introducing a low-frequency filter to enhance the robustness of the diagnostic system in a noisy environment.

[0060] Specifically, for the last layer of the network, instead of using the method of Flatten connected to a fully connected network, the method of Global Average Pooling 1D connected to softmax is adopted. Compared with the redundant method of flattening all features, the global average pooling layer can effectively alleviate overfitting, and using this structure can further reduce the network parameters.

[0061] The method of introducing depthwise separable convolution is used to establish a multi-sensor information fusion framework, as shown in the appendix Figure 4 Specifically: for feature extraction of each channel, depthwise convolution is adopted; for feature fusion, pointwise convolution is adopted. Specifically, multiple 1×1 convolution kernels are used, all with zero padding, and L2 regularization is used to control the network weights to reduce overfitting caused by noise, and the regularization parameter is selected as 0.01.

[0062] The multi-scale feature extraction module is as shown in the appendix Figure 2 Specifically, convolution kernels covering large, medium, and small feature scales are used, which are sensitive to information at each stage frequency; the large-scale convolution kernel has a convolution size of 27×1, the medium-scale convolution kernel has a convolution size of 9×1, and the small-scale convolution kernel has a convolution size of 3×1. To make the extracted feature dimensions the same, the convolution stride is set to 2 for all.

[0063] After the convolution operation, Batch Normalization is used to standardize the data of the current training batch, which can accelerate the convergence speed and improve the generalization ability, and the momentum parameter is set to 0.99.

[0064] The ReLU function is selected as the convolution activation function. While avoiding the vanishing gradient, it can alleviate overfitting and further reduce the computational amount. The formula of the ReLU function is:

[0065]

[0066] where: x is the input feature to be activated.

[0067] The max pooling method is selected for pooling. For the demagnetization diagnosis classification, it can highlight the favorable features, avoid overfitting, and further reduce the data dimension. The pooling stride is set to 4.

[0068] The method of stacking and integrating multi-scale extracted features means that the extracted features are processed by Concatenate in a stacking manner, rather than an adding manner, so as to maximize the preservation of the feature information extracted by each multi-scale module, and finally the feature of the stacking layer is obtained as merge;

[0069] The channel attention mechanism is specifically as follows: After extracting feature parameters through the multi-scale module, the stacked features respectively pass through one-dimensional global average pooling (GlobalAveragePooling1D) and one-dimensional global max pooling (GlobalMaxPooling1D) to integrate global feature information.

[0070] After adjusting the dimensions, the results of the two poolings are respectively subjected to one-dimensional convolution without bias terms. The purpose is to utilize the good cross-channel information integration ability of convolution to capture the mutual relationship between the features of each channel. The size of the one-dimensional convolution kernel without bias is determined by the number of channels according to the following formula:

[0071]

[0072] In the formula: k is the size of the convolution kernel, log is the logarithmic function, ch is the number of channels of the input convolution kernel, and abs is the absolute value.

[0073] Then, the results of the two-way convolution are added together, which is beneficial for comprehensively considering the overall features and local prominent features; the added result passes through the sigmoid activation function to generate adaptive weights for each channel; the obtained adaptive weights are multiplied by the stacked layer features (merge) and enter the feature fusion link.

[0074] S5: According to the division of the field loss degree interval, determine the output type of the network; perform network configuration and network training.

[0075] According to the division of the field loss degree interval, determine the output type of the network, specifically: divide the field loss degree interval into three stages, namely mild field loss (field loss percentage is less than 30%), moderate field loss (field loss percentage is higher than 30% and lower than 60%), and severe field loss (field loss percentage is higher than 60%). According to this, determine the output type of the network as 4 diagnostic labels: no field loss situation, mild field loss, moderate field loss, and severe field loss. Endow the processed feature signal data with the above 4 diagnostic labels, and adopt the one-hot encoding form for the diagnostic labels.

[0076] The network configuration is specifically that the optimizer selects the Adaptive Moment Estimation (Adma) method, which can adaptively adjust the learning rate of each parameter. The learning rate parameter lr is set to 0.001, the exponential decay rate parameter beta_1 of the first moment estimation is set to 0.9, and the exponential decay rate parameter beta_2 of the second moment estimation is set to 0.99. Combined with the cross-entropy as the loss function, it can accelerate the convergence speed when the model effect is poor, avoid the phenomenon that the model convergence becomes slow due to too small partial derivatives, and make the training data distribution close to the real data distribution. The formula of the loss function is:

[0077]

[0078] Where: n is the number of categories; y i (x) is the one-hot training data label; is the network prediction label; Loss is the cross entropy loss function.

[0079] The network training involves dividing the processed feature signal data into a training set and a test set proportionally, i.e., 70% of the data is allocated to the training set and 30% to the test set. The training set data and the test set data are randomly sorted and fed into the network, and noise is injected during training to improve the network's noise resistance. Preferably, 0.2% of the training set data is used as the validation set, and the training effect is verified after each batch of training.

[0080] Training in small batches is beneficial to network training. The training data is trained in batches of 32.

[0081] To ensure the consistency of the training parameter dimension features, each channel uses data of the same length for training, and all are padded with zeros; the characteristic signal data length N is taken to have characteristics that are independent of the initial sampling phase, including at least one signal cycle, and under the same sampling frequency, the lowest characteristic frequency in the fault space signal is used as the benchmark.

[0082] Furthermore, considering that elevators are special equipment for carrying people, in order to ensure the safety and reliability of system operation, a diagnostic method outside the system is used to extract signal features and diagnose traction machine demagnetization without interfering with the operation of the traction system.

[0083] In summary, the present invention constructs a one-dimensional multi-scale separable convolutional network for online diagnosis of demagnetization of permanent magnet synchronous traction machines. Compared with traditional one-dimensional convolutional networks, the present invention makes full use of multi-sensor information, eliminates the need to build a recognition framework, and takes into account both noise processing and demagnetization diagnosis. Samples are collected under typical working conditions, and a more representative demagnetization feature database is built. The present invention uses a multi-scale module to extract the channel features of each sensor, enriches the multi-scale features, and introduces a channel attention mechanism to adaptively assign weights to each channel, thereby improving the network's adaptability. The present invention introduces a separable network framework, which can increase the adjustment and optimization of the network structure, take into account the depth and width of the network, greatly reduce network parameters, and obtain a network with better accuracy and real-time performance.

Claims

1. A multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution, characterized in that The specific steps are as follows: S1: Collect the demagnetization manifestation forms of permanent magnet synchronous traction machines from existing data, screen out the demagnetization-sensitive characteristic parameters under the demagnetization manifestations, and combine with the elevator control system to determine the types of required multi-channel sensors and the sensor arrangement forms; S2: Set the determined multi-channel sensors on the permanent magnet synchronous traction machine according to the arrangement form, collect the demagnetization characteristic signal data of the permanent magnet synchronous traction machine under typical working conditions through the multi-channel sensors, and construct a representative demagnetization characteristic database; S3: According to the physical properties and sensitive manifestation forms of each demagnetization-sensitive characteristic, preprocess the characteristic signal data in the demagnetization characteristic signal database, and update the demagnetization characteristic database; S4: Build a lightweight network model, introduce the depthwise separable method to establish a multi-sensor information fusion framework, build a multi-scale feature extraction module for each sensor channel, use stacking to integrate multi-scale extracted features, and introduce a channel attention mechanism to guide the network to focus on the characteristic information for demagnetization classification; S5: Determine the output types of the network according to the demagnetization degree interval division; perform network configuration and network training; The demagnetization-sensitive characteristic parameters include the demagnetization fault characteristic phase current of the permanent magnet synchronous traction machine, the load-end temperature rise change rate, the output speed pulsation of the traction wheel, the dynamic output torque of the traction wheel during normal operation, and the body vibration parameters; The types of required multi-channel sensors and the sensor arrangement forms are specifically as follows: for the permanent magnet synchronous traction machine control system using id=0 control, set a speed monitoring sensor in the permanent magnet synchronous traction machine control system, set a rotor position monitoring sensor, set a current sensor on the three-phase power inverter, set a housing temperature sensor on the housing, set a torque sensor on the rotating shaft, and set a body vibration acceleration sensor on the body; The taken typical working conditions are: full-load start-up and rise, full-load constant-speed rise, no-load start-up and rise, no-load constant-speed rise, full-load rise braking, no-load rise braking, full-load start-up and fall, full-load constant-speed fall, no-load start-up and fall, no-load constant-speed fall, full-load fall braking, and no-load fall braking; The demagnetization characteristic signal data are the phase currents of the permanent magnet synchronous traction machine under each typical operating condition sampled by the current sensor, the load-end temperature rise changes of the permanent magnet synchronous traction machine under each typical operating condition sampled by the housing temperature sensor, the output speed pulsation of the traction wheel of the permanent magnet synchronous traction machine under each typical operating condition pre-sampled by the rotor position monitoring sensor, the dynamic output torque of the traction wheel of the permanent magnet synchronous traction machine under each typical operating condition pre-sampled by the torque sensor, and the body vibration signal data of the permanent magnet synchronous traction machine under each typical operating condition pre-sampled by the inertial vibration measuring instrument; The preprocessing of the characteristic signal data in the demagnetization characteristic database refers to processing according to the physical characteristics and sensitive manifestation forms of the data, specifically as follows: For the fuselage vibration signal data, a window function is used for processing, that is, a window with a length of L is overlapped and sampled on the feature signal data with a length of S at a sliding step size, and normalization is performed; N ​ For the dynamic output torque and rotational speed pulsation characteristic signal data, it is divided into start-stop process signals and stable operation pulsation signals. For the start-stop process signal data, the processing method of dividing by the maximum torque of the stable operation without losing magnetism is adopted; for the stable operation pulsation characteristic signal data, the same processing method as the body vibration is adopted, that is, a window with a length of L is sampled overlapped on the characteristic signal data with a length of S at a sliding step size of N and normalized; For the phase current characteristic signal data, perform normalization processing on the stable operation signal; For the load-end temperature rise change characteristic signal data, adopt the processing method of dividing by the temperature rise in the non-demagnetization stable operation state; The multi-scale feature extraction module uses convolution kernels covering large, medium, and small three feature scales and is sensitive to the information of each stage frequency; The adoption of stacked integration for multi-scale feature extraction means processing the extracted features in a stacked manner instead of an additive manner to maximize the preservation of the feature information extracted by each multi-scale module; The introduced depth - separable method is used to establish a multi - sensor information fusion framework, specifically: Feature fusion adopts the method of point - by - point convolution, and after point - by - point convolution, global pooling is adopted to reduce network parameters; Before softmax classification, cancel Flatten the way of connecting the fully - connected layer, and instead use global average pooling to further reduce the number of parameters; The channel attention mechanism is specifically as follows: After the feature parameters are extracted by the multi-scale module, the stacked features are respectively subjected to one-dimensional global average pooling and one-dimensional global max pooling to integrate the global feature information; After adjusting the dimensions, the results of the two poolings are respectively subjected to one-dimensional convolution without bias terms to capture the mutual relationship between the features of each channel. Then, the results of the two-way convolution are added and passed through an activation function to generate adaptive weights for each channel. The obtained adaptive weights are multiplied by the stacked layer features and enter the feature fusion link.

2. The multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution according to claim 1, wherein: According to the division of the loss-of-excitation degree interval, the output type of the network is determined. Specifically, the loss-of-excitation degree interval is divided into three stages, namely, mild loss of excitation, moderate loss of excitation, and severe loss of excitation stages. Based on this, the output type of the network is determined to be 4 diagnostic labels: no loss of excitation, mild loss of excitation, moderate loss of excitation, and severe loss of excitation. The processed characteristic signal data are assigned the above 4 diagnostic labels, and the diagnostic labels are one-hot encoded.

3. A multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution according to claim 1, characterized in that: The network configuration specifically uses the Adaptive Moment Estimation optimizer Adma method, which can adaptively adjust the learning rate of each parameter. Combined with the cross-entropy as the loss function, it can accelerate the convergence speed when the model performance is poor, making the training data distribution closer to the real data distribution. The network training refers to dividing the processed feature signal data into a training set and a test set according to a ratio, randomly sorting the training data and the test data respectively, feeding them into the network, and injecting noise during training to improve the network's anti-noise ability; To ensure the consistency of the dimensional features of the training parameters, the same length of data is used for training in each channel, and all-zero padding is adopted; the length of the feature signal data N takes values with the property independent of the sampling initial phase, including at least one signal cycle, and is based on the lowest frequency feature in the fault space signal under the same sampling frequency.

4. A multi-sensor traction machine demagnetization monitoring method based on one-dimensional multi-scale convolution according to any one of claims 1 to 3, characterized in that: Considering that the elevator is a special equipment for carrying people, to ensure the safe and reliable operation of the system, a diagnostic method for the periphery of the system is adopted to extract the features of the signal and diagnose the demagnetization of the traction machine without disturbing the operation of the traction system.

Images