Laminating equipment fault diagnosis method based on multi-physical-quantity feature fusion

Abstract

The invention discloses a press fit equipment fault diagnosis method based on multi-physical-quantity feature fusion. The method is characterized by comprising the four steps of A, multi-source data synchronous acquisition and preprocessing, B, multi-dimensional feature extraction, C, fault diagnosis model construction and fault diagnosis, and D, hierarchical diagnosis. According to the method, the strong correlation between current and torque can be effectively decoupled, a comprehensive and accurate motor state characteristic space is constructed, and high-precision and real-time diagnosis of press-fit equipment faults is realized.

Description

Technical Field

[0001] This invention relates to the field of automobile manufacturing, and in particular to a fault diagnosis method for pressing equipment based on the fusion of multiple physical quantity characteristics. Background Technology

[0002] In the automobile manufacturing process, pressing equipment is used to press-fit automotive body panels. The stability of the pressing equipment's operation plays a decisive role in the quality of automotive products and the efficiency of the production line. In existing technologies, fault diagnosis of pressing equipment mainly relies on vibration detection or detection methods based on motor current or torque. Vibration detection requires additional sensors, increasing equipment cost and maintenance complexity. While detection methods based on motor current or torque are easy to integrate, they are difficult to accurately distinguish between electrical faults (such as motor phase loss or winding short circuits) and mechanical faults (such as bearing wear or rotor eccentricity) based on changes in a single physical quantity. The diagnostic results are often vague and cannot provide a reliable basis for equipment maintenance.

[0003] In existing technologies, the integration of key operating parameters such as voltage and speed is often neglected when constructing the motor state feature space. Due to the lack of these key parameters, the feature space is not fully constructed and cannot comprehensively reflect the actual operating state, resulting in low diagnostic accuracy and poor adaptability in complex production environments with frequently changing operating conditions, making it difficult to meet the high requirements of modern automobile manufacturing for equipment reliability and efficiency.

[0004] Therefore, a method or apparatus is needed to solve the above problems. Summary of the Invention

[0005] This invention aims to address the aforementioned shortcomings of existing technologies by proposing a multi-physical quantity feature fusion-based fault diagnosis method for pressing equipment. This method effectively decouples the strong correlation between current and torque, constructs a comprehensive and accurate motor state feature space, and enables high-precision, real-time diagnosis of pressing equipment faults.

[0006] The technical solution of this invention is: a fault diagnosis method for pressing equipment based on multi-physical quantity feature fusion, characterized in that: the method includes four major steps: A. synchronous acquisition and preprocessing of multi-source data, B. multi-dimensional feature extraction, C. construction of fault diagnosis model and fault diagnosis, and D. hierarchical diagnosis. The operation process of the multi-source data synchronous acquisition and preprocessing step A is as follows: A1. Using three-phase current sensors, torque sensors, voltage sensors, and encoders, synchronous high-speed data acquisition of motor current, torque, three-phase voltage, and speed is performed at a sampling frequency of no less than 10kHz. Simultaneously, the pressing pressure and displacement data of the pressing equipment are also acquired synchronously. A2. Wavelet transform is used to filter the acquired current and torque data. Specifically, a 5-level decomposition is performed using the db4 wavelet basis, and electromagnetic interference in the industrial environment is removed using a soft thresholding denoising method. The threshold calculation formula in the soft thresholding denoising method is as follows: Where σ is the noise standard deviation and N is the signal length. A3. Perform precise time alignment on the data collected in step A1, controlling the time synchronization error within ±1ms to ensure consistency of different physical quantity data across time scales, thereby constructing a spatiotemporal feature matrix. The B-dimensional feature extraction step includes B1 time-domain feature extraction, B2 frequency-domain and time-domain feature extraction, and B3 fusion feature construction. The B1 time-domain feature extraction step quickly identifies abnormal changes in the operating state of equipment by analyzing the instantaneous or statistical characteristics of physical quantities in the time dimension, providing direct criteria for preliminary fault screening. The specific operation steps are as follows: B1.1 Based on measured three-phase voltage values, the ratio of the negative-sequence component to the positive-sequence component is calculated to reflect the degree of three-phase voltage imbalance. When this ratio exceeds 5%, a motor phase loss fault warning is triggered. B1.2 Performs linear fitting on the speed and torque data within the same process and solves for the slope value. When the fluctuation of the obtained slope value is greater than 5% compared with the slope value during normal operation, a mechanical jamming fault warning is triggered. B1.3 Calculate the ratio of the standard deviation to the mean of the peak current over any 24 hours, and quantify the degree of peak current fluctuation. When the fluctuation exceeds 8%, trigger an abnormal peak current fluctuation warning. B1.4 Calculate the ratio of the measured pressure to the average pressure. When the fluctuation exceeds 20%, trigger an abnormal pressure fluctuation warning. B1.5 Calculate the difference between the maximum and minimum torque values ​​to determine the torque fluctuation range. When the torque fluctuation range exceeds 10% of the rated torque, trigger an abnormal torque fluctuation range warning. B1.6 Calculate the absolute value of the difference between the measured displacement value and the preset process displacement value, and statistically analyze the maximum and average deviations. If the maximum deviation is greater than 0.3 mm and the average deviation exceeds ±0.1 mm, an abnormal transmission clearance warning will be triggered. The extraction steps of the B2 frequency domain and time-frequency domain features reveal the harmonic components and time-frequency distribution patterns of physical quantities through frequency domain analysis, accurately locate the characteristic frequencies associated with faults, and distinguish between electrical and mechanical fault types. The specific operation steps are as follows: B2.1 performs a Fourier transform on the voltage signal, extracts the amplitude of the third harmonic component, and compares it with the fundamental amplitude. When the ratio of the two is greater than 15%, a short-circuit warning for the motor winding is triggered. B2.2 decomposes the speed signal using Fourier transform and extracts the amplitude corresponding to twice the rotational frequency. When this amplitude reaches more than three times the amplitude under normal conditions, a rotor eccentricity warning is triggered. B2.3 performs a Fourier transform on the current signal to extract the amplitude of the 5th harmonic component. When the ratio of each harmonic to the fundamental frequency is greater than 8%, an electrical malfunction warning is triggered. B2.4 Identify the peak frequency points of voltage and current signals in the spectrum graph, record their frequency values ​​and corresponding amplitudes. If the frequency value is found to fall within ±30% of the equipment's structural frequency, trigger an abnormal frequency source warning. B2.5 quantifies the distribution of signal energy in the frequency domain through power spectral density analysis, extracts the power spectral density value corresponding to the peak frequency, reflecting the energy intensity of that frequency component. When the peak power spectral density value at the corresponding frequency increases by more than 10 dB compared to the normal state, an abnormal power spectral density value warning is triggered. B2.6 performs 3D short-time Fourier transform on current, voltage, and torque signals, setting the window size to 1024 points and the overlap rate to 75%. It constructs three-dimensional time-frequency-physical quantity (current / voltage / torque) spectra for each signal and extracts the spatiotemporal distribution characteristics of feature frequencies from these spectra. The construction steps of the B3 fusion feature can integrate the correlation of multiple physical quantities, theoretical model deviations, and chaotic characteristics, breaking through the limitations of a single physical quantity and improving the identification and diagnostic accuracy of fault features. The specific operation steps are as follows: B3.1 uses Hilbert transform to extract the instantaneous phase of current, voltage, and torque, calculates the standard deviation of the phase difference, and triggers an electro-mechanical coupling fault warning when the standard deviation is found to be greater than 15°. B3.2 calculates the instantaneous power time series based on voltage and current signals, and uses the box-counting method to calculate the fractal dimension. The phase space reconstruction parameters are: a delay time of 10ms and an embedding dimension of 7. Under normal operating conditions, the fractal dimension ranges from 1.8 to 2.2. If the fractal dimension deviates from the above range, a power fractal dimension anomaly warning is triggered. B3.3 performs frequency domain analysis on the torque signal, extracting the energy value in the 2000Hz frequency band. When the obtained energy value is found to be more than three times the normal value, a high-frequency energy anomaly warning for torque is triggered. The construction and fault diagnosis steps of the C fault diagnosis model are as follows: C1 constructs a bidirectional LSTM-GAN combined model, extracting key features from physical quantities such as current, torque, voltage, and speed. Specifically, it selects voltage imbalance, speed-torque slope, current peak fluctuation coefficient, and speed twice the rotational frequency to form a hybrid input vector. The generator uses a bidirectional LSTM network, with the input dimension determined based on the hybrid input vector dimension. The hidden layer has 128 nodes, and the output dimension is consistent with the input dimension. The discriminator uses a fully connected network with the same input and output dimensions as the generator. It has 64 hidden layer nodes and an output dimension of 1. The training parameters are set as follows: batch_size=32, epoch=100, Adam optimizer, and learning rate=0.001. The generator output data is compared with the actual acquired data. When the difference between the output data and the actual acquired data is found to be greater than 3 times the standard deviation (3σ1), a corresponding anomaly warning is triggered. These anomaly warnings include abnormal current peak fluctuations, abnormal voltage imbalance, abnormal speed-torque slope, abnormal speed twice the rotational frequency, abnormal temperature, and abnormal physical quantity trends. σ1 represents the standard deviation of the normal data error. C2 uses a 3D-CNN model to perform depth processing on the 3D temporal spectrogram constructed in step B2.6. The 3D-CNN model comprises four convolutional layers and two pooling layers. The convolutional layers have 3×3×3 kernels, with the number of kernels being 32, 64, 128, and 256 respectively. It automatically extracts spatial and temporal features through multi-layer convolutional operations. The training parameters for the 3D-CNN model are: batch_size=16, epoch=80, and learning rate=0.0005. Using a three-dimensional time-spectrum graph as input, the features extracted through multi-layer convolutional operations are matched against a preset fault feature library. When the matching degree is ≥90%, a corresponding fault confirmation warning is triggered. These fault confirmation warnings include motor phase loss fault warnings, winding fault warnings, rotor eccentricity fault warnings, and bearing fault warnings, achieving accurate identification of fault types. The operation process of the D-layer diagnostic step is as follows: First, anomaly detection is performed using a bidirectional LSTM model for real-time monitoring. An alert is triggered when anomalies are detected for three consecutive process cycles. Next, fault type identification is performed, using a 3D-CNN model to distinguish between electrical and mechanical faults. If, as in step B3.3, the energy value in the 2000Hz frequency band is greater than 1 times but less than 2 times the normal value, then it indicates minor gear wear. If the energy value in the 2000Hz frequency band is more than twice but less than three times the normal value, it indicates moderate gear wear. If the energy value in the 2000Hz frequency band is more than three times the normal value, it indicates severe gear wear. If, as in step B1.6, the maximum deviation is between 0.3-0.5 mm and the average deviation is between 0.1-0.2 mm, then it indicates slight leadscrew wear. If the maximum deviation is between 0.5-1.0 mm and the average deviation is between 0.2-0.3 mm, then it indicates moderate lead screw wear. If the maximum deviation is greater than 1.0 mm and the average deviation is greater than 0.3 mm, it indicates severe lead screw wear.

[0007] Compared with the prior art, the present invention has the following advantages: The fault diagnosis method for pressing equipment disclosed in this invention, by simultaneously collecting multi-source data such as current, torque, voltage, speed and process parameters, and extracting multi-dimensional features, constructs a comprehensive motor state feature space, which makes up for the defect of incomplete feature space construction in the prior art and can fully reflect the real operating status of the equipment. Meanwhile, because it adopts a multi-physical quantity feature fusion method, it can effectively decouple the strong correlation between current and torque, solve the problem that a single physical quantity is difficult to accurately distinguish between electrical faults and mechanical faults, and improve the accuracy of fault diagnosis.

[0008] Furthermore, in the construction of the fault diagnosis model, various features obtained from feature extraction are fully utilized. Anomaly detection is performed through a bidirectional LSTM-GAN combined model, and the accurate identification of fault types is achieved by combining 3D-CNN and cross-modal attention mechanism, thereby improving the model's adaptability to complex working conditions and diagnostic accuracy.

[0009] In conclusion, this diagnostic method has many advantages and is particularly suitable for widespread application in this field, with a very broad market prospect.

[0010] The specific embodiments of the present invention will be described below.

[0011] A fault diagnosis method for pressing equipment based on multi-physical quantity feature fusion is characterized by comprising four main steps: A) synchronous acquisition and preprocessing of multi-source data, B) multi-dimensional feature extraction, C) construction of fault diagnosis model and fault diagnosis, and D) hierarchical diagnosis. The operation process of the multi-source data synchronous acquisition and preprocessing step A is as follows: First, a three-phase current sensor and a torque sensor are configured at the motor output end to collect current and torque data during motor operation. A three-phase voltage sensor is deployed at the motor power input end to acquire three-phase voltage signals. An encoder is installed at the rotating shaft end of the pressing actuator to collect speed data. Simultaneously, pressure sensors and displacement sensors are planned and deployed at the pressing mold location to collect pressing pressure and displacement data, respectively. All sensors must meet industrial-grade accuracy requirements to ensure the accuracy of data acquisition.

[0012] Then, a data acquisition system with synchronous acquisition function is adopted, with a sampling frequency of no less than 10kHz to ensure that the dynamic characteristics of the equipment during operation can be captured. The acquisition cycle is consistent with the process and cycle of the pressing equipment to ensure complete acquisition of equipment operation data in each process. The types of data acquired include multi-source data such as three-phase current, torque, three-phase voltage, speed, pressure, and displacement.

[0013] Then, the collected current and torque data are subjected to a 5-level wavelet transform using the db4 wavelet basis. Electromagnetic interference in the industrial environment is removed using a soft thresholding denoising method, preserving the effective signal components. The threshold calculation formula in the soft thresholding denoising method is as follows: Where σ is the noise standard deviation and N is the signal length. Based on the process characteristics of the pressing process, the data of each cycle is segmented according to the process. Using the encoder timestamp as a reference, the multi-dimensional data is precisely time-aligned, and the time synchronization error is controlled within ±1ms to ensure the consistency of different physical quantity data on the time scale, thereby constructing a spatiotemporal feature matrix (dimension: time × physical quantity type).

[0014] The B-dimensional feature extraction step includes B1 time-domain feature extraction, B2 frequency-domain and time-domain feature extraction, and B3 fusion feature construction. The B1 time-domain feature extraction step quickly identifies abnormal changes in the operating state of equipment by analyzing the instantaneous or statistical characteristics of physical quantities in the time dimension, providing direct criteria for preliminary fault screening. The specific operation steps are as follows: B1.1 Based on measured three-phase voltage values, the negative-sequence and positive-sequence components are separated using the symmetrical component method, and the voltage unbalance is calculated as the ratio of the negative-sequence component to the positive-sequence component. This feature can be used to monitor whether a motor has a phase loss fault. When a phase is lost, the voltage unbalance increases significantly. When this ratio exceeds 5%, a phase loss fault warning for the motor is triggered. B1.2 Within the same process, linear fitting is performed on the speed and torque data, and the slope of the fitted line is calculated using the least squares method. This slope reflects the dynamic relationship between speed and torque. When a mechanical jamming fault occurs, the speed-torque slope will change significantly. When the fluctuation of the calculated slope value compared with the slope value during normal operation is greater than 5%, a mechanical jamming fault warning is triggered. B1.3 Calculate the ratio of the standard deviation to the mean of the peak current over any 24-hour period, and obtain the peak current fluctuation coefficient. This coefficient quantifies the degree of fluctuation in the peak current. During electrical faults, the peak current fluctuation coefficient typically exceeds the normal range. When the fluctuation exceeds 8%, an abnormal peak current fluctuation warning is triggered. B1.4 Calculate the absolute value of the difference between the measured pressure value at each sampling time and the average pressure within that process segment to assess the stability of the pressing pressure. Excessive pressure deviation may lead to product quality abnormalities. When the fluctuation exceeds 20%, an abnormal pressing pressure fluctuation warning is triggered. B1.5 From the collected torque data, identify the maximum and minimum torque values, calculate the difference between them, and obtain the torque fluctuation amplitude. This indicator reflects the smoothness of torque output; when a fault occurs in the mechanical transmission system, the torque fluctuation amplitude will increase. When the torque fluctuation amplitude exceeds 10% of the rated torque, an abnormal torque fluctuation amplitude warning is triggered. B1.6 Calculate the absolute value of the difference between the measured displacement value and the preset process displacement value, and statistically analyze the maximum or average deviation within this process segment. Displacement accuracy deviation can assess the accuracy of the die displacement control. Mechanical faults such as lead screw wear and excessive transmission clearance will cause this deviation to increase. If the maximum deviation is greater than 0.3mm and the average deviation exceeds ±0.1mm, an alarm for abnormal lead screw wear and transmission clearance will be triggered. The B2 frequency domain and time domain feature extraction steps reveal the harmonic components and time-frequency distribution patterns of physical quantities through frequency domain analysis, accurately locate the characteristic frequencies associated with faults, and distinguish between electrical and mechanical fault types. The specific operation steps are as follows: B2.1 performs a Fourier transform on the voltage signal, converting the time-domain signal to the frequency domain, extracting the amplitude of the third harmonic component, and comparing it with the fundamental amplitude. When a short-circuit fault occurs in the motor windings, the ratio of the third harmonic amplitude to the fundamental amplitude will increase significantly. When this ratio is found to be greater than 15%, a short-circuit warning for the motor windings is triggered. B2.2 Calculate the rotational frequency based on the rotational speed data, perform a Fourier transform on the rotational speed signal, extract the amplitude corresponding to twice the rotational frequency, and analyze its changing trend. Rotor eccentricity faults will cause an abnormal increase in the amplitude of the rotational speed component that is twice the rotational frequency. When this amplitude reaches more than three times the amplitude under normal conditions, a rotor eccentricity warning is triggered. B2.3 performs a Fourier transform on the current signal to extract the amplitude of the 5th harmonic components. During normal operation, the ratio of each harmonic to the fundamental frequency is relatively small. However, when an electrical abnormality occurs, the amplitude of these harmonics increases significantly. When the ratio of each harmonic to the fundamental frequency exceeds 8%, an electrical abnormality warning is triggered. B2.4 Identify the peak frequency points of voltage and current signals in the spectrum graph, record their frequency values ​​and corresponding amplitudes, and analyze whether the peak frequency matches the structural frequency of the equipment (such as the natural frequency of the leadscrew) to locate the abnormal frequency source and help distinguish between electrical interference and mechanical vibration faults. If the frequency value is found to fall within ±30% of the equipment's own structural frequency, an abnormal frequency source warning is triggered. B2.5 quantifies the signal energy distribution in the frequency domain through power spectral density analysis, extracting the power spectral density value corresponding to the peak frequency to reflect the energy intensity of that frequency component. Under fault conditions, the peak power spectral density at the corresponding fault characteristic frequency will significantly increase. When the peak power spectral density at the corresponding frequency increases by more than 10 dB compared to the normal state, an abnormal power spectral density warning is triggered. B2.6 performs 3D short-time Fourier transform on current, voltage, and torque signals, setting the window size to 1024 points and the overlap rate to 75%, constructing a three-dimensional time-frequency-physical quantity spectrogram, and extracting the spatiotemporal distribution characteristics of feature frequencies in the three-dimensional time-frequency spectrogram. The construction steps of the B3 fusion feature can integrate the correlation of multiple physical quantities, theoretical model deviations, and chaotic characteristics, breaking through the limitations of a single physical quantity and improving the identification and diagnostic accuracy of fault features. The specific operation steps are as follows: B3.1 extracts the instantaneous phase of current, voltage, and torque using Hilbert transform and calculates the standard deviation of the phase difference. This feature can reflect abnormal phase relationships between multiple physical quantities. Electro-mechanical coupling faults will cause an increase in the standard deviation of the phase difference. When the standard deviation is found to be greater than 15°, an electro-mechanical coupling fault abnormality warning is triggered. B3.2 calculates the instantaneous power time series based on voltage and current signals. Using the box-dimensional method or correlation dimension method from fractal theory, the power time series is mapped to phase space, constructing a reconstructed phase space matrix. By statistically analyzing the number of "boxes" required to cover the phase space at different scales, the power fractal dimension is fitted. The phase space reconstruction parameters are: a delay time of 10ms and an embedding dimension of 7. Under normal operating conditions, the fractal dimension ranges from 1.8 to 2.2. If the fractal dimension deviates from this range, a power fractal dimension anomaly warning is triggered. B3.3 performs frequency domain analysis on the torque signal, extracts the energy value in the 2000Hz frequency band, and compares it with the energy value under normal conditions. When gear wear faults occur, the energy value in this frequency band will increase significantly. When the obtained energy value is found to be more than three times that of the normal state, a high-frequency energy anomaly warning for torque is triggered. The construction and fault diagnosis steps of the C fault diagnosis model are as follows: C1 constructs a bidirectional LSTM model, extracting key features from physical quantities such as current, torque, voltage, speed, and temperature. Specifically, it selects voltage imbalance, speed-torque slope, current peak fluctuation coefficient, speed twice the rotational frequency, and temperature to form a hybrid input vector. The generator uses a bidirectional LSTM network, with the input dimension determined by the hybrid input vector dimension. The hidden layer has 128 nodes, and the output dimension is consistent with the input dimension. The discriminator uses a fully connected network with the same input and output dimensions as the generator. It has 64 hidden layer nodes and an output dimension of 1. The training parameters are set as follows: batch_size=32, epoch=100, Adam optimizer, and learning rate=0.001. The generator output data is compared with the actual acquired data. When the difference between the output data and the actual acquired data is found to be greater than 3 times the standard deviation (3σ1), a corresponding anomaly warning is triggered. These anomaly warnings include abnormal current peak fluctuations, abnormal voltage imbalance, abnormal speed-torque slope, abnormal speed twice the rotational frequency, abnormal temperature, and abnormal physical quantity trends. σ1 represents the standard deviation of the normal data error. C2 uses a 3D-CNN model to perform depth processing on the 3D temporal spectrogram constructed in step B2.6. The 3D-CNN model comprises four convolutional layers and two pooling layers. The convolutional layers have 3×3×3 kernels, with the number of kernels being 32, 64, 128, and 256 respectively. It automatically extracts spatial and temporal features through multi-layer convolutional operations. The training parameters for the 3D-CNN model are: batch_size=16, epoch=80, and learning rate=0.0005. Using a three-dimensional time-spectrum graph as input, the features extracted through multi-layer convolutional operations are matched against a preset fault feature library. When the matching degree is ≥90%, a corresponding fault confirmation warning is triggered. These fault confirmation warnings include motor phase loss fault warnings, winding fault warnings, rotor eccentricity fault warnings, bearing fault warnings, and mechanical jamming fault warnings, achieving accurate identification of fault types. The operation process of the D-layer diagnostic step is as follows: First, anomaly detection is performed using a bidirectional LSTM model for real-time monitoring. An alert is triggered when anomalies are detected for three consecutive process cycles. Next, fault type identification is performed, using a 3D-CNN model to distinguish between electrical and mechanical faults. If, as in step B3.3, the energy value in the 2000Hz frequency band is greater than 1 times but less than 2 times the normal value, then it indicates minor gear wear. If the energy value in the 2000Hz frequency band is more than twice but less than three times the normal value, it indicates moderate gear wear. If the energy value in the 2000Hz frequency band is more than three times the normal value, it indicates severe gear wear. If, as in step B1.6, the maximum deviation is between 0.3-0.5 mm and the average deviation is between 0.1-0.2 mm, then it indicates slight leadscrew wear. If the maximum deviation is between 0.5-1.0 mm and the average deviation is between 0.2-0.3 mm, then it indicates moderate lead screw wear. If the maximum deviation is greater than 1.0 mm and the average deviation is greater than 0.3 mm, it indicates severe lead screw wear.

Claims

1. A fault diagnosis method for pressing equipment based on the fusion of multiple physical quantity features, characterized in that: The method comprises four main steps: A) synchronous acquisition and preprocessing of multi-source data; B) multi-dimensional feature extraction; C) construction of fault diagnosis models and fault diagnosis; and D) hierarchical diagnosis. The operation process of the multi-source data synchronous acquisition and preprocessing step A is as follows: A1. Using three-phase current sensors, torque sensors, voltage sensors, and encoders, synchronous high-speed data acquisition of motor current, torque, three-phase voltage, and speed is performed at a sampling frequency of no less than 10kHz. Simultaneously, the pressing pressure and displacement data of the pressing equipment are also acquired synchronously. A2. Wavelet transform is used to filter the acquired current and torque data. Specifically, a 5-level decomposition is performed using the db4 wavelet basis, and electromagnetic interference in the industrial environment is removed using a soft thresholding denoising method. The threshold calculation formula in the soft thresholding denoising method is as follows: Where σ is the noise standard deviation and N is the signal length. A3. Perform precise time alignment on the data collected in step A1, controlling the time synchronization error within ±1ms to ensure consistency of different physical quantity data across time scales, thereby constructing a spatiotemporal feature matrix. The B-dimensional feature extraction step includes B1 time-domain feature extraction, B2 frequency-domain and time-domain feature extraction, and B3 fusion feature construction. The B1 time-domain feature extraction step quickly identifies abnormal changes in the operating state of equipment by analyzing the instantaneous or statistical characteristics of physical quantities in the time dimension, providing direct criteria for preliminary fault screening. The specific operation steps are as follows: B1.1 Based on measured three-phase voltage values, the ratio of the negative-sequence component to the positive-sequence component is calculated to reflect the degree of three-phase voltage imbalance. When this ratio exceeds 5%, a motor phase loss fault warning is triggered. B1.2 Performs linear fitting on the speed and torque data within the same process and solves for the slope value. When the fluctuation of the obtained slope value is greater than 5% compared with the slope value during normal operation, a mechanical jamming fault warning is triggered. B1.3 Calculate the ratio of the standard deviation to the mean of the peak current over any 24 hours, and quantify the degree of peak current fluctuation. When the fluctuation exceeds 8%, trigger an abnormal peak current fluctuation warning. B1.4 Calculate the ratio of the measured pressure to the average pressure. When the fluctuation exceeds 20%, trigger an abnormal pressure fluctuation warning. B1.5 Calculate the difference between the maximum and minimum torque values ​​to determine the torque fluctuation range. When the torque fluctuation range exceeds 10% of the rated torque, trigger an abnormal torque fluctuation range warning. B1.6 Calculate the absolute value of the difference between the measured displacement value and the preset process displacement value, and statistically analyze the maximum and average deviations. If the maximum deviation is greater than 0.3 mm and the average deviation exceeds ±0.1 mm, an abnormal transmission clearance warning will be triggered. The extraction steps of the B2 frequency domain and time-frequency domain features reveal the harmonic components and time-frequency distribution patterns of physical quantities through frequency domain analysis, accurately locate the characteristic frequencies associated with faults, and distinguish between electrical and mechanical fault types. The specific operation steps are as follows: B2.1 performs a Fourier transform on the voltage signal, extracts the amplitude of the third harmonic component, and compares it with the fundamental amplitude. When the ratio of the two is greater than 15%, a short-circuit warning for the motor winding is triggered. B2.2 decomposes the speed signal using Fourier transform and extracts the amplitude corresponding to twice the rotational frequency. When this amplitude reaches more than three times the amplitude under normal conditions, a rotor eccentricity warning is triggered. B2.3 performs a Fourier transform on the current signal to extract the amplitude of the 5th harmonic component. When the ratio of each harmonic to the fundamental frequency is greater than 8%, an electrical malfunction warning is triggered. B2.4 Identify the peak frequency points of voltage and current signals in the spectrum graph, record their frequency values ​​and corresponding amplitudes. If the frequency value is found to fall within ±30% of the equipment's structural frequency, trigger an abnormal frequency source warning. B2.5 quantifies the distribution of signal energy in the frequency domain through power spectral density analysis, extracts the power spectral density value corresponding to the peak frequency, reflecting the energy intensity of that frequency component. When the peak power spectral density value at the corresponding frequency increases by more than 10 dB compared to the normal state, an abnormal power spectral density value warning is triggered. B2.6 performs 3D short-time Fourier transform on current, voltage, and torque signals, setting the window size to 1024 points and the overlap rate to 75%. It constructs three-dimensional time-frequency-physical quantity (current / voltage / torque) spectra for each signal and extracts the spatiotemporal distribution characteristics of feature frequencies from these spectra. The construction steps of the B3 fusion feature can integrate the correlation of multiple physical quantities, theoretical model deviations, and chaotic characteristics, breaking through the limitations of a single physical quantity and improving the identification and diagnostic accuracy of fault features. The specific operation steps are as follows: B3.1 uses Hilbert transform to extract the instantaneous phase of current, voltage, and torque, calculates the standard deviation of the phase difference, and triggers an electro-mechanical coupling fault warning when the standard deviation is found to be greater than 15°. B3.2 calculates the instantaneous power time series based on voltage and current signals, and uses the box-counting method to calculate the fractal dimension. The phase space reconstruction parameters are: a delay time of 10ms and an embedding dimension of 7. Under normal operating conditions, the fractal dimension ranges from 1.8 to 2.

2. If the fractal dimension deviates from the above range, a power fractal dimension anomaly warning is triggered. B3.3 performs frequency domain analysis on the torque signal, extracting the energy value in the 2000Hz frequency band. When the obtained energy value is found to be more than three times the normal value, a high-frequency energy anomaly warning for torque is triggered. The construction and fault diagnosis steps of the C fault diagnosis model are as follows: C1 constructs a bidirectional LSTM-GAN combined model, extracting key features from physical quantities such as current, torque, voltage, and speed. Specifically, it selects voltage imbalance, speed-torque slope, current peak fluctuation coefficient, and speed twice the rotational frequency to form a hybrid input vector. The generator uses a bidirectional LSTM network, with the input dimension determined based on the hybrid input vector dimension. The hidden layer has 128 nodes, and the output dimension is consistent with the input dimension. The discriminator uses a fully connected network with the same input and output dimensions as the generator. The hidden layer has 64 nodes, and the output dimension is 1. Training parameters are set as follows: batch_size=32, epoch=100, the optimizer is Adam, and the learning rate is 0.

001. The generator output data is compared with the actual acquired data. When the difference between the output data and the actual acquired data is found to be greater than 3 times the standard deviation (3σ1), a corresponding anomaly warning is triggered. These anomaly warnings include abnormal current peak fluctuations, abnormal voltage imbalance, abnormal speed-torque slope, abnormal speed twice the rotational frequency, abnormal temperature, and abnormal physical quantity trends. σ1 represents the standard deviation of the normal data error. C2 uses a 3D-CNN model to perform depth processing on the 3D temporal spectrogram constructed in step B2.

6. The 3D-CNN model comprises four convolutional layers and two pooling layers. The convolutional layers have 3×3×3 kernels, with the number of kernels being 32, 64, 128, and 256 respectively. It automatically extracts spatial and temporal features through multi-layer convolutional operations. The training parameters for the 3D-CNN model are: batch_size=16, epoch=80, and learning rate=0.0005. Using a three-dimensional time-spectrum graph as input, the features extracted through multi-layer convolutional operations are matched against a preset fault feature library. When the matching degree is ≥90%, a corresponding fault confirmation warning is triggered. These fault confirmation warnings include motor phase loss fault warnings, winding fault warnings, rotor eccentricity fault warnings, and bearing fault warnings, achieving accurate identification of fault types. The operation process of the D-layer diagnostic step is as follows: First, anomaly detection is performed using a bidirectional LSTM model for real-time monitoring. An alert is triggered when anomalies are detected for three consecutive process cycles. Next, fault type identification is performed, using a 3D-CNN model to distinguish between electrical and mechanical faults. If, as in step B3.3, the energy value in the 2000Hz frequency band is greater than 1 times but less than 2 times the normal value, then it indicates minor gear wear. If the energy value in the 2000Hz frequency band is more than twice but less than three times the normal value, it indicates moderate gear wear. If the energy value in the 2000Hz frequency band is more than three times the normal value, it indicates severe gear wear. If, as in step B1.6, the maximum deviation is between 0.3-0.5 mm and the average deviation is between 0.1-0.2 mm, then it indicates slight leadscrew wear. If the maximum deviation is between 0.5-1.0 mm and the average deviation is between 0.2-0.3 mm, then it indicates moderate lead screw wear. If the maximum deviation is greater than 1.0 mm and the average deviation is greater than 0.3 mm, it indicates severe lead screw wear.