Motor installation eccentric orientation error correction method based on recurrent neural network

Abstract

The invention discloses a motor installation eccentric orientation error correction method based on a recurrent neural network, and relates to the technical field of motor installation eccentric orientation error correction, and the method comprises the following specific steps: initially installing a motor, and collecting a motor current signal; carrying out eccentric azimuth diagnosis model training; inputting the preprocessed signal into an eccentric azimuth diagnosis model for azimuth diagnosis; the output label judges whether an eccentric fault exists or not; the main eccentric direction is judged, the motor is installed and adjusted in the reverse direction, and calibration verification is conducted through a laser centering instrument; time window sliding grouping is carried out on collected eccentric fault data, time domain features, frequency domain features and wavelet packet energy reconstruction of the data are extracted, and fusion features are used as input signals of a neural network model; an improved genetic optimization algorithm is adopted to seek an optimal combination for parameters such as the number of network layers and the number of units of each layer, the accuracy rate of eccentric orientation diagnosis is improved to 95% or above, and the anti-interference capability to noise is improved.

Description

Technical Field

[0001] This invention relates to the field of motor installation eccentricity correction technology, specifically a motor installation eccentricity correction method based on recurrent neural networks. Background Technology

[0002] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network. Unlike typical feedforward neural networks, LSTMs can analyze inputs using time series data. Genetic Algorithms (GA) are optimization search algorithms inspired by natural selection. These algorithms mimic biological evolution, iteratively improving a "population" of candidate solutions through selection, crossover, and mutation. By simulating natural selection, they explore different regions of the parameter space to find the global optimum or a near-global optimum, rather than getting trapped in local optima. However, LSTM models have a large number of hyperparameters, which significantly impact network performance. Manually adjusting hyperparameters is a complex and time-consuming process, especially with complex datasets and large networks. Eccentricity causes motors to vibrate at specific frequencies. The choice of diagnostic method depends on various factors, such as the severity of the fault, cost, and available diagnostic equipment. Traditional diagnostic methods cannot identify the location of eccentricity faults in motors.

[0003] In the actual operating environment of motors, various noise sources are unavoidable. These noises may come from the motor's own mechanical vibration, electrical noise, or the background noise of the external environment such as the operation of industrial equipment. These noise environments pose challenges to the performance monitoring and fault diagnosis of motors. Summary of the Invention

[0004] The purpose of this invention is to provide a method for correcting motor installation eccentricity based on a recurrent neural network. By performing time window sliding grouping on the collected eccentricity fault data, the time domain features, frequency domain features, and wavelet packet energy reconstruction of the data are extracted, and the fused features are used as the input signal of the neural network model. An improved genetic optimization algorithm is used to seek the optimal combination of parameters such as the number of network layers and the number of units per layer, thereby improving the accuracy of eccentricity diagnosis to over 95% and enhancing the anti-interference ability against noise, thus solving the problems in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for correcting motor mounting eccentricity based on recurrent neural networks includes the following specific steps: Step 1: Initial installation of the motor and acquisition of motor current signal; Step 2: Train the eccentric orientation diagnosis model; Step 3: Input the preprocessed signal into the eccentric azimuth diagnostic model for azimuth diagnosis; Step 4: Output the label to determine if there is an eccentricity fault; Step 5: Determine the main eccentricity and install and adjust the motor in the opposite direction, then verify the calibration using a laser alignment instrument; After calibration, repeat steps one through four. When there is no eccentricity fault, the error correction process terminates; when there is still an eccentricity fault, repeat step five.

[0006] Furthermore, the implementation method of step one is as follows: A permanent magnet synchronous motor drives a load through a flexible coupling. The load is loaded by a magnetic powder brake, which enables data acquisition under different loading conditions. The motor is mounted on a three-dimensional support on the test bench. The eccentricity and degree of the motor can be adjusted through the three-dimensional support. The laser alignment instrument is used to quantify and calibrate the eccentricity. The motor's current signal, torque signal, and speed data signal are acquired by the motor incremental encoder and driver. The driver is a GSK six-axis servo driver. When the motor's torque and speed are stable, the motor's current signal is collected, and the fault location is identified through the motor's current signal.

[0007] Furthermore, the rated power of the permanent magnet synchronous motor is 750W.

[0008] Furthermore, the motor drive control strategy in step one adopts a typical dual closed-loop speed control, including an inner current loop and an outer speed loop.

[0009] Furthermore, the eccentric orientation diagnosis model in step two is an optimized model based on the LSTM model. By combining GA and LSTM, the optimal combination of hyperparameters is found during the global search process, thereby improving model performance and increasing the error correction and diagnosis rate of eccentric orientation. The specific implementation method is as follows: 1) After the eccentric data is collected, the fused features extracted by time-frequency domain features and wavelet packet features are used as the input vector of the model; the true corresponding labels 0-7 of the 8 eccentric directions are input, and the predicted label data of the network model is used as the output. 2) Set the initialization parameters for the genetic optimization algorithm, including population size and crossover probability. Probability of mutation Number of iterations, etc. Represents the crossover probability of the previous generation. This represents the magnitude of the change in the crossover probability. Represents the mutation probability of the previous generation. Represents the magnitude of the change in the probability of mutation. This represents the average fitness of contemporary individuals. This represents the optimal fitness of a contemporary individual; the specific constraint formula is as follows: , ; 3) Create N initial populations. After crossover and mutation, evaluate the fitness of each individual using the evaluation function F. Select the individual with the highest fitness in each generation as the optimal solution for that generation, as shown in the following formula: , For the true value, Here, n represents the predicted value, and n represents the amount of data. 4) By repeatedly creating and updating the population, and performing a continuous iterative process, the maximum number of iterations is reached; at the end of the iteration, the best-performing solution is selected and applied to determine the architecture of the GA-LSTM neural network model.

[0010] Furthermore, the parameters to be set in the optimization process of the eccentric orientation diagnosis model in step two include the number of LSTM layers, the number of neurons in the LSTM layers, the number of fully connected layers, the number of neurons in each layer, and the batch size. The search range for the number of LSTM layers is [1,3], the search range for the number of units per LSTM layer is [32,512], the search range for the number of fully connected layers is [1,3], the search range for the number of neurons in each layer is [32,512], and the search range for the batch size is [16,128]. The number of LSTM layers, the number of neurons in the LSTM layers, the number of fully connected layers, the number of neurons in each layer, and the batch size are used as the search space for GA optimization, and the optimal combination of these parameters is found through a genetic algorithm.

[0011] Furthermore, the processing method for the preprocessed signal in step three is as follows: 1) The eccentric orientation was divided into 8 groups of fault points: up, down, left, right, upper left, lower left, upper right, and lower right. The motor sampling frequency was 4 kHz, the sampling time was 10 seconds, and there were a total of 8 orientation points with 40,000 current data points in each orientation. 2) The collected data is grouped into sliding groups using a time window function, i.e., the sliding window length and the moving step size are set according to actual needs; 3) Extract 15 time-domain features from each group of data after grouping by the sliding window function: maximum value, maximum absolute value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin, impulse, waveform, and peak index; and 4 frequency-domain features: average frequency, centroid frequency, root mean square frequency, and frequency variance; set the wavelet packet decomposition level to three levels, i.e., use the energy of 8 frequency bands as feature vectors; and fuse the above time-frequency domain features with the wavelet packet features and concatenate them together as the input signal of the neural network.

[0012] Furthermore, the decomposition method of the wavelet packet is as follows: 1) Input the original current signal and determine the wavelet basis function and the three-level decomposition hierarchy; 2) Perform wavelet transform on each node to decompose it into high-frequency and low-frequency sub-bands; 3) Determine if it is the final level; 4) If it is the final level, extract the wavelet packet decomposition function, calculate the energy value of each frequency band and normalize it; 5) If it is not the final level, repeat steps 2) and 3) until it is determined to be the final level.

[0013] Furthermore, the eccentric orientation diagnosis model trained in step two is used to diagnose the orientation of the preprocessed signal. The input signal obtained in step three, which is fused and normalized, is used as a test sample to train the eccentric orientation diagnosis model. The orientation prediction label is used as the output of the model. The network parameters are updated using the backpropagation algorithm based on the error between the true label and the predicted label of the sample.

[0014] Furthermore, the output labels in step four include 0, 1, 2, 3, 4, 5, 6 and 7, which correspond to eight sets of fault points in the eccentric orientation: up, down, left, right, upper left, lower left, upper right and lower right. The motor is installed and adjusted in the opposite direction using a three-dimensional bracket based on the main eccentric orientation of the fault point. The adjusted motor is then calibrated and verified using a laser alignment instrument.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method for correcting motor installation eccentricity based on recurrent neural networks. By performing time window sliding grouping on the collected eccentricity fault data, the time domain features, frequency domain features, and wavelet packet energy reconstruction of the data are extracted and the fused features are used as the input signal of the neural network model. An improved genetic optimization algorithm is used to seek the optimal combination of parameters such as the number of network layers and the number of units per layer, thereby improving the accuracy of eccentricity diagnosis to over 95% and enhancing the anti-interference ability against noise. Attached Figure Description

[0016] Figure 1This is a flowchart of the error correction process of the present invention; Figure 2 This is an overall diagnostic block diagram of the present invention; Figure 3 This is a block diagram of the GA-LSTM diagnostic system of the present invention; Figure 4 This is a flowchart of the wavelet packet decomposition process of the present invention; Figure 5 This is a schematic diagram of the original motor signal and the noise-added signal in this invention; Figure 6 This is a schematic diagram comparing the diagnostic rates under different signal-to-noise ratios in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figures 1 to 6 As shown, a method for correcting motor mounting eccentricity based on recurrent neural networks includes the following specific steps: Step 1: Initial installation of the motor and acquisition of motor current signal; Step 2: Train the eccentric orientation diagnosis model; Step 3: Input the preprocessed signal into the eccentric azimuth diagnostic model for azimuth diagnosis; Step 4: Output the label to determine if there is an eccentricity fault; Step 5: Determine the main eccentricity and install and adjust the motor in the opposite direction, then verify the calibration using a laser alignment instrument; After calibration, repeat steps one through four. When there is no eccentricity fault, the error correction process terminates; when there is still an eccentricity fault, repeat step five.

[0019] The implementation method of step one is as follows: A 750W permanent magnet synchronous motor drives a load through a flexible coupling. The load is loaded by a PB-0.6 magnetic powder brake, which can collect data under different loading conditions. The motor is mounted on a three-dimensional support on the test bench. The eccentricity and degree of the motor can be adjusted through the three-dimensional support. The TKSA31 SKF laser alignment instrument is used to quantify and calibrate the eccentricity. The motor's current signal, torque signal, and speed data signal are acquired by the motor incremental encoder and driver. The driver is a GSK six-axis servo driver. When the motor's torque and speed are stable, the motor's current signal is collected, and the fault location is identified through the motor's current signal.

[0020] The motor drive control strategy in step one adopts a typical dual closed-loop speed control, including an inner current loop and an outer speed loop.

[0021] The eccentric orientation diagnosis model in step two is an optimized model based on the LSTM model. The method of combining GA and LSTM can find the optimal combination of hyperparameters during the global search process, further improving model performance and increasing the error correction and diagnosis rate of eccentric orientation. The specific implementation method is as follows: 1) After eccentricity data acquisition, the fused features extracted from time-frequency domain features and wavelet packet features are used as the input vector of the model. The true corresponding labels 0-7 for the eight eccentricity directions are input, and the predicted label data of the network model is used as the output. 2) Set the initialization parameters for the genetic optimization algorithm, including population size and crossover probability. Probability of mutation Number of iterations, etc. Represents the crossover probability of the previous generation. This represents the magnitude of the change in the crossover probability. Represents the mutation probability of the previous generation. Represents the magnitude of the change in the probability of mutation. This represents the average fitness of contemporary individuals. This represents the optimal fitness of a contemporary individual; the specific constraint formula is as follows: , ; 3) Create N initial populations. After performing crossover and mutation genetic operations, evaluate the fitness of each individual using the evaluation function F. Select the individual with the highest fitness in each generation as the optimal solution for that generation. For the true value, Here is the predicted value, and n is the amount of data; the formula is as follows: ; 4) The updated population is repeatedly created and iterated continuously until the maximum number of iterations is reached. At the end of the iteration, the best-performing solution is selected and applied to determine the architecture of the GA-LSTM neural network model.

[0022] The parameters to be set in the optimization process of the eccentric orientation diagnosis model in step two include the number of LSTM layers, the number of neurons in the LSTM layers, the number of fully connected layers, the number of neurons in each layer, and the batch size. The search range for the number of LSTM layers is [1,3], the search range for the number of units per LSTM layer is [32,512], the search range for the number of fully connected layers is [1,3], the search range for the number of neurons in each layer is [32,512], and the search range for the batch size is [16,128]. The number of LSTM layers, the number of neurons in the LSTM layers, the number of fully connected layers, the number of neurons in each layer, and the batch size are used as the search space for GA optimization, and the optimal combination of these parameters is found through a genetic algorithm.

[0023] The processing method for the preprocessed signal in step three is as follows: 1) The eccentric orientation was divided into 8 groups of fault points: up, down, left, right, upper left, lower left, upper right, and lower right. The motor sampling frequency was 4 kHz, the sampling time was 10 seconds, and there were a total of 8 orientation points with 40,000 current data points in each orientation. 2) The collected data is grouped into sliding groups using a time window function, i.e., the sliding window length and the moving step size are set according to actual needs; 3) Extract 15 time-domain features from each group of data after grouping by the sliding window function: maximum value, maximum absolute value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin, impulse, waveform, and peak index; and 4 frequency-domain features: average frequency, centroid frequency, root mean square frequency, and frequency variance; set the wavelet packet decomposition level to three levels, i.e., use the energy of 8 frequency bands as feature vectors; and fuse the above time-frequency domain features with the wavelet packet features and concatenate them together as the input signal of the neural network. like Figure 4 As shown, the decomposition method of the wavelet packet is as follows: 1) Input the original current signal and determine the wavelet basis function and the three-level decomposition hierarchy; 2) Perform wavelet transform on each node to decompose it into high-frequency and low-frequency sub-bands; 3) Determine if it is the final level; 4) If it is the final level, extract the wavelet packet decomposition function, calculate the energy value of each frequency band and normalize it; 5) If it is not the final level, repeat steps 2) and 3) until it is determined to be the final level.

[0024] The eccentric orientation diagnosis model trained in step two is used to diagnose the orientation of the preprocessed signal. The input signal obtained in step three, which has been processed by fusion feature extraction and normalization, is used as a test sample to train the eccentric orientation diagnosis model. The orientation prediction label is used as the output of the model. The network parameters are updated by using the backpropagation algorithm based on the error between the true label and the predicted label of the sample, thereby improving the accuracy of the eccentric orientation diagnosis model in automatically correcting and identifying eccentric orientation.

[0025] The output labels in step four include 0, 1, 2, 3, 4, 5, 6 and 7, which correspond to eight sets of fault points in the eccentric orientation: up, down, left, right, upper left, lower left, upper right and lower right. The motor is installed and adjusted in the opposite direction using a three-dimensional bracket based on the main eccentric orientation of the fault point. The adjusted motor is then calibrated and verified using a laser alignment instrument.

[0026] like Figure 5 and Figure 6 As shown, to further verify the anti-interference and robustness of the GA-LSTM model in noisy environments, Gaussian white noise was artificially introduced into the collected fault signals, specifically as follows: Figure 5 As shown; Gaussian white noise is a common mathematical model used to simulate random noise in real-world environments. Gaussian noise follows a normal distribution, meaning it is symmetrically distributed around a mean, and most values ​​fall within a specific standard deviation. Gaussian noise has a uniform power density across all frequencies, and there is no correlation between any two values. This means that each sample in the noise is independent and does not depend on other samples. Its formula can be expressed as: , In the formula N For the generated noise, The mean, Standard deviation randn ( n ) for generation n A random number that conforms to a standard normal distribution.

[0027] The signal-to-noise ratio (SNR) of Gaussian noise is a key indicator of signal quality. It is the ratio of signal power to noise power, usually expressed in decibels (dB). A high SNR indicates less noise and better signal quality; conversely, a low SNR indicates higher noise levels and poorer signal quality. Converting the SNR to a linear form yields the noise factor F; a higher noise factor indicates a higher noise content in the signal. ,

[0028] To test the noise immunity of GA-LSTM, Gaussian noise was introduced into the eccentric fault signal, and the percentage performance of the model under different noise factors or signal-to-noise ratios was compared: Table 1 Comparison of GA-LSTM performance metrics noise factor accuracy Accuracy Recall rate F1 score 0.001 94.87 95.07 94.89 94.90 0.005 93.27 93.73 93.35 93.34 0.01 91.03 91.09 90.95 90.99 0.05 85.42 85.56 85.38 85.28 0.1 77.88 79.67 78.06 78.08 As shown in Table 1, under low noise conditions (noise factors of 0.001 and 0.005), the model exhibits high performance, with all four performance indicators exceeding 93%. Specifically, at a noise factor of 0.001, all indicators exceed 94.5%, demonstrating the model's strong robustness under low noise conditions. Under medium noise conditions (noise factors of 0.01 and 0.05), all performance indicators begin to decline. Particularly at a noise factor of 0.05, the accuracy drops to 85.42%, indicating that medium noise environments negatively impact the model's robustness, although a high diagnostic rate can still be maintained. Under high noise conditions (noise factor of 0.1), the diagnostic rate further decreases, suggesting that the model's robustness in high noise environments needs further improvement.

[0029] To further demonstrate the effectiveness of the GA algorithm in selecting optimal parameters for LSTM models, the diagnostic rates of models with and without the GA algorithm were compared under noise environments with different signal-to-noise ratios. The results are as follows: Figure 6 As shown.

[0030] Overall, GA-LSTM demonstrates superior performance across the entire signal-to-noise ratio range compared to traditional LSTM models. This indicates that the addition of GA optimizes the network structure and weights, helping the LSTM model better adjust its parameters and enhance its robustness to noise, thus enabling more accurate diagnosis in noisy environments.

[0031] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for correcting motor mounting eccentricity based on recurrent neural networks, characterized in that: The specific steps include the following: Step 1: Initial installation of the motor and acquisition of motor current signal; Step 2: Train the eccentric orientation diagnosis model; Step 3: Input the preprocessed signal into the eccentric azimuth diagnostic model for azimuth diagnosis; Step 4: Output the label to determine if there is an eccentricity fault; Step 5: Determine the main eccentricity and install and adjust the motor in the opposite direction, then verify the calibration using a laser alignment instrument; After calibration, repeat steps one through four. When there is no eccentricity fault, the error correction process terminates; when there is still an eccentricity fault, repeat step five.

2. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 1, characterized in that: The implementation method of step one is as follows: A permanent magnet synchronous motor drives a load through a flexible coupling. The load is loaded by a magnetic powder brake, which enables data acquisition under different loading conditions. The motor is mounted on a three-dimensional support on the test bench. The eccentricity and degree of the motor can be adjusted through the three-dimensional support. The laser alignment instrument is used to quantify and calibrate the eccentricity. The motor's current signal, torque signal, and speed data signal are acquired by the motor incremental encoder and driver. The driver is a GSK six-axis servo driver. When the motor's torque and speed are stable, the motor's current signal is collected, and the fault location is identified through the motor's current signal.

3. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 2, characterized in that: The rated power of the permanent magnet synchronous motor is 750W.

4. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 2, characterized in that: The motor drive control strategy in step one adopts a typical dual closed-loop speed control, including an inner current loop and an outer speed loop.

5. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 1, characterized in that: The eccentric orientation diagnosis model in step two is an optimized model based on the LSTM model. By combining GA and LSTM, the optimal combination of hyperparameters is found during the global search process, improving model performance and increasing the error correction and diagnosis rate of eccentric orientation. The specific implementation method is as follows: 1) After the eccentric data is collected, the fused features extracted by time-frequency domain features and wavelet packet features are used as the input vector of the model; the true corresponding labels 0-7 of the 8 eccentric directions are input, and the predicted label data of the network model is used as the output. 2) Set the initialization parameters for the genetic optimization algorithm, including population size, crossover probability, etc. Probability of mutation Number of iterations, etc. Represents the crossover probability of the previous generation. This represents the magnitude of the change in the crossover probability. Represents the mutation probability of the previous generation. Represents the magnitude of the change in the probability of mutation. This represents the average fitness of contemporary individuals. This represents the optimal fitness of a contemporary individual; the specific constraint formula is as follows: ， ， 3) Create N initial populations. After crossover and mutation, evaluate the fitness of each individual using the evaluation function F. Select the individual with the highest fitness in each generation as the optimal solution for that generation, as shown in the following formula: ， For the true value, Here, n represents the predicted value, and n represents the amount of data. 4) By repeatedly creating and updating the population, and performing a continuous iterative process, the maximum number of iterations is reached; at the end of the iteration, the best-performing solution is selected and applied to determine the architecture of the GA-LSTM neural network model.

6. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 1, characterized in that: The parameters to be set in the optimization process of the eccentric orientation diagnosis model in step two include the number of LSTM layers, the number of neurons in the LSTM layers, the number of fully connected layers, the number of neurons in each layer, and the batch size. The search range for the number of LSTM layers is [1,3], the search range for the number of units per LSTM layer is [32,512], the search range for the number of fully connected layers is [1,3], the search range for the number of neurons in each layer is [32,512], and the search range for the batch size is [16,128]. The number of LSTM layers, the number of neurons in the LSTM layers, the number of fully connected layers, the number of neurons in each layer, and the batch size are used as the search space for GA optimization, and the optimal combination of these parameters is found through a genetic algorithm.

7. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 1, characterized in that: The processing method for the preprocessed signal in step three is as follows: 1) The eccentric orientation was divided into 8 groups of fault points: up, down, left, right, upper left, lower left, upper right, and lower right. The motor sampling frequency was 4 kHz, the sampling time was 10 seconds, and there were a total of 8 orientation points with 40,000 current data points in each orientation. 2) The collected data is grouped into sliding groups using a time window function, i.e., the sliding window length and the moving step size are set according to actual needs; 3) Extract 15 time-domain features from each group of data after grouping by the sliding window function: Maximum value, maximum absolute value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis, skewness, margin, impulse, waveform, and peak index; four frequency domain features: average frequency, centroid frequency, root mean square frequency, and frequency variance; the wavelet packet decomposition level is set to three levels, i.e., the energy of eight frequency bands is used as feature vectors; the above time-frequency domain features are fused with the wavelet packet features and then concatenated as the input signal of the neural network; The wavelet packet decomposition method is as follows: 1) Input the original current signal and determine the wavelet basis function and the three-level decomposition hierarchy; 2) Perform wavelet transform on each node to decompose it into high-frequency and low-frequency sub-bands; 3) Determine if it is the final level; 4) If it is the final level, extract the wavelet packet decomposition function, calculate the energy value of each frequency band and normalize it; 5) If it is not the final level, repeat steps 2) and 3) until it is determined to be the final level.

8. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 6, characterized in that: The eccentric orientation diagnosis model trained in step two is used to diagnose the orientation of the preprocessed signal. The input signal obtained in step three, which has been processed by fusion feature extraction and normalization, is used as a test sample to train the eccentric orientation diagnosis model. The orientation prediction label is used as the output of the model. The network parameters are updated by using the backpropagation algorithm based on the error between the true label and the predicted label of the sample.

9. The method for correcting motor mounting eccentricity based on recurrent neural networks according to claim 1, characterized in that: The output labels in step four include 0, 1, 2, 3, 4, 5, 6 and 7, which correspond to eight sets of fault points in the eccentric orientation: up, down, left, right, upper left, lower left, upper right and lower right. The motor is installed and adjusted in the opposite direction using a three-dimensional bracket based on the main eccentric orientation of the fault point. The adjusted motor is then calibrated and verified using a laser alignment instrument.

Images