Fault identification and analysis method of permanent magnet synchronous motor for belt conveyor

Abstract

The invention relates to the technical field of motor fault identification, in particular to a permanent magnet synchronous motor fault identification and analysis method for a belt conveyor, which comprises the following steps of: extracting a characteristic change rule of a stator current signal and a vibration acceleration signal of a permanent magnet synchronous motor under a typical fault condition; collecting a vibration signal and a current signal of the fault motor; carrying out noise reduction processing on the collected vibration signals and current signals; converting the vibration signal and the current signal after noise reduction into a two-dimensional image data set; vertically splicing and fusing the vibration signal image and the current signal image; dividing the fused image data into a training set and a test set according to a preset proportion; establishing a convolutional neural network model; and inputting the training set into the convolutional neural network model. According to the permanent magnet synchronous motor fault diagnosis method based on the data layer fusion strategy, the feature information of multiple fault modes of the permanent magnet synchronous motor is comprehensively captured, and the accuracy and robustness of fault recognition are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of motor fault identification, and in particular to a permanent magnet synchronous motor fault identification and analysis method for a belt conveyor. BACKGROUND

[0002] Permanent magnet synchronous motors are widely used in large-scale continuous operation equipment such as belt conveyors due to their high efficiency and high power density. However, due to the harsh working conditions of high load, high dust and continuous operation of the belt conveyor, the permanent magnet synchronous motor is prone to typical faults such as demagnetization of the permanent magnet, short circuit of the stator winding, rotor eccentricity and bearing failure. The initial characteristics of these faults are weak and easy to be covered by noise. Traditional diagnosis methods rely on manual feature extraction and are difficult to meet the requirements of real-time and accuracy. With the progress of signal detection and processing technology and the popularization and application of artificial intelligence, fault diagnosis methods based on signals and traditional machine learning have gradually become the research focus due to their high efficiency, fast processing and high precision. However, most of the existing methods still rely on manual feature extraction and are sensitive to noise. In addition, they lack effective fusion mechanisms when dealing with multi-source heterogeneous signals, resulting in insufficient diagnosis accuracy and robustness in complex working conditions. In particular, the parameters of traditional variational mode decomposition depend on experience, which limits its adaptive application in engineering. Therefore, a new method that can adaptively reduce noise, fuse multiple signals and perform end-to-end intelligent diagnosis is needed. SUMMARY

[0003] The application aims to solve the above problems and provides a permanent magnet synchronous motor fault identification and analysis method for a belt conveyor. The application combines speed pause particle swarm optimization with variational mode decomposition to achieve adaptive noise reduction of current and vibration signals. Continuous wavelet transform and Markov transition field are used to convert one-dimensional signals into two-dimensional images. A dual-channel convolutional neural network based on data layer fusion is constructed to fuse vibration and current information, significantly improving the accuracy and robustness of fault diagnosis. The technical scheme adopted is as follows: A permanent magnet synchronous motor fault identification and analysis method for a belt conveyor, comprising the following steps, S1. Extracting the characteristic variation law of the stator current signal and the vibration acceleration signal of the permanent magnet synchronous motor under typical fault conditions; S2. Collecting the vibration signal and the current signal of the faulty motor; S3. Noise reduction processing of the collected vibration signal and current signal; S4. Using continuous wavelet transform and / or Markov transition field to convert the noise-reduced vibration signal and current signal into a two-dimensional image dataset; S5. Vertically splicing and fusing the vibration signal image and the current signal image generated using the same method; S6. Dividing the fused image data into a training set and a test set according to a predetermined proportion; S7. Establishing a convolutional neural network model; including three convolutional modules arranged in sequence, and the number of convolutional kernels contained in each convolutional module is 32, 16 and 8 respectively, the size of each convolutional kernel is 3X3, and each convolutional module is embedded with a pooling layer; further comprising a flattening layer and two fully connected layers; S8. Inputting the training set into the convolutional neural network model, and setting labels for different typical fault types respectively; The step S3 comprises: S3-1. Performing variational modal decomposition on the collected signals; S3-2. Optimizing the decomposition mode number and the penalty factor in step S3-1 based on a velocity pause particle swarm optimization algorithm; S3-3. Calculating the threshold value of each modal component in the signal according to energy entropy, and selecting the modal components greater than the threshold value for feature reconstruction.

[0004] On the basis of the above scheme, the typical faults include inter-turn short circuit fault, air gap fault, permanent magnet demagnetization fault, bearing inner ring fault, bearing rolling element fault and bolt loosening fault.

[0005] On the basis of the above scheme, the step S2 comprises: S2-1. Building an experimental platform, the experimental platform comprising a motor base, permanent magnet synchronous motors of different fault types, a belt conveyor simulation load, current sensors and vibration acceleration sensors, the permanent magnet synchronous motor being installed on the motor base, the number of current sensors being 3, and the current sensors being arranged on the three phases of the stator wire of the permanent magnet synchronous motor, the number of vibration acceleration sensors being 3, and the vibration acceleration sensors being arranged on the H, A and V directions of the driving end of the permanent magnet synchronous motor; S2-2. According to the different fault types corresponding to the signals to be collected, installing only the permanent magnet synchronous motor of the corresponding fault on the experimental platform, and collecting the time domain and frequency domain change signals of the current and vibration acceleration.

[0006] Preferably, the step S3-1 comprises: The modal components of each modal component Performing Hilbert transform to obtain the corresponding analytical signal, and then extracting the one-sided spectrum of the modal component, the formula being as follows (3-1) Wherein is a pulse function, t is time, the unit is s, j is an imaginary unit, and * represents convolution; An exponential term is introduced in each one-sided spectrum So that the spectrum modulation of the modal component is within the base frequency band, the formula being as follows (3-2) A bandwidth evaluation criterion is constructed based on the squared 2-norm of the demodulated signal gradient. Simultaneously, mode reconstruction constraints are introduced to ensure the completeness of the decomposition, thereby constructing a variational optimization model, as shown in the following formula. (3-3) In the formula, , , This represents the partial derivative of the function with respect to t; When solving the variational model of equation (3-3), Lagrange multipliers are introduced. and secondary penalty factor This allows for the transformation from constrained to unconstrained variational problems, ultimately yielding the extended Lagrange form: (3-4) The algorithm is continuously updated and iterated using the alternating direction multiplier algorithm. To find the saddle point of equation (3-4), the iterative method is as follows: (3-5) (3-6) (3-7) (3-8) In the formula, n is the number of iterations. , The update factor is ε, which is a positive number greater than 0, representing the precision. By using the Plancheral Fourier isometry transform, we can solve equation (3-8) to obtain... (3-9) (3-10) (3-11) The frequency domain of each modal component is continuously updated and then converted to the time domain through inverse Fourier transform.

[0007] Preferably, the optimization principle of step S3-2 includes: An adaptive speed adjustment is achieved using a speed holding mechanism, as shown in equation (3-16). (3-16) in Let be the velocity of the i-th particle at time t. and Here, ω is the learning factor, r1 is the bearing inner diameter, and r2 is the bearing outer diameter. A random number within the interval [0,1]. Maintain a probability threshold for speed; A velocity pause coefficient 'a' is introduced as a key regulating factor for particle motion characteristics. When a > 1, the particle follows a velocity update paradigm as shown in equation (3-17). (3-17) When a ≤ 1, the particle velocity is constrained, exhibiting approximately constant motion characteristics, i.e. (3-18) A dual-population cooperative evolution strategy is adopted to improve overall optimization performance. The first population N1 uses a velocity-pausing particle swarm optimization algorithm, while the second population N2 moves based on the globally optimal position without calculating velocity. (3-19) X i (t) represents the displacement of the i-th particle at time t. f r Here, is the synchronous frequency of the permanent magnet synchronous motor, a(t) is the acceleration response, and r is the stator radius; Using the envelope spectrum peak factor E c As the fitness function, the signal envelope spectrum amplitude sequence is assumed to be... ,but (3-20); in, For envelope spectrum in Amplitude within the frequency range in vibration signals The value should be greater than the rotational frequency of the permanent magnet synchronous motor; The maximum fault frequency of the transmission system is in the current signal. The value should be greater than the power supply base frequency. f r This refers to the rotational frequency of the permanent magnet synchronous motor. f i The rotational frequency of the rolling element; The optimization steps include: S3-2-1. Define parameters and initialize particle position and velocity; S3-2-2. Update the velocities of the particles in the first group according to equation (3-16); S3-2-3. Update the positions of the particles in the first group according to equation (3-17); S3-2-4. Update the positions of the particles in the second group according to equation (3-18); S3-2-5. Evaluate the fitness of all particles; S3-2-6. Update the individual optimal and global optimal values ​​of particles in the first group; S3-2-7. Update the global optimum of particles in the second group; S3-2-8. Determine if the maximum number of iterations has been reached; if not, repeat the subsequent steps from step S3-2-2; if the maximum number of iterations has been reached, proceed to step S3-2-9. S3-2-9. Return to the global optimum.

[0008] Preferably, the continuous wavelet transform uses Morlet wavelet basis functions, which are defined as follows: (4-1) In the formula, t is time and ω0 is angular frequency.

[0009] Preferably, the construction of the Markov transition field includes dividing the time series into Q quantile bins and calculating the transition matrix W and the Markov transition matrix M.

[0010] Preferably, in step S5, the vibration signal image and the current signal image are vertically stitched together to form a dual-channel fused image, with the image size uniformly set to 224×224 pixels and then normalized.

[0011] Preferably, in step S7, a ReLU activation function is used after each convolutional layer, max pooling is used for the pooling layer, and a Dropout layer is introduced after the fully connected layer with a dropout rate of 0.5.

[0012] The beneficial effects of this invention are as follows: This method combines velocity-pause particle swarm optimization with variational mode decomposition (VMD) and applies it to denoising current and vibration acceleration signals of permanent magnet synchronous motors (PMSMs). It effectively solves the problem of feature extraction difficulties caused by noise interference and overcomes the limitation of manually setting the decomposition modulus and penalty factor in traditional VMD algorithms, thus improving signal quality and the reliability of feature extraction. The fault signal is denoised by using continuous wavelet transform and Markov transform field to convert the one-dimensional signal into a two-dimensional feature map, which is then used as the input data for the convolutional neural network. A fault diagnosis method for permanent magnet synchronous motors based on a data layer fusion strategy was developed. This method comprehensively captures the feature information of various fault modes of permanent magnet synchronous motors, effectively integrates vibration signals and current signals, and constructs an improved dual-channel input architecture of a deep convolutional neural network. This significantly improves the accuracy and robustness of fault identification and provides a new path for the diagnosis of permanent magnet synchronous motors. Attached Figure Description

[0013] Figure 1 Flowchart of the steps of this invention; Figure 2This invention is based on a velocity-pause particle swarm optimization algorithm flowchart; Figure 3 : A comparative analysis chart of the fitness of the particle swarm optimization algorithm of this invention; Figure 4 : Schematic diagram of multi-source information image fusion of the present invention; Figure 5 : A schematic diagram of the fault diagnosis process of this invention; Detailed Implementation The present invention will be further described below with reference to the accompanying drawings and embodiments: In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0014] A method for fault identification and analysis of permanent magnet synchronous motors used in belt conveyors, such as... Figure 1 and Figure 5 As shown, it includes the following steps: S1. Extract the characteristic variation patterns of stator current signals and vibration acceleration signals of permanent magnet synchronous motors under typical fault conditions; the typical faults include inter-turn short circuit faults, air gap faults, permanent magnet demagnetization faults, bearing inner ring faults, bearing rolling element faults, and bolt loosening faults; the characteristic variation patterns of the above typical faults can be found in references such as "A Diagnostic Method for Inter-turn Short Circuit Faults of Permanent Magnet Synchronous Motors Based on Optimized Neural Network Architecture", "A New Method for Early Fault Diagnosis of Permanent Magnet Synchronous Motors under Multiple Working Conditions", and "Fault Diagnosis of Demagnetization Faults and Sensor Faults of Permanent Magnet Synchronous Motors". S2. Collect vibration and current signals of the faulty motor; Further, step S2 includes: S2-1. Construct an experimental platform, which includes a motor base, permanent magnet synchronous motors with different fault types, current sensors and vibration acceleration sensors. The permanent magnet synchronous motor is mounted on the motor base. There are three current sensors, which are respectively set on the three phases of the stator wires of the permanent magnet synchronous motor. There are three vibration acceleration sensors, which are respectively set on the H, A and V directions of the drive end of the permanent magnet synchronous motor. Optionally, the simulated load on the belt conveyor is achieved through an adjustable mechanical loading device that uses the gravity of weights to generate a constant frictional resistance torque, which is applied to the drive shaft of the motor to simulate the load borne by the belt conveyor under stable operating conditions. Considering load variations, the experimental platform should have the capability to adjust the load during testing to ensure successful testing under different load conditions. S2-2. Depending on the fault type corresponding to the signal to be collected, install a permanent magnet synchronous motor with only the corresponding fault on the experimental platform to collect the time-domain and frequency-domain variation signals of current and vibration acceleration.

[0015] For each type of motor fault, a brand-new motor is used for targeted fault handling: When simulating a stator inter-turn short circuit fault, a variable resistor is connected in series between two stator windings in the three-phase windings, and the current will flow from one winding through the resistor to the other winding; when simulating an air gap fault, an eccentric bearing is installed on the permanent magnet synchronous motor, and the outer ring of the eccentric bearing is laser-cut to a thickness of 1mm; when simulating a rotor bar breakage fault, holes are drilled in the rotor bars, and a certain amount of asphalt is added to the drilled holes to correct the eccentricity; when simulating a loose base bolt fault, two bolts on the same side of the permanent magnet synchronous motor are loosened.

[0016] Current and vibration signals of permanent magnet synchronous motors under normal conditions and motors with various fault types were collected under the same loading force and speed. On the one hand, the characteristics of typical faults were extracted, analyzed and verified, and on the other hand, preparations were made for the subsequent construction of datasets and fault identification verification.

[0017] S3. Noise reduction processing is performed on the collected vibration and current signals; Further, step S3 includes: S3-1. Perform variational mode decomposition on the acquired signal; variational mode decomposition optimizes the variational model to decompose the complex original signal x into K independent modal components. In this process, each mode can be explicitly defined with its unique center frequency. and frequency band range.

[0018] Step S3-1 includes: For each mode The Hilbert transform is applied to obtain the corresponding analytic signal, thereby enabling the extraction of the one-sided spectrum of the modal components. The formula is as follows: (3-1) in Here, t represents time (in seconds), j represents the imaginary unit, and * denotes convolution. Introduce an exponential term into each one-sided spectrum. This ensures that the spectral modulation of each mode is within the fundamental frequency band, as shown in the following formula. (3-2) A bandwidth evaluation criterion is constructed based on the squared 2-norm of the demodulated signal gradient. Simultaneously, mode reconstruction constraints are introduced to ensure the completeness of the decomposition, thereby constructing a variational optimization model, as shown in the following formula. (3-3) In the formula, , , This represents the partial derivative of the function with respect to t; When solving the variational model of equation (3-3), Lagrange multipliers are introduced. and secondary penalty factor This allows for the transformation from constrained to unconstrained variational problems, ultimately yielding the extended Lagrange form: (3-4) The algorithm is continuously updated and iterated using the alternating direction multiplier algorithm. To find the saddle point of equation (3-4), the iterative method is as follows: (3-5) (3-6) (3-7) (3-8) In the formula, n is the number of iterations. , The update factor is ε, which is a positive number greater than 0, representing the precision. By using the Plancheral Fourier isometry transform, we can solve equation (3-8) to obtain... (3-9) (3-10) (3-11) The frequency domain of each modal component is continuously updated and then converted to the time domain through inverse Fourier transform.

[0019] S3-2. In the variational mode decomposition algorithm, the decomposition modulus K and the penalty factor α are two important parameters affecting the denoising results. The choice of K directly impacts the accuracy of the decomposition results. The value of α mainly affects the decomposition accuracy: too small a value leads to decreased accuracy, while too large a value increases the computational burden and may even cause the algorithm to fail to converge. Therefore, the decomposition modulus and penalty factor in step S3-1 are optimized based on the velocity-pausing particle swarm optimization algorithm. In a D-dimensional search space, each particle seeks the optimal solution through continuous iteration. Each particle contains two basic pieces of information during each iteration: a position vector and a velocity vector. The position of the c-th particle is represented by a vector... Velocity is represented by a vector. This means that the particle dynamically adjusts these two vectors based on its own optimal position (individual extremum) and the optimal position found by the entire population (global extremum), thereby updating its position and velocity. The update formula is as follows: (3-12) (3-13) in h is the number of iterations. Let be the velocity in dimension d of particle o in the h-th iteration. , where c1 and c2 are inertia weights, and c1 and c2 are learning factors. For those within the interval random numbers, Let be the position of the individual extreme point of particle o in the d-th dimension during the h-th iteration. Let be the current position of particle o in dimension d during the h-th iteration. Let be the position of the global extremum of the entire population in dimension d during the h-th iteration.

[0020] Inertia weights in particle swarm optimization Following the strategy of concave function variation, since the constant is fixed, the inertia weight in this scheme... According to the changes in the concave function, the adjustment of the concave function strategy can be expressed as follows: (3-14) in, This represents the maximum weight. Let h be the minimum weight, h be the number of iterations, and H be the maximum number of iterations. (3-15).

[0021] The optimization principle of step S3-2 includes: In the Velocity Pause-Based Particle Swarm Optimization (VVPSO) algorithm, a velocity preservation mechanism is innovatively introduced, enabling particles to exhibit diverse motion patterns during iteration. Unlike traditional Particle Swarm Optimization (PSO) algorithms, which force velocity updates with each iteration, the Velocity Pause-Based PVPSO algorithm allows particles to retain their previous velocity values ​​based on specific conditions. This design results in three unique velocity states: Acceleration state: When environmental conditions are favorable, particles can gain a large velocity increment, accelerating the search speed; Deceleration state: When approaching the potential optimal solution region, the velocity will decrease, enhancing the accuracy of the local search; Uniform state: By maintaining the current velocity, the stability of the search space can be achieved.

[0022] This adaptive velocity adjustment mechanism significantly improves the algorithm's search efficiency. During the global search phase, particles can rapidly explore unknown regions using an accelerated state; after finding a potential optimal solution region, they can switch to a deceleration or constant velocity state for refined searching. This dynamic balancing mechanism effectively overcomes the standard particle swarm optimization algorithm's tendency to get trapped in local optima, improving its performance in complex optimization problems.

[0023] An adaptive speed adjustment is achieved using a speed holding mechanism, as shown in equation (3-16). (3-16) in Let be the velocity of the i-th particle at time t. and Here, ω is the learning factor, r1 is the bearing inner diameter, and r2 is the bearing outer diameter. A random number within the interval [0,1]. A probability threshold is maintained for velocity; this probabilistic velocity maintenance mechanism introduces more uncertainty into the particle's trajectory, enhances population diversity, and thus improves the algorithm's ability to escape local optima.

[0024] A velocity pause coefficient 'a' is introduced as a key regulating factor for particle motion characteristics. When a > 1, the particle follows a velocity update paradigm as shown in equation (3-17). (3-17) When a ≤ 1, the particle velocity is constrained, exhibiting approximately constant motion characteristics, i.e. (3-18) A dual-population co-evolutionary strategy is adopted to improve overall optimization performance. The first population, N1, uses a velocity-pausing particle swarm optimization algorithm. By adjusting the leading coefficient of the velocity equation and introducing a velocity-pausing mechanism, the adaptive behavior of particles in the search space is enhanced. The second population, N2, moves based on the globally optimal position without calculating velocity. (3-19) X i (t) represents the displacement of the i-th particle at time t. f r Here, is the synchronous frequency of the permanent magnet synchronous motor, a(t) is the acceleration response, and r is the stator radius; This clustering structure allows the algorithm to maintain exploration diversity while effectively suppressing the local optimum problem caused by premature aggregation of particle swarms.

[0025] Using the envelope spectrum peak factor E c As the fitness function, the signal envelope spectrum amplitude sequence is assumed to be... ,but (3-20); in, For envelope spectrum in Amplitude within the frequency range in vibration signals The value should be greater than the rotational frequency of the permanent magnet synchronous motor; The maximum fault frequency of the transmission system is in the current signal. The value should be greater than the power supply base frequency. f r This refers to the rotational frequency of the permanent magnet synchronous motor. f i The rotational frequency of the rolling element; The optimization steps include: S3-2-1. Define parameters and initialize particle position and velocity; S3-2-2. Update the velocities of the particles in the first group according to equation (3-16); S3-2-3. Update the positions of the particles in the first group according to equation (3-17); S3-2-4. Update the positions of the particles in the second group according to equation (3-18); S3-2-5. Evaluate the fitness of all particles; S3-2-6. Update the individual optimal and global optimal values ​​of particles in the first group; S3-2-7. Update the global optimum of particles in the second group; S3-2-8. Determine if the maximum number of iterations has been reached; if not, repeat the subsequent steps from step S3-2-2; if the maximum number of iterations has been reached, proceed to step S3-2-9. S3-2-9. Return to the global optimum.

[0026] Optimization results are as follows Figure 3 As shown, Figure 3 Figure -a shows the vibration signal optimization fitness curve. Figure 3 Figure -b shows the fitness curve for current signal optimization. It can be seen that the minimum fitness and number of iterations of the velocity-pause particle swarm optimization algorithm are both lower than those of the traditional particle swarm optimization algorithm, demonstrating the superiority of the velocity-pause particle swarm optimization algorithm.

[0027] S3-3. Calculate the threshold of each modal component in the signal based on the energy entropy, select the modal components that are greater than the threshold for feature reconstruction, and then obtain the fault feature signal with noise reduction completed.

[0028] The fault signal is subjected to variational mode decomposition with optimal parameters to obtain K sets of modal components. The energy entropy reflects the energy distribution of each modal component in the signal. Typically, modal components containing the main characteristic information of the signal have a higher energy proportion, and therefore their energy entropy is also larger, while spurious and noise components have a smaller energy proportion and therefore their energy entropy is also smaller.

[0029] S4. Use continuous wavelet transform and / or Markov transform field to convert the denoised vibration signal and current signal into a two-dimensional image dataset; Wavelet transform was performed on the vibration and current signals of normal state, inter-turn short circuit fault, outer ring fault, rolling element fault, and rolling bearing fault at 1480 rpm to obtain wavelet transform-based time-frequency diagrams. Two-dimensional matrix transform of Markov transferred field was performed on the vibration and current signals of normal state, inter-turn short circuit fault, air gap fault, permanent magnet demagnetization fault, bearing inner ring fault, bearing rolling element fault, and bolt loosening fault to obtain Markov transferred field images based on vibration and current signals.

[0030] The continuous wavelet transform uses Morlet wavelet basis functions, which are defined as follows: (4-1) In the formula, t is time and ω0 is angular frequency.

[0031] The construction of the Markov transition field involves dividing the time series into Q quantile bins and calculating the transition matrix W and the Markov transition matrix M.

[0032] Continuous wavelet transform calculates the "similarity" of a signal at different times and frequencies by translating a Morlet wavelet function on the time axis, scaling it on the frequency axis, and performing an inner product with the original signal. Markov transformed fields encode signals from a completely different perspective. They do not concern themselves with the signal's frequency but focus on the dynamic statistical transition characteristics of time series values. First, the entire time series is divided into Q quantile bins. Then, a transition matrix is ​​constructed to form the Markov transition matrix, and finally, a Markov transformed field image is generated.

[0033] Continuous wavelet transform can simultaneously preserve information in both the time and frequency domains. Unlike Fourier transform or short-time Fourier transform, continuous wavelet transform offers higher time resolution for high-frequency components and higher frequency resolution for low-frequency components. It is sensitive to non-stationary signals and closely matches the physical mechanisms of faults. Markov transformed fields can capture the dynamic evolution patterns of signals; different faults can cause signals to exhibit different "walking" patterns. They are insensitive to absolute amplitude but sensitive to changing patterns, and can also visualize long-term dependencies.

[0034] S5. Vertically stitch and fuse the vibration signal image and the current signal image generated using the same method; The collected vibration and current signals are normalized and converted into two-dimensional images, such as... Figure 4 This is a schematic diagram of the fused image, where Figure 4 Figure (a) is a schematic diagram of continuous wavelet transform image fusion. Figure 4 Figure (b) is a schematic diagram of Markov transformed field image fusion. The upper part of each image is a two-dimensional image obtained by converting the current signal under the rated test parameters, and the lower part is a two-dimensional image obtained by converting the vibration signal under the same test conditions. Then, the images generated by the two signals are vertically stitched and fused to form the input data of the convolutional neural network, which is finally input into the model for diagnosis.

[0035] Specifically, the vibration signal image and the current signal image are vertically stitched together to form a dual-channel fused image. The image size is uniformly 224×224 pixels and is normalized.

[0036] S6. Divide the fused image data into a training set and a test set according to a predetermined ratio; S7. Establish a convolutional neural network model, including three convolutional modules set in sequence, with the number of convolutional kernels in each module being 32, 16, and 8 respectively, and the size of each convolutional kernel being 3x3. Each convolutional module embeds one pooling layer. It also includes one flattening layer and two fully connected layers. Each convolutional layer is followed by the ReLU activation function, the pooling layer uses max pooling, and a Dropout layer is introduced after the fully connected layer with a dropout rate of 0.5.

[0037] S8. Input the training set into the convolutional neural network model, and assign labels to different typical fault types; the labels are shown in the table below. Table of Fault Classification Labels for Permanent Magnet Synchronous Motors

[0038] Based on vibration signals, current signals, and the fusion results of vibration and current signals, the test set obtained by fusing the two-dimensional Fourier transform graph with the multi-source data is used as input to the convolutional neural network model to extract different fault features of various faults of permanent magnet synchronous motors. The convergence curves and fitting effects of the above two types of inputs and the fusion results of the multi-source data obtained after the two types of transformations are compared.

[0039] The present invention has been described above by way of example, but the present invention is not limited to the specific embodiments described above. Any modifications or variations made based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for fault recognition and analysis of a permanent magnet synchronous motor for a belt conveyor, characterized by, Comprising the following steps, S1. Extract the characteristic variation law of the stator current signal and the vibration acceleration signal of the permanent magnet synchronous motor under typical fault conditions; S2. Collect the vibration signal and the current signal of the fault motor; S3. Perform noise reduction processing on the collected vibration signal and current signal; S4. Convert the noise-reduced vibration signal and current signal into a two-dimensional image data set using continuous wavelet transform and / or Markov transform field; S5. Vertically splice and fuse the vibration signal images and current signal images generated using the same method; S6. Divide the fused image data into a training set and a test set according to a predetermined ratio; S7. Establish a convolutional neural network model; including three convolutional modules set in sequence, and the number of convolutional kernels contained in each convolutional module is 32, 16 and 8 respectively, the size of each convolutional kernel is 3X3, and each convolutional module is embedded with a pooling layer; further comprising a flattening layer and two fully connected layers; S8. Input the training set into the convolutional neural network model, and set labels for different typical fault types respectively; The step S3 comprises: S3-1. Perform variational modal decomposition on the collected signal; S3-2. Optimize the decomposition mode number and the penalty factor in step S3-1 based on the speed pause particle swarm optimization algorithm; S3-3. Calculate the threshold value of each modal component in the signal according to energy entropy, and select the modal components greater than the threshold value for feature reconstruction.

2. The method according to claim 1, characterized in that, The typical faults include turn-to-turn short circuit fault, air gap fault, permanent magnet demagnetization fault, bearing inner ring fault, bearing rolling element fault and bolt loosening fault.

3. The belt conveyor permanent magnet synchronous motor fault identification and analysis method according to claim 2, characterized in that, The step S2 comprises: S2-1. Build an experimental platform, the experimental platform comprises a motor base, permanent magnet synchronous motors of different fault types, a belt conveyor simulation load, current sensors and vibration acceleration sensors, the permanent magnet synchronous motor is installed on the motor base, the number of current sensors is three, which are respectively arranged on the three phases of the permanent magnet synchronous motor stator wire, the number of vibration acceleration sensors is three, which are respectively arranged in the H, A and V directions of the driving end of the permanent magnet synchronous motor; S2-2. According to the different fault types corresponding to the signals to be collected, install only the permanent magnet synchronous motor with the corresponding fault on the experimental platform, and collect the time domain and frequency domain variation signals of the current and vibration acceleration.

4. The belt conveyor permanent magnet synchronous motor fault identification and analysis method according to claim 1, characterized in that, The step S3-1 comprises: For each modality The Hilbert transform is implemented to obtain the corresponding analytic signal, and then the single-sided spectrum of the modal component is extracted, and the formula is as follows （3-1） wherein is a pulse function, t is time in seconds, j is the imaginary unit, and * denotes convolution; An exponential term is introduced in each single side spectrum such that the spectrum modulation of the individual modes is within the base band, the formula is as follows （3-2） A bandwidth evaluation criterion is constructed based on the square of the two-norm of the demodulation signal gradient, and a modal reconstruction constraint condition is introduced to ensure the completeness of the decomposition, thereby constructing a variational optimization model, as follows （3-3） wherein , , denotes the partial derivative of the function with respect to t; In solving the variational model of equation (3-3), by introducing Lagrange multiplier and quadratic penalty factor , the conversion of constraint to non-constraint variational problem can be realized, and finally the extended Lagrange form is obtained as follows: (3-4) The iteration is updated by the alternating direction multiplier algorithm The saddle point of equation (3-4) is sought, and the iteration method is （3-5） （3-6） （3-7） （3-8） In the formula, n is the number of iterations, , is an update factor, and ε is a positive number greater than 0, representing accuracy. Solving formula (3-8) by using Plancheral Fourier equidistant transformation can obtain （3-9） （3-10） （3-11） The frequency domain of each modal component is updated constantly, and inverse Fourier transform is performed to the time domain.

5. The belt conveyor permanent magnet synchronous motor fault identification and analysis method according to claim 1, characterized in that, The optimization principle of step S3-2 comprises: Adopting a speed maintaining mechanism for adaptive speed adjustment, as shown in formula (3-16) （3-16） wherein is the velocity of the i-th particle at time t, and is a learning factor, ω is the angular frequency of the power supply, r1 is the inner diameter of the bearing, r2 is the outer diameter of the bearing, is a random number in the interval [0, 1], is a velocity maintenance probability threshold; Introducing a speed pause coefficient a as a key control factor of particle motion characteristics, when a>1, the particle follows the speed update norm as shown in formula (3-17) （3-17） When a≤1, the particle speed is constrained, showing an approximate constant motion characteristic, that is, （3-18） A double population co-evolution strategy is adopted to improve the overall optimization performance, the first population N1 adopts the speed pause particle swarm optimization algorithm, and the second population N2 adopts the global optimal position based movement without speed calculation, i.e. （3-19） X i (t) is the displacement of the i-th particle at time t, f r is the synchronous frequency of the permanent magnet synchronous motor, a(t) is the acceleration response, and r is the stator radius; The envelope spectrum peak factor E is adopted c As a fitness function, the signal envelope spectrum amplitude sequence is assumed to be then （3-20）； wherein, is the amplitude of the envelope spectrum in the frequency range, is the amplitude of the envelope spectrum in the frequency range, is greater than the rotational frequency of the permanent magnet synchronous motor; is the maximum fault frequency of the drive system, is greater than the power supply fundamental frequency, f r is the rotational frequency of the permanent magnet synchronous motor, f i is the rotational frequency of the rolling element The optimization step includes: S3-2-1. Define parameters and initialize particle position and velocity; S3-2-2. Update the velocity of the particles in the first population according to formula (3-16); S3-2-3. Update the position of the particles in the first population according to formula (3-17); S3-2-4. Update the position of the particles in the second population according to formula (3-18); S3-2-5. Evaluate the fitness of all particles; S3-2-6. Update the individual optimal and global optimal of the particles in the first population; S3-2-7. Update the global optimal of the particles in the second population; S3-2-8. Determine whether the maximum number of iterations is reached; if not, loop the subsequent steps from step S3-2-2; if yes, proceed to step S3-2-9; S3-2-9. Return the global optimal.

6. The belt conveyor permanent magnet synchronous motor fault identification and analysis method according to claim 1, characterized in that, The continuous wavelet transform adopts a Morlet wavelet base function, which is defined as: （4-1） In the formula, t is time, and ω0 is an angular frequency.

7. The belt conveyor permanent magnet synchronous motor fault identification and analysis method according to claim 1, characterized in that, The construction of the Markov transition field includes dividing the time series into Q quantile boxes, calculating the transition matrix W and the Markov transition matrix M.

8. The belt conveyor permanent magnet synchronous motor fault identification and analysis method according to claim 1, characterized in that, In the step S5, the vibration signal image and the current signal image are vertically spliced to form a dual-channel fusion image, the image size is unified to 224x224 pixels, and normalization processing is performed.

9. The belt conveyor permanent magnet synchronous motor fault identification and analysis method according to claim 1, characterized in that, In the step S7, a ReLU activation function is used after each convolutional layer, a maximum pooling is used in the pooling layer, a Dropout layer is introduced after the full connection layer, and the dropout rate is 0.5.

Images