Motor Fault Detection Method Based on Group Sparse Auto-Encoding and Swarm Intelligence

Abstract

The present invention discloses a method for motor fault detection based on group sparse auto - encoding and swarm intelligence, including: (1) For the DC motor fault signal, the output current of the tension controller is used to control the magnetic powder brake to adjust the motor load torque, and the motor fault signal is collected; (2) Establish a framework for implementing a group sparse auto - encoder and solve it using the M - M method. The sparse feature extraction network is used to extract features from the original input motor fault data; (3) Improve the swarm intelligence algorithm and propose the SS - PSO algorithm; (4) Establish an SS - PSO - ANN deep classifier model; (5) For the high - quality sparse features extracted by the network based on the group sparse auto - encoder implementation framework, the SS - PSO - ANN classifier is used for effective diagnosis of motor faults. The present invention aims at the common faults of inter - turn short - circuit and rotor demagnetization of DC motors, realizes the autonomous fault diagnosis of motors, and thus avoids the situation that the operation efficiency of the motor is affected by faults or even accidents are caused by motor damage.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault diagnosis, and particularly relates to a motor fault detection method based on group sparse auto-encoding and swarm intelligence. Background Art

[0002] Motors have become the most widely used basic power equipment in industrial production. Motor faults not only affect production, but may also cause major safety accidents. Therefore, how to timely diagnose and eliminate faults, prevent accidents, and ensure the safe, reliable, and efficient operation of motors is particularly important.

[0003] Traditional fault diagnosis methods include methods based on wavelet packet analysis, eigenvector methods, or methods based on Hilbert-Huang transform, etc. However, due to the complex fault mechanism of motors, the fault characteristic signals are not obvious, and the fault signals are non-linear and non-stationary, making the above traditional feature extraction methods all have some defects. For example, the calculation of the eigenvector method is too complex and overly dependent on prior knowledge. Signal processing methods such as Hilbert-Huang transform or wavelet transform have certain constraints on their generalization ability in the extraction of fault characteristics and can no longer meet the requirements of current motor fault monitoring. With the development of artificial intelligence, deep learning has become an important branch in the research of machine learning algorithms. The deep learning model is a model based on the hierarchical structure of multi-layer neural networks, which can automatically learn from big data and express the essential and implicit laws of data features. For the problems in the feature extraction of motor fault signals, deep learning can provide a better solution. Summary of the Invention

[0004] Object of the Invention: Aiming at the problems existing in the prior art, the present invention provides a motor fault detection method based on group sparse auto-encoding and swarm intelligence, which overcomes the influence of people's subjective thinking on the maintenance process and fault diagnosis basis in fault diagnosis, and improves the accuracy and efficiency of fault diagnosis.

[0005] Technical Solution: The present invention provides a motor fault detection method based on group sparse auto-encoding and swarm intelligence, including the following steps:

[0006] Step 1, collect fault signals from the motor, including DC bus current and phase current under three conditions of normal motor state, inter-turn short circuit of the winding, and local demagnetization of the rotor, and establish a fault case library based on the fault data;

[0007] Step 2, establish a realization framework of a group sparse auto-encoder (GSAE) and use the Majorization-Minimization (M-M) method to solve it, and use this sparse feature extraction network to extract features from the original input motor fault data; The specific realization framework of the group sparse auto-encoder is as follows:

[0008] 1) Set input parameters and categories

[0009] Let X be the input data, then:

[0010] X = {X1,..., X c}

[0011] Among them, X1 is the first class (class1), X2 is the second class (class2), c is the number of classes, {n1, n2,... n c} is the number of data points in each class. The rule of data organization is: the columns belonging to class 1 appear first, then the data columns of class 2, and so on, until the data columns of the last class c;

[0012] 2) Define the loss function

[0013] Introduce the l 2，1 -norm based on regularization to define the loss function, which is defined as follows:

[0014]

[0015] Among them, ||·|| 2，1 = ∑ j ||Z j→ ||2 is the sum of the rows of the l2-norm, specified by j; φ is a non-linear activation function, W and U are the encoding and decoding weights respectively, and λ is a parameter. Therefore, the l 2，1 -norm of the second term applies to two cases, namely φ(WX) or only to WX;

[0016] Step 3, improve the swarm intelligence algorithm to implement the particle swarm optimization algorithm (SS-PSO) that combines its own and social factors; the SS-PSO algorithm randomly uses the optimal positions g best of other subgroups, and retains the optimal position g best of each subgroup itself, and at the same time uses the two optimal positions for updating the velocity and optimal position of the particles in the subgroup;

[0017] Step 4, establish an SS-PSO-ANN deep classifier model, and use the improved particle swarm optimization algorithm SS-PSO to optimize the weights and thresholds of the ANN network;

[0018] Step 5, for the high-quality sparse features extracted based on the GSAE network, use the SS-PSO-ANN classifier to effectively diagnose motor faults.

[0019] Further, the specific operation of using the Majorization-Minimization (M-M) method to solve sparse auto-encoding in step 2 is as follows:

[0020] 1) Obtain the decoding weight U, and use the closed-form linear least squares regression problem to obtain the decoding weight U:

[0021]

[0022] where k represents the number of iterations, that is, the k-th iteration;

[0023] 2) Obtain the encoding weight W, and use the M-M algorithm to obtain the encoding weight W:

[0024]

[0025] 2.1) Construct the smoothing function G0(W):

[0026] Let J(W) be the minimization objective function. For the initial point ω0, construct the smoothing function G0(W) through ω0. Its value is larger than J(W) when far from ω0, and the same at the point ω0, that is, construct the smoothing function G0(W) that is easy to minimize; in each iteration, G k (W) is minimized to obtain the next iteration value, thereby defining G k (Z) of the actual loss function J(W), as follows:

[0027]

[0028] where a is the largest eigenvalue of the matrix , and I is the identity matrix. By simplifying G k (Z), we can get:

[0029]

[0030] 2.2) Rewrite the optimization function

[0031] Let:

[0032] The optimization function is rewritten as:

[0033] 2.3) Replace the non-linear problem with gradient descent of a simple linear problem

[0034]

[0035] In the formula, σ is the step size of gradient descent;

[0036] Let: V be Xc T The row correlation degree, D = diag(|VW T | -1 ), solve the coding weight W: W T = P - V T T.

[0037] Furthermore, the complete update formula of the SS - PSO algorithm is as follows:

[0038]

[0039] Among them, β1 = 1 - η, β2 = η, η is a random number between (0, 1); β1 is a linearly decreasing function less than 1; β2 is a linearly increasing function less than 1; w is called the inertia weight, c1 and c2 are called learning factors, r1, r2 ∈ (0, 1) are two independent random numbers, v i (t), v i (t + 1) is the velocity of the i - th particle at the t - th and (t + 1) - th iterations, x i (t), x i (t + 1) is its position at the t - th and (t + 1) - th iterations, p best is the optimal position of the particle, g best is the optimal position of the subgroup itself;

[0040] At the same time, its fitness function is defined as:

[0041]

[0042] Among them, O is the target output vector, and Y is the actual output vector of the network.

[0043] Furthermore, the specific steps of the SS - PSO - ANN deep classifier model in step 4 are as follows:

[0044] Step 4.1: Construct and initialize a three - layer BP neural network. The number of network layers is designed to be 3 layers, namely the input layer, output layer, and hidden layer; design the number of input layer nodes to be 4, representing the motor bus current and three - phase currents collected by the Hall current sensor, as well as the torque and speed measured in real - time by the speed - torque meter in cooperation with the speed - torque sensor; determine the number of hidden layer nodes to be 10; based on the application object of motor fault diagnosis, that is, diagnosing three types: normal motor state, inter - turn short - circuit of the winding, and local demagnetization of the rotor, design the number of output nodes to be 3;

[0045] Step 4.2: Initialize the particle swarm, determine the acceleration coefficients c1, c2, the inertia factor ω, determine the number of particles, the number of iterations, design the values of r1, r2 to be random numbers between [0, 1], and the particle dimension d;

[0046] The dimension of the particle is: d = p + n2 + q + n3

[0047] Where p is the number of connection weights between the input / hidden layer; q is the number of connection weights between the hidden layer / output layer; n2 is the number of threshold values in the hidden layer; n3 is the number of threshold values in the output layer;

[0048] Step 4.3: Calculate the individual optimal value pi t , the global optimal value pg t , and the subgroup optimal value pr t ;

[0049] Step 4.4: Update the current velocity and position of the particle, and update the velocity v of each particle using the SS-PSO algorithm update formula t+1 and the position x t+1 information;

[0050] Step 4.5: Update the optimal value, and compare the current optimal value with the global optimal value pg t , the individual optimal value pi t and the subgroup optimal value pr t according to the SS-PSO algorithm fitness function formula. If the current optimal value is better than any of these parameters, replace it;

[0051] Step 4.6: If the number of iterations k is greater than the maximum number of iterations k1 or the evaluation error value is greater than the given value, the program stops iterating and goes to Step 4.7. Otherwise, the program goes to the new round of particle state update, that is, Step 4.4;

[0052] Step 4.7: Save the global optimal value of this group. Compare the global optimal particle positions of each group according to the fitness function, and map the global optimal position of the group with the best position to the neural network weight and threshold values.

[0053] Furthermore, when calculating the individual optimal value pi t , the global optimal value pg t , and the subgroup optimal value pr t in Step 4.3, first initialize the initial positions and velocities of each particle as random numbers. Substitute the connection weights or threshold values represented by different dimensions of each particle into the structure parameter calculation formula of the BP neural network, calculate the outputs of the hidden layer nodes and output layer nodes of the network respectively, calculate the mean square error between the actual output and the target output of the network through the fitness function, and finally obtain the initialized global optimal point pg t , the individual optimal point pi t and the subgroup optimal point pr t through comparison.

[0054] Furthermore, Step 1 includes:

[0055] (1) The braking torque of the magnetic powder brake is controlled by the output current of the tension controller;

[0056] (2) The Hall current sensor is used to collect current signals, and the rotational speed torque sensor is used to measure the real-time torque and rotational speed;

[0057] (3) The data acquisition card is combined with the Hall current sensor to collect the motor bus current and three-phase current, and the rotational speed torque meter is used in cooperation with the rotational speed torque sensor to measure the real-time torque and rotational speed.

[0058] Beneficial effects:

[0059] 1. The present invention studies the Sparse Auto-Encoder (SAE) model in the deep learning model, improves the swarm intelligence algorithm, proposes the Particle Swarm Optimization algorithm considering both self and social factors (SS-PSO), and establishes the SS-PSO-ANN deep classifier. For the high-quality sparse features extracted based on the (GSAE) network, the SS-PSO-ANN classifier is used for effective diagnosis of motor faults. The deep learning model used can automatically perform feature learning, discover the deep-seated fault essential features that are not easily extracted, thereby improving the diagnosis accuracy, solving the complexity of manually selecting features in the traditional network, and changing the disadvantage of the separation of feature extraction and classification recognition.

[0060] 2. The present invention realizes the autonomous fault diagnosis ability of the motor for the common faults of inter-turn short circuit and rotor demagnetization of the DC motor, thereby avoiding the situation that the operation efficiency of the motor is affected by the fault or even the motor is damaged causing accidents. Using this algorithm to extract and learn the deeper fault features of the motor, thereby realizing the fine diagnosis and on-line monitoring of motor faults, thus eliminating the necessity of manual feature selection and better realizing the automation of motor fault diagnosis. Description of the drawings

[0061] Figure 1 is a schematic diagram of the motor model fault diagnosis process;

[0062] Figure 2 is a schematic diagram of the iterative solution process of the M-M method;

[0063] Figure 3 is a schematic diagram of particle position update;

[0064] Figure 4 is a schematic diagram of the implementation process of the SS-PSO algorithm;

[0065] Figure 5 is the test function y = 1 - cos(3x)e (-x) curve graph;

[0066] Figure 6It is a schematic diagram of the initialization of the function optimized by the SS-PSO algorithm;

[0067] Figure 7 It is the diagram of the first position update of the function optimized by the SS-PSO algorithm;

[0068] Figure 8 It is the diagram of the second position update of the function optimized by the SS-PSO algorithm;

[0069] Figure 9 It is the diagram of the 21st position update of the function optimized by the SS-PSO algorithm;

[0070] Figure 10 It is the diagram of the final (30th) position update of the particle swarm of the function optimized by the SS-PSO algorithm;

[0071] Figure 11 It is the flow chart of the ISS-PSO-ANN training algorithm;

[0072] Figure 12 It is a schematic diagram of the motor fault diagnosis and prediction based on the SS-PSO-ANN deep classifier;

[0073] Figure 13 It is a schematic diagram of the accuracy rate of the learning model on the training set and the test set. Specific implementation manners

[0074] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and cannot be used to limit the protection scope of the present invention.

[0075] The present invention proposes a motor fault detection method based on group sparse autoencoder and swarm intelligence. By establishing an implementation framework of a group sparse autoencoder (GSAE) and using the M-M method to solve, the feature extraction of the original input motor fault data is realized; a particle swarm algorithm (SS-PSO) that combines its own and social factors is proposed; an SS-PSO-ANN deep classifier model is established; the SS-PSO-ANN classifier is used for effective diagnosis of motor faults. The present invention can automatically learn the inherent features of motor fault data, excavate the deep abstract features of the data and use them as the feature quantities for motor fault diagnosis, thereby realizing the autonomous fault intelligent diagnosis of the motor, reducing human intervention, and improving the automation degree of the motor fault diagnosis process.

[0076] This embodiment is based on a short-circuit fault, demagnetization fault data and normal state data of a DC motor. A disclosed motor fault detection method based on group sparse autoencoder and swarm intelligence includes the following steps:

[0077] Step 1: Collect fault signals from the motor. Among them, the output current of the tension controller is used to control the magnetic powder brake to adjust the load torque of the motor, and signal acquisition is carried out by adjusting the change of the load torque, including the DC bus current and phase current in three cases: normal state of the motor, inter-turn short circuit of the winding, and partial demagnetization of the rotor, and a fault case library based on fault data is established.

[0078] Use the output current of the tension controller to control the braking torque of the magnetic powder brake; use a Hall current sensor to collect current signals, and use a speed-torque sensor to measure real-time torque and speed; use a data acquisition card combined with a Hall current sensor to collect the motor bus current and three-phase current, and use a speed-torque meter in cooperation with a speed-torque sensor to measure real-time torque and speed.

[0079] Step 2: Establish a group sparse autoencoder (GSAE) implementation framework, and use the Majorization-Minimization (M-M) method to solve it. Use this sparse feature extraction network to extract features from the original input motor fault data.

[0080] Step 2.1: Establishment of the group sparse autoencoder framework:

[0081] The present invention proposes a new stack sparse autoencoder modeling method to retain the sparsity of the group. During the learning process, the features of a single category have the same sparse signature, that is, the non-zero values in the features occur at the same positions in the same category, which is obtained here by combining l 2，1 -norm regularization. The implementation process of the proposed algorithm is as follows:

[0082] 1) Set input parameters and categories

[0083] Let X be the input data, then:

[0084] X = {X1,..., X c} (1)

[0085] Among them, X1 is the first class (class1), X2 is the second class (class2), which correspond to various fault types of the motor described above in this patent; c is the number of classes, {n1, n2,... n c} is the number of data points in each class. The rule of data organization is: the columns belonging to class 1 appear first, then the data columns of class 2, and so on, until the data columns of the last class c.

[0086] 2) Define the loss function

[0087] In the learning of a single-layer generative autoencoder model, the loss function is defined as:

[0088]

[0089] Among them, φ is a non-linear activation function, such as the sigmoid function, and W and U are the encoding and decoding weights respectively. High-order representations can be achieved by stacking multiple layers and through intelligent training. R(W) is a regularization function, controlled by the parameter λ. In the learning of the weight matrix, an additional constraint term is introduced to avoid overfitting. Common regularization functions include:

[0090] (A) LASSO or l1-norm, which enforces sparse learning of the weights;

[0091] (B) Euclidean or l2-norm, which adds a higher penalty to the peak weights to enforce diffusive learning of the weights;

[0092] (C) Elastic net or l1 + l2-norm, which adds the two norms in the optimization;

[0093] In the group sparse autoencoder framework proposed in this patent, a regularization-based l 2，1 -norm is introduced to define the loss function, which is defined as follows:

[0094]

[0095] Among them, ||·|| 2，1 = ∑ j ||Z j→ ||2 is the sum of the rows of the l2-norm (specified by j). The internal l2-norm promotes the generation of a dense (non-zero) solution within the selected rows; the external l1-norm (sum) enforces sparsity when selecting rows. In the improved formula, regularization enforces group sparsity by adding a constraint (i.e., features in the same group / class have similar sparse signatures).

[0096] Note that φ(·) is an element-wise applied clipping function. Therefore, the l 2，1 -norm of the second term applies to both cases, i.e., φ(WX) or only to WX, because both cases promote row sparsity. This enables obtaining optimized supervision when considering that the information of the class labels is used in training. However, no discriminatory measures are implemented on the attribute features, because it is not enforced that features in different groups have different sparse signatures.

[0097] Step 2.2: Use the Majorization-Minimization (M-M) method to solve the sparse autoencoding:

[0098] The objective function in formula (3) is a non-convex optimization problem, which can be solved by the Majorization-Minimization (M-M) algorithm. The M-M algorithm is an important method in the field of optimization. It is more of an algorithm framework than a specific algorithm because many specific algorithms can be deduced as the M-M framework. This method requires two-step iteration. By continuously finding the optimal solutions of approximate problems, it approaches the solution of the original problem. Mathematically speaking, the core idea of the M-M algorithm is successive upper bound minimization. By designing a series of approximate optimization functions to control the upper bound of the original function, it converges to the optimal solution of the original objective through minimizing the sequence.

[0099] In this patent, it is used to solve for W and U, that is, the encoding and decoding weights of the sparse autoencoder. When using the M-M method to solve, at any k th th iteration, the solution to the above non-convex problem can be divided into the following two steps:

[0100] 1) Obtain the decoding weight U

[0101] Use a closed-form linear least squares regression problem to obtain the decoding weight U:

[0102]

[0103] 2) Obtain the encoding weight W

[0104] Use the M-M algorithm to obtain the encoding weight W:

[0105]

[0106] 2.1) Construct the smoothing function G0(W).

[0107] Let J(W) be the minimized objective function. For the initial point ω0, construct the smoothing function G0(W) through ω0. Its value is larger than J(W) when far from ω0, and has the same value at the point ω0, that is, construct a smoothing function G0(W) that is easy to minimize. In each iteration, G k (W) is minimized to obtain the next iteration value. Thus, the result of each iteration will be closer to the actual result.

[0108] Formula (4) can be rewritten as:

[0109]

[0110] In the formula, Z c = WX c , Z is obtained by stacking Z c in columns. In this optimization problem, only the least squares regression term needs to be optimized, and the penalty condition does not affect. In the minimization step, define Gk (Z) is as follows:

[0111]

[0112] Here, a is the largest eigenvalue of the matrix , I is the identity matrix, and simplify G k (Z) can be obtained as:

[0113]

[0114]

[0115] 2.2) Rewrite the optimization function

[0116] Let:

[0117] Equation (8) can be written as:

[0118] In the formula, ε is a constant term.

[0119] Using the formula: Equation (9) can be rewritten as:

[0120]

[0121] Remove the constant term and rewrite W, then the optimization function can be written as:

[0122]

[0123] 2.3) Replace the non - linear problem with gradient descent for a simple linear problem:

[0124] All matrices are represented in transposed form, and the activation function is calculated element - by - element. The steps to replace the above non - linear problem with a simple linear problem for gradient descent are as follows:

[0125]

[0126] In the formula, σ is the step size of gradient descent. Rearrange equation (12) and remove the summation term to get the following formula:

[0127]

[0128] where V is the row correlation degree of X c T Taking the derivative of formula (13) and setting it to zero, we can get:

[0129]

[0130] where:

[0131] D = diag(|VW T | -1 )

[0132]

[0133] Using the principle of matrix inversion, we can obtain:

[0134]

[0135] Let: Then equation (16) becomes W T = P - V T T

[0136] In addition, by adding cT (c is the largest eigenvalue of V T V) to both sides of the equation, and then subtracting VW T VW T from it, the T value can be obtained.

[0137]

[0138] The iterative process of the entire algorithm is as Figure 2 .

[0139] Step 3: Improve the swarm intelligence algorithm to implement the particle swarm optimization algorithm (SS-PSO) that combines self and social factors.

[0140] Implement the improved particle swarm optimization algorithm (SS-PSO). Based on the application object of motor fault diagnosis, the PSO algorithm is improved to propose the particle swarm optimization algorithm (SS-PSO) that combines self and social factors. The core idea of the initial improvement is to randomly use the optimal position g best of other subgroups, and retain the optimal position g best of each subgroup itself. At the same time, the two optimal positions are used to update the velocity and optimal position of the particles in the subgroup. Thus, the global and local search capabilities of the algorithm are improved.

[0141] The initial update formula of the improved particle swarm optimization algorithm is defined as:

[0142]

[0143] It can be seen from equation (18) that the velocity update formula consists of four parts: the first part wv i (t) is the inheritance of the magnitude and direction of the velocity of the particle at the previous moment; the second part c1r1[p best - x i (t)] is the influence of the historical best position of this subgroup on the current position, reflecting the local search ability of the particle; the third part c2r2[gbest -x i (t) represents the influence of the best position of this subgroup on the current position, reflecting the global search ability of the particle; the fourth part c2r2[g best (r)-x i (t) represents the global search ability of other subgroups. It can be seen from this that the particles in the first and second parts consider the influence of their own factors, which is the "cognitive" part, reflecting the global search ability of the particles; the third and fourth parts consider the influence of the social factors of the particles, which is the "social" part, and is the sharing of social information among particles, reflecting the local search ability of the particles.

[0144] For global optimization problems, research shows that: in the initial stage of algorithm iteration, in order to obtain better particles, the particle's own factors are mainly considered, and better particles are obtained through the global search ability; in the later stage of algorithm iteration, in order to accelerate the convergence speed of the algorithm, the influence of the particle's social factors is mainly considered, and better ability is obtained through the local search ability. In order to balance the roles of the particle's optimal position pbest, the subgroup's own optimal position gbest, and the optimal position gbest(r) of other subgroups, the velocity update formula in Equation (18) is rewritten in the following form:

[0145] v i (t + 1) = wv i (t) + β1c1r1[p best -x i (t)] + β2[c2r2[g best -x i (t)] + c2r2[g best (r)-x i (t)]] (19)

[0146] In the formula, β1 is a linearly decreasing function less than 1, so that the population considers the particle's own factors in the initial stage of evolution and selects more ideal particles through its own search ability; β2 is a linearly increasing function less than 1, so that the population considers social factors in the later stage of evolution and accelerates the convergence speed of the algorithm.

[0147] According to the above analysis, the complete formula of the improved PSO algorithm (SS - PSO) is as follows:

[0148]

[0149] Among them, β1 = 1 - η, β2 = η, and η is a random number between (0, 1).

[0150] At the same time, its fitness function is defined as:

[0151]

[0152] Where: O is the target output vector, and Y is the actual output vector of the network.

[0153] The implementation steps of the improved particle swarm optimization algorithm are as follows:

[0154] (1) Initialize the SS-PSO parameters, define the dimensions and ranges of the velocity and position, initialize the velocities and positions of all subgroup particles, and design the flight positions and velocity vectors of the training particles to be two-dimensional. Among them, the rows are the parameters to be learned, and the columns are the particles for training flight, that is:

[0155] X = (X1, X2,..., X n )

[0156] where n is the number of particles, the position of the i-th particle is X i = (X i1 , X i2 ,...X id ) T , the velocity is V i = (V i1 , V i1 ,,...V iD ) T , the individual optimal value is P i = (P i1 , P i2 ,...P iD ) T , the global optimal value is P g = (P g1 , P g2 ,...P gD ) T .

[0157] (2) Calculate the fitness function values of all particles according to formula (21), and update the current velocity and position of each particle through the position and velocity formulas (20) of the particle swarm algorithm;

[0158] (3) Update the individual extreme value and the global extreme value, that is, calculate the optimal positions p best and g best of each subgroup, and store the smallest g best in all subgroups;

[0159] (4) Randomly generate the particle number i from 1 to n;

[0160] (5) If f(x i ) < f(p best ), then update p best of the i-th particle in the subgroup;

[0161] (6) If f(p best ) < f(gbest ) then update \(g\) of the subgroup best ;

[0162] (7) Save the smallest \(g\) among all subgroups best ;

[0163] (8) Repeat (2) - (7), judge the algorithm termination condition until the set maximum number of loop simulation steps or reach the maximum predetermined error;

[0164] The schematic diagram of particle position update is as shown in Figure 3 , and the implementation process of the SS - PSO algorithm is as shown in Figure 4 .

[0165] Specific verification of the SS - PSO algorithm:

[0166] Let the function \(y = 1-\cos(3x)e\) (-x) , and it is as shown in Figure 5 in the interval \([0, 4]\). When \(x = 0.9350 - 0.9450\), the function reaches the maximum value \(y = 1.3706\)

[0167] Under the Matlab environment, use the SS - PSO algorithm to optimize it, and the effects of different optimization times are as shown in Figures 6 to 10 . Figure 6 Select two points in the interval, calculate the corresponding function values, and at the same time, these two points update their positions according to a certain speed, and then calculate the function values corresponding to the new two points to obtain Figure 7 , and so on until the maximum value of the function is obtained and the update stops.

[0168] It can be obtained from these figures that the particles are relatively dispersed at the initial update of the particle swarm. As the number of updates increases, more and more particles concentrate at \(x = 0.9350 - 0.9450\), that is, at the maximum value. The effect is more obvious after 20 updates. After each update, the optimal solution of the function is also more and more accurate. Figure 10 It can be seen from [the figure] that finally all points concentrate at the maximum value.

[0169] Step 4: Establish the SS - PSO - ANN deep classifier model.

[0170] Construct the SS - PSO - ANN deep classifier model. Introduce the improved particle swarm algorithm SS - PSO to optimize the weights and thresholds of the ANN network. The specific implementation process is as follows:

[0171] The first step is to construct and initialize a three - layer BP neural network:

[0172] 1) Design of the number of layers: The number of network layers is designed as 3 layers, namely the input layer, the output layer, and the hidden layer.

[0173] 2) Design of the number of input layer nodes: Based on the application object of motor fault diagnosis, in this embodiment, the number of input layer nodes is designed to be 4, representing the motor bus current and three-phase current collected by the Hall current sensor, as well as the torque and speed measured in real time by the speed-torque meter in cooperation with the speed-torque sensor.

[0174] 3) Design of the number of hidden layer nodes

[0175] Too many hidden layer nodes will cause a decline in the network generalization ability and is prone to overfitting. However, if the number of hidden layer nodes is too small, the network is not easy to converge. In this embodiment, the number of hidden layer nodes is determined to be 10 through simulation experiments.

[0176] 4) Design of the number of output layer nodes

[0177] Based on the application object of motor fault diagnosis, that is, diagnosing three types: normal state of the motor, inter-turn short circuit of the winding, and local demagnetization of the rotor, so the number of output nodes is designed to be 3 in this embodiment.

[0178] The second step is to initialize the particle swarm:

[0179] 1) Design of the acceleration coefficients c1 and c2: It is determined through simulation that c1 = c2 = 2.05.

[0180] 2) Design of the inertia factor ω: It is determined through simulation that ω = 0.9.

[0181] 3) Design of the number of particles: It is determined through simulation that the number of particles N = 160.

[0182] 4) Design of the number of iterations: It is determined through simulation that the number of iterations: k1 = 100.

[0183] 5) Design of r1 and r2: According to empirical data, the design values are random numbers between [0, 1].

[0184] 6) Design of the particle dimension: Based on the processing objective of optimizing the weights and thresholds of the BP network using SS-PSO, the particle dimension d is designed as: d = p + n2 + q + n3.

[0185] Among them, p is the number of input / hidden layer connection weights; q is the number of hidden layer / output layer connection weights; n2 is the number of hidden layer thresholds; n3 is the number of output layer thresholds.

[0186] The third step is to use the fitness function to calculate the individual optimal value, pi t and the global optimal value pg t and the subgroup optimal value pr t .

[0187] During implementation, first initialize the initial positions and velocities of each particle as random numbers. Substitute the connection weights or thresholds represented by different dimensions of each particle into the calculation formula of the structural parameters of the BP neural network, calculate the outputs of the hidden layer nodes and output layer nodes of the network respectively, calculate the mean square error between the actual output and the target output of the network through the fitness function. Finally, obtain the initialized global optimal point pg through comparison t , individual optimal point pi t and subgroup optimal point pr t .

[0188] The fourth step is to update the current velocity and position of the particle.

[0189] Update the velocity v of each particle using the improved particle swarm optimization algorithm update formula (20) t+1 and position x t+1 information.

[0190] The fifth step is to update the optimal value.

[0191] Compare the current optimal value with the global optimal value pg t , individual optimal value pi t and subgroup optimal value pr t according to the fitness function formula (21) of the improved particle swarm optimization algorithm. If the current optimal value is better than any of these parameters, replace it.

[0192] The sixth step is to check the end condition.

[0193] If the number of iterations k is greater than the maximum number of iterations k1 or the evaluation error value is greater than the given value, the program stops iterating and goes to the seventh step. Otherwise, the program goes to the new round of particle state update (the fourth step).

[0194] The seventh step is to save this set of global optimal values.

[0195] The eighth step is to compare the global optimal particle positions output by each group according to the fitness function, and map the global optimal position of the group with the best position to the weights and thresholds of the neural network.

[0196] The flow of the ISS-PSO-optimized neural network training algorithm is as Figure 11 .

[0197] Step 5: For the high-quality sparse features extracted based on the GSAE network, use the SS-PSO-ANN classifier to effectively diagnose motor faults.

[0198] The specific process of using the constructed SS-PSO-ANN deep classifier model to diagnose and predict motor faults is as Figure 12 .

[0199] Use a sparse feature extraction network to extract features from the original input data. Then, based on the high-quality sparse features extracted by the network, use the constructed SS-PSO-ANN deep classifier model to predict motor faults. In implementation, first, the collected motor fault signals are trained through sparse autoencoders to obtain parameters. The improved particle swarm optimization algorithm (SS-PSO) is used to optimize the neural network, and the SS-PSO-ANN deep classifier model is constructed. The classifier is trained using the collected motor fault data. Finally, the trained model is used to identify motor faults, thereby realizing effective diagnostic prediction of motor fault signals using this classification model. The experimental results show that: compared with traditional diagnostic methods, the proposed algorithm has higher accuracy and stability.

[0200] In the implementation process, use the unsupervised learning algorithm of sparse autoencoders to pre-train the sample data. The optimal feature expression weight values obtained from the hidden layer: the decoding weight U and the encoding weight W are used to initialize the deep neural network. Then, the deep neural network is trained and optimized to further improve the local optimum problem and improve the classification accuracy and training efficiency of the model.

[0201] Using a motor fault detection method based on group sparse autoencoders and swarm intelligence proposed in this patent, first use a fault signal acquisition platform, cooperate with a data acquisition card and a Hall current sensor to collect the motor bus current and three-phase current, and use a speed torque meter and a speed torque sensor to measure the real-time torque and speed to obtain motor fault signals, establish a database, and then use the group sparse autoencoders and swarm intelligence deep learning model to complete the diagnosis and prediction of motor faults. In implementation, the motor output state types include three: normal, inter-turn short circuit fault, and rotor demagnetization fault. The motor working state labels are shown in Table 1. In the implementation process of model establishment, first, design Experiment 1 to obtain the accuracy distribution of this deep learning model on the training set and the test set as Figure 13 shown. Select 12 groups of data, representing normal motor, inter-turn short circuit fault, and rotor demagnetization fault respectively, to obtain the model prediction output results shown in Table 2. It can be seen that the model of this method can stably and accurately identify motor fault signals. Finally, compare the model of this method (SAE+SS-PSO-ANN) with several common fault diagnosis methods such as support vector machine (Support vector machine, SVM), BP, and sparse autoencoder (SAE) to obtain the results shown in Table 3. The above experimental results once again prove the effectiveness of the deep learning model described in this patent in motor fault diagnosis. Therefore, the method of this patent can avoid situations where the operation efficiency of the motor is affected by faults or even accidents caused by motor damage.

[0202] Table 1 Motor working state labels

[0203] Motor operating state Encoding Label Normal [0 0 1] 0 Short - circuit fault [0 1 0] 1 Demagnetization fault [1 0 0] 2

[0204] Table 2 Results of Model Prediction Labels and True Labels

[0205]

[0206]

[0207] Table 3 Comparison of the Effects of Different Fault Diagnosis Methods

[0208] Experiment number Model type Diagnostic accuracy rate 1 SAE 85％ 2 SVM 90.2％ 3 BP 68.5％ 4 SAE + SS - PSO - ANN 95％

[0209] The above embodiments are only for illustrating the technical concept and characteristics of the present invention, and the purpose is to enable those who are familiar with this technology to understand the content of the present invention and implement it accordingly, and it cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit and essence of the present invention should be covered within the protection scope of the present invention.

Claims

1. A motor fault detection method based on group sparse auto - encoding and swarm intelligence, characterized in that It includes the following steps: Step 1: Collect fault signals from the motor, including the DC bus current and phase current under three conditions: normal motor state, inter-turn short circuit of the winding, and local demagnetization of the rotor, and establish a fault case library based on the fault data; Step 2: Establish a Group Sparse Autoencoder (GSAE) implementation framework and solve it using the Majorization-Minimization (M-M) method. Use this sparse feature extraction network to extract features from the original input motor fault data; The specific implementation framework of the Group Sparse Autoencoder is as follows: 1) Set input parameters and categories Let X be the input data, then: X = {X1, …, X c} Among them, X1 is the first class (class1), X2 is the second class (class2), c is the number of classes, {n1, n2, … n c} is the number of data points in each class. The rule for data organization is that the columns belonging to class 1 appear first, followed by the data columns of class 2, and so on, until the data columns of the last class c; 2) Define the loss function Introduce the regularization-based l 2,1 -norm to define the loss function, which is defined as follows: where, ||·|| 2,1 = ∑ j |||Z j→ |||2 is the sum of the rows of the l2-norm, specified by j; φ is a non-linear activation function, W and U are the encoding and decoding weights respectively, and λ is a parameter. Therefore, the l 2,1 -norm of the second term applies to two cases, namely φ(WX) or only to WX; Step 3: Improve the swarm intelligence algorithm to implement the social and self particle swarm optimization algorithm (SS-PSO); the SS-PSO algorithm randomly uses the optimal position g of other subgroups best and retains the optimal position g of each subgroup itself best ; meanwhile, the two optimal positions are used to update the velocity and optimal position of the particles in the subgroup The complete update formula of the SS-PSO algorithm is as follows: Among them, β1 = 1 - η, β2 = η, where η is a random number between (0, 1); β1 is a linearly decreasing function less than 1; β2 is a linearly increasing function less than 1; w is called the inertia weight, c1 and c2 are called learning factors, r1, r2 ∈ (0, 1) are two independent random numbers, v i (t), v i (t + 1) is the velocity of the i-th particle at the t-th and (t + 1)-th iterations, x i (t), x i (t + 1) is its position at the t-th and (t + 1)-th iterations, p best is the optimal position of the particle, g best is the optimal position of the subgroup itself; At the same time, define its fitness function as: where, O is the target output vector, and Y is the actual output vector of the network; Step 4: Establish an SS-PSO-ANN deep classifier model, and use the improved particle swarm algorithm SS-PSO to optimize the weights and thresholds of the ANN network; Step 5: For the high-quality sparse features extracted based on the GSAE network, use the SS-PSO-ANN classifier to effectively diagnose motor faults.

2. The motor fault detection method based on group sparse auto - encoding and swarm intelligence according to claim 1, wherein, The specific operation of using the Majorization-Minimization (M-M) method to solve sparse autoencoding in Step 2 is as follows: 1) Obtain the decoding weight U, and use the closed-form linear least squares regression problem to obtain the decoding weight U: where, k represents the number of iterations, that is, the kth iteration; 2) Obtain the encoding weight W, and use the M-M algorithm to obtain the encoding weight W: 2.1) Construct the smoothing function G0(W): Let \(J(W)\) be the objective function to be minimized. For the initial point \(\omega_0\), construct a smoothed function \(G_0(W)\) through \(\omega_0\). Its value is larger than \(J(W)\) when far away from \(\omega_0\), and has the same value at the point \(\omega_0\), that is, construct a smoothed function \(G_0(W)\) that is easy to minimize; in each iteration, \(G\) k (W) is minimized to obtain the next iteration value, thereby defining \(G\) k (Z) of the actual loss function \(J(W)\) as follows: where a is the maximum eigenvalue of the matrix , I is the identity matrix, and by simplifying G k (Z), we can obtain: 2.2) Rewrite the optimization function Let: The optimization function is rewritten as: 2.3) Replace the non-linear problem with the gradient descent of a simple linear problem wherein, σ is the step size of gradient descent; Let: Let V be the row correlation of X c T and D = diag(|VW T | -1 ). Solve for the coding weight W: W T = P - V T T.

3. The motor fault detection method based on group sparse auto-encoding and swarm intelligence according to claim 1, characterized in that, The specific steps of the SS-PSO-ANN deep classifier model in Step 4 are as follows: Step 4.1: Construct and initialize a three-layer BP neural network. The number of network layers is designed to be 3 layers, namely the input layer, output layer, and hidden layer; Design the number of input layer nodes to be 4, representing the motor bus current and three-phase current collected by the Hall current sensor, as well as the torque and speed measured in real time by the speed torque meter in cooperation with the speed torque sensor; Determine the number of hidden layer nodes to be 10; Based on the application object of motor fault diagnosis, that is, to diagnose three types: normal motor state, inter-turn short circuit of the winding, and local demagnetization of the rotor, design the number of output nodes to be 3; Step 4.2: Initialize the particle swarm, determine the acceleration coefficients c1, c2, the inertia factor ω, determine the number of particles, the number of iterations, design the values of r1 and r2 to be random numbers between [0, 1], and the particle dimension d; The particle dimension is: d = p + n2 + q + n3 where, p is the number of input / hidden layer connection weights; q is the number of hidden layer / output layer connection weights; n2 is the number of hidden layer thresholds; n3 is the number of output layer thresholds; Step 4.3: Calculate the individual optimal value pi using the fitness function t , the global optimal value pg t , and the subgroup optimal value pr t ; Step 4.4: Update the current velocity and position of the particles, and update the velocity v of each particle using the update formula of the SS-PSO algorithm t+1 and the position x t+1 information; Step 4.5: Update the optimal value, and compare the current optimal value with the global optimal value pg according to the fitness function formula of the SS-PSO algorithm t , the individual optimal value pi t , and the subgroup optimal value pr t to determine the relationship among the three. If the current optimal value is better than any of these parameters, replace it Step 4.6: If the number of iterations k is greater than the maximum number of iterations k1 or the evaluation error value is greater than the given value, the program stops iterating and goes to Step 4.7, otherwise, the program goes to the new round of particle state update, that is, Step 4.4; Step 4.7: Save this group of globally optimal values, compare the globally optimal particle positions of each group according to the fitness function, and map the globally optimal position of the group with the optimal position to the weights and thresholds of the neural network.

4. The motor fault detection method based on group sparse auto - encoding and swarm intelligence according to claim 3, characterized in that, In step 4.3, calculate the individual optimal value pi t , the global optimal value pg t , and the subgroup optimal value pr t : First, initialize the initial positions and velocities of each particle as random numbers. Substitute the connection weights or thresholds represented by different dimensions of each particle into the calculation formula of the structural parameters of the BP neural network, calculate the outputs of the hidden layer nodes and output layer nodes of the network respectively, calculate the mean square error between the actual output and the target output of the network through the fitness function. Finally, obtain the initialized global optimal point pg t , the individual optimal point pi t , and the subgroup optimal point pr t .

5. The motor fault detection method based on group sparse auto - encoding and swarm intelligence according to any one of claims 1 to 4, characterized in that, The said Step 1 includes: (1) Use the output current of the tension controller to control the braking torque of the magnetic powder brake; (2) Use a Hall current sensor to collect current signals, and use a speed-torque sensor to measure the real-time torque and speed; (3) Use a data acquisition card combined with a Hall current sensor to collect the motor bus current and three-phase current, and use a speed-torque meter in cooperation with a speed-torque sensor to measure the real-time torque and speed.

Images