Asynchronous induction motor fault intelligent detection method based on multi-modal time sequence

Abstract

The invention belongs to the technical field of motor fault diagnosis, and particularly relates to an asynchronous induction motor fault intelligent detection method based on a multi-modal time sequence. The method comprises the following steps: acquiring various signals of a motor, including three-phase current, vibration, equipment sound, environment sound and the like, and constructing a multi-modal data set; thirdly, constructing a Transform network classified by a single time sequence, and on the basis of the Transform network, connecting a plurality of the networks in parallel to obtain a lumped Transform network; and then, features of different modal data are dynamically integrated by using a self-attention mechanism, and accurate recognition of a motor fault is realized. In a network training process, a hyper-parameter of the lumped Transform network is optimized by adopting a chaos WOA algorithm; finally, different types of faults of the asynchronous motor can be accurately detected and judged, the situation that different fault features of the motor cannot be accurately represented by a single time sequence is avoided, and the accuracy and reliability of fault detection of the asynchronous motor are guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of motor fault diagnosis, and in particular relates to an intelligent fault detection method for an asynchronous induction motor based on a multimodal time series and an integrated Transformer network. Background Art

[0002] As a widely used driving device in industrial production, the reliability and stability of the motor have a significant impact on the smooth progress of the production process. However, affected by factors such as long-term operation and external environment, the motor may have various fault problems, such as bearing failure, rotor fracture and stator inter-turn short circuit. These fault characteristics are diverse and have a strong nonlinear relationship. Therefore, it is difficult to extract fault features, the data requirement is huge, the algorithm complexity is high, and the accuracy of fault type identification is low. The above studies are all based on single modal data, which may have problems such as incomplete information, feature redundancy and poor fault diagnosis robustness. Therefore, studying multimodal data fault diagnosis methods has become the current focus.

[0003] In the prior art, Chinese patent CN114065804A (application date: October 25, 2021, held by Shanghai University of Electric Power) describes a method and system for diagnosing short-circuit faults in the windings of an electrically excited double-pole motor. The patent uses multi-classification support vector machine (SVM) and convolutional neural network (CNN) to improve the accuracy of fault diagnosis, but may affect the diagnosis results when the data quality is not high. This patent effectively overcomes the above problems by introducing the DS evidence theory algorithm to fuse the diagnostic results, and has better diagnostic reliability and accuracy. Chinese patent CN118381415A (application date: June 25, 2024, held by Rongrun (Zhejiang) Textile Technology Co., Ltd.) describes an independent inverter system and its control method. The patent uses a deep learning optimization control strategy to improve the accuracy of motor control, but there are still challenges in fault detection and processing under complex working conditions.

[0004] In the paper "Broken rotor bar fault diagnosis method for variable frequency powered induction motor using rectification technology" (Wang Wanting, Song Xiangjin, 2022, published in "Transactions of China Electrotechnical Society"), a broken rotor bar fault diagnosis method for variable frequency powered induction motor based on rectification technology is proposed, which effectively improves the recognition accuracy of fault characteristic frequency. However, this technology still has limitations in processing motor current signals under nonlinear loads and variable working conditions, and cannot effectively solve the fault diagnosis problem under complex working conditions. In the paper "Broken rotor bar fault detection of cage asynchronous motor based on ESPRIT and Duffing system" (Xu Boqiang, Sun Liling, 2020, published in "Electric Power Automation Equipment"), a broken rotor bar fault detection method for cage asynchronous motor combining ESPRIT spectrum analysis and Duffing system is proposed to improve the accuracy of fault detection. However, this technology still has limitations in processing non-ideal signals and noise interference in actual motor operation, and cannot effectively solve the problem of fault feature extraction in high noise environment. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an intelligent detection method for asynchronous induction motor faults based on multimodal time series. The method of the present invention combines time series analysis, Transformer network and multimodal data fusion technology. First, multiple signals of the motor are collected, including three-phase current, vibration, equipment sound and environmental sound, etc., to construct a multimodal data set; then, a Transformer network for single time series classification is constructed, and on this basis, multiple networks are connected in parallel to obtain a lumped Transformer network; then, the self-attention mechanism is used to dynamically integrate the features of different modal data to achieve accurate identification of motor faults. During the network training process, the chaotic WOA algorithm is used to optimize the hyperparameters of the lumped Transformer network; finally, the present invention can accurately detect and judge different types of asynchronous motor faults, avoiding the inability of a single time series to accurately characterize the different fault characteristics of the motor, and ensuring the accuracy and reliability of asynchronous motor fault detection.

[0006] The technical solution of the present invention is: an intelligent detection method for asynchronous induction motor faults based on multi-modal time series, characterized in that it comprises the following steps:

[0007] Step 1: The steps of the multimodal data acquisition module, synchronously collect different types of data,

[0008] Step 2: Merge the data into a lumped Transformer network consisting of multiple parallel sets of time series; the specific steps are as follows:

[0009] Step 21: Token embedding; time series Where M and N represent the length and dimension of the time series, C represents the number of channels, and R represents the set of real numbers; then, it is flattened into Cut into a block sequence along the time direction [x 1 ,x 2 ,…,x M ]and Since, the block sequence is unfolded and mapped to D dimensions through a trainable linear projection;

[0010] Step 22: Classify tokens; add a learnable embedding to the chunk sequence [x 1 ,x 2 ,…,x M ], we get a new embedded block sequence z0=[x class ;x 1 E;x 2 E;…;x N E], where x class a classification token representing a sequence of blocks;

[0011] Step 23: Position embedding: Using a learnable 1-D position embedding method, z0 is expanded to obtain an embedding vector sequence z0 = [x class ;x 1 E;x 2 E;…;x N E]+E positon ,in, The resulting embedding vector sequence is used as the input of the encoder;

[0012] Step 24 encoder: use Transformer encoder, take the embedding vector sequence z0 as input; each Transformer block contains multiple self-attention layers, addition & normalization layers and feed-forward neural network layers;

[0013] The feedforward neural network layer is represented by the following formula:

[0014] z l ′=LayerNorm(MSA(z l-1 )+z l-1 ) l=1,2,…,N (1)

[0015] z l =LayerNorm(FFN(z l ' -1 )+z l ' -1 ) l=1,2,…,N (2)

[0016] Where MSA(·) is a multi-self-attention layer; FFN(·) is a feed-forward neural network layer, and the hidden state of each encoder output is It is defined as:

[0017]

[0018] In the formula, β n＝1,2,,N is a hyperparameter;

[0019] Self-attention layer; sparse attention is defined as follows:

[0020]

[0021] In the formula, is the hash function of the key and query, B is the hash bucket, d h Indicates dimension; Q, K, and V represent query, key, and value respectively; represents the attention weight between the i-th query and the j-th key in the sparse attention mechanism; V j Denotes the value of the j-th key vector K j the associated value vector; is satisfied The attention weights are defined by parameterizing the function Implementing a hash function The specific process is as follows:

[0022]

[0023] In the formula, H Q and H K is an arbitrary parameterized function; a single linear layer is applied to obtain the hidden state

[0024] Addition & Normalization Layer: The hidden state output by the self-attention layer Add and keep the original input information in the network;

[0025] Step 25: Classifier; use each time series Transformer to encode the input data to generate feature representation; the classifier maps the features output by the addition & normalization layer to the probability distribution of the category through the feedforward multilayer perceptron and the softmax activation function, as shown below:

[0026]

[0027] In the formula, x is the flattened data; is the hidden layer feature; where W1 is the weight matrix and b1 is the bias vector. The purpose of the GeLU function is to introduce nonlinearity to help the network learn more complex patterns. After the GeLU activation function, the result is multiplied by another weight matrix W2 and added with the bias vector b2.

[0028] Step 3. Steps of 3D chaotic composite mapping;

[0029] Step 4 Chaos-Whale Optimization Algorithm, the implementation steps are as follows:

[0030] Step 4.1 Parameter initialization

[0031] (1) Initialization task dimension D; WOA population size N D ; Maximum number of iterations T; Boundary constraints for different tasks B D ∈[B1,B2] D ;

[0032] (2) Initialize the control parameter c of the modified 3D Logistic-Sine composite mapping; variables x1 = rand(·), x2 = rand(·), x3 = rand(·), where rand(·) is a uniformly distributed random function and rand(·)∈(0,1);

[0033] (3) Initialize the initial position of each WOA in different populations As shown below:

[0034]

[0035]

[0036]

[0037] Where CRand is the spatial distance between the trajectory point of the chaotic composite mapping and the origin (0,0,0), and CRand∈[0,1];

[0038] (4) Initialize the random vector r1, r2, p of WOA, and r1, r2, p = CRand;

[0039] Step 4.2 Hunting process

[0040] (1) Surrounding the prey

[0041] The nonlinear parameter a is used as follows:

[0042]

[0043] Where, t is the current iteration number; Tmax is the maximum iteration number; is a random disturbance, and

[0044] The strategy of chaotic composite mapping is as follows:

[0045]

[0046] In the formula, is the variable of the modified 3D Logistic-Sine composite mapping;

[0047] (2) Hunting stage

[0048] The logarithmic spiral update rule for 3D is as follows:

[0049]

[0050] Where b is a constant that defines the logarithmic spiral shape; l∈[-1,1] is a random number, and the spatial point M(0,0,0) spirals up to the spatial point M(X spiral ,Y spiral ,Z spiral ), and the distance between two points in space is:

[0051]

[0052] The spiral update mechanism of the standard WOA is improved using formula (16) as follows:

[0053]

[0054] The parameter D represents the task dimension, and the random numbers l, X(t+1), X * (t) represents the current optimal value and the previous optimal value of the pigeon respectively;

[0055] (3) Searching for prey

[0056] Using the Cauchy mutation strategy, the mathematical model is as follows:

[0057]

[0058] D Rand The random distance representing the pigeon's position, X CRand( t) The spatial distance between the pigeon’s current position and the origin M(0,0,0), represents the Hadamard operation; Cauchy is the Cauchy operator;

[0059] Step 4.3 Fitness function

[0060] Construct the corresponding fitness function as follows:

[0061]

[0062] Where, Fitness is the fitness function of the E-Transformer network; loss is the loss function during the Transformer network training process; fitness is the fitness function of each independent Transformer network in the E-Transformer network, i=1...5;

[0063] Step 5. Train and test the E-Transformer network based on chaos-whale optimization.

[0064] In order to optimize the hyperparameters of the lumped Transformer network, the present invention introduces the whale optimization algorithm (WOA). WOA has the advantages of effectively searching in the solution space and seeking the global optimal solution or the approximate optimal solution. However, the traditional WOA is prone to fall into the local optimum, and the global search capability and local development capability are unbalanced. Therefore, this paper improves the traditional WOA and introduces the modified 3D-Logistic-Sine composite chaotic mapping to improve the global search capability and convergence of WOA. The improved WOA is applied to optimize the hyperparameters of the lumped Transformer network to achieve excellent classification performance, and then accurately judge the motor fault and its type. The main innovations of the present invention can be summarized as follows:

[0065] (1) Use multiple sensors to collect the motor's three-phase current signal, three-axis vibration signal, two-dimensional sound signal, and one-dimensional environmental noise signal. These multimodal time series signals obtained from different sensors will be directly used as training and testing data sets for the diagnostic model to ensure that the model can fully capture the motor's operating status and potential fault information;

[0066] (2) A time series-based parallel Transformer network, namely the lumped Transformer network, was constructed. This network is specifically used to train and test multi-modal motor fault data, thereby obtaining robust and accurate classification results;

[0067] (3) The whale optimization algorithm is introduced, and combined with 3D chaotic composite mapping to improve the global search ability and convergence speed of the whale optimization algorithm. It is also used to optimize the hyperparameters of the lumped Transformer network to achieve the best classification performance. The beneficial effects of the present invention are: BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The framework diagram of the motor fault diagnosis algorithm.

[0069] Figure 2is a lumped Transformers network.

[0070] Figure 3 Transformers network for a single time series.

[0071] Figure 4 This is the Transformers network structure diagram.

[0072] Figure 5 Chaotic orbit of the modified 3D Logistic-Sine composite mapping.

[0073] Figure 6 is the Lyapunov exponent of the modified 3D Logistic-Sine composite mapping.

[0074] Figure 7 It is a logarithmic spiral update mechanism.

[0075] Name explanation: Transformer: Transformer network; Lyapunov: Lyapunov; Logistic-Sine: Logistic-Sine. DETAILED DESCRIPTION

[0076] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0077] The overall structure of the intelligent fault detection method for asynchronous induction motors based on multimodal time series is as follows: Figure 1 As shown, it mainly includes a multimodal data acquisition module, a lumped Transformer network, a classifier, etc. The lumped Transformer network of the present invention is composed of a plurality of time series Transformer networks in parallel. In view of network performance, the improved WOA algorithm is used to search for the optimal hyperparameters of each Transformer network.

[0078] The present invention discloses a lumped Transformer network formed by a parallel collection of multiple time series. Multimodal time series have a variety of data structures and expressions, and a single Transformer network cannot fully extract its features. To this end, this section proposes an innovative lumped Transformer network, which assigns a dedicated Transformer network to data of different modes in a multimodal time series, forming a Transformer network formed by a parallel collection of multiple Transformer networks. Figure 2 As shown in Figure 1, the overall architecture of the lumped Transformer network is shown. Figure 2In the example, the entire network includes multiple time series Transformer networks, Add&Norm modules, and Classifiers. Figure 3 The structure of a single time series Transformer network is presented in detail. Its design fully considers the characteristics of time series data and can effectively extract useful information from complex data.

[0079] The intelligent detection method for asynchronous induction motor fault based on multimodal time series of the present invention comprises the following steps:

[0080] Step 1: Multimodal data acquisition module steps. Integrate multiple sensors and data acquisition technologies to synchronously collect different types of data, including sound, vibration, current and other different modal data.

[0081] Specifically, motor fault data of 5 different modes are collected (motor fault data of 5 different modes: environmental sound, equipment sound, front-end vibration, rear-end vibration, and current), as shown in Table 1; then, the length of data of different dimensions in each mode is set to 1024; then, the data set is randomly scattered, 60% of the data is used as a training set, 20% of the data is used as a validation set, and 20% of the data is used as a test set;

[0082] Table 1 Multimodal signals of motor faults

[0083]

[0084] Step 2: Merge the data into a lumped Transformer network consisting of multiple parallel sets of time series. The multidimensional time series of the training set, validation set, and test set are divided into small blocks of fixed size, then linearly embedded, position embedding is added, and the resulting vector sequence is fed to the standard Transformer encoder. For classification, the standard method is used here, that is, adding an additional learnable object "classification token" to the sequence. Its detailed structure is as follows:

[0085] Step 21: Token embedding. The standard Transformer network receives a one-dimensional sequence as a token embedding. However, the time series of different modes in a multimodal time series may be multidimensional, that is, the time series Where M and N represent the length and dimension of the time series, C represents the number of channels, and R represents the set of real numbers; then, it is flattened into That is, x is the flattened data, and it is cut into a block sequence [x 1 ,x 2 ,…,x M ]and Since, in all layers of the Transformer, a constant latent vector of size D is used. Therefore, the block sequence is unfolded and mapped to D dimensions through a trainable linear projection, and the output of the projection is called token embedding;

[0086] Step 22: Class token. Similar to BERT's [class] token, a learnable embedding Add to block sequence [x 1 ,x 2 ,…,x M ], (x class : The classification token representing the block sequence is taken according to the number of classifications), and a new embedded block sequence z0 = [x class ;x 1 E;x 2 E;…;x N E], where

[0087] Step 23: Position embeddings. In order to preserve the position information, position embeddings are added to the block sequence. Here, the standard learnable 1-D position embedding method is used, and z0 is expanded to obtain the embedding vector sequence z0 = [x class ;x 1 E;x 2 E;…;x N E]+E positon ,in, The resulting embedding vector sequence is then used as input to the encoder.

[0088] Step 24: Encoder. A Transformer encoder is used, such as Figure 4 As shown in Figure 1. It consists of N Transformer blocks and takes the embedding vector sequence z0 as input. In addition, each Transformer block contains a multi-headed self-attention layer, an addition & normalization layer, and a feed-forward neural network layer.

[0089] The feedforward neural network layer can be expressed by the following formula:

[0090] z l ′=LayerNorm(MSA(z l-1 )+z l-1 ) l=1,2,…,N (1)

[0091] zl =LayerNorm(FFN(z l ' -1 )+z l ' -1 ) l=1,2,…,N (2)

[0092] In the formula, MSA(·) is a multi-headed self-attention layer; FFN(·) is a feed-forward neural network layer. In addition, since each time series data of a multimodal time series has its own feature range. The hidden state of each encoder output is With different ranges of values, this diversity makes the subsequent modules unstable. Therefore, the context normalization model is introduced, which is defined as:

[0093]

[0094] In the formula, β n＝1,2,,N is a hyperparameter that determines The weight of .

[0095] Self-Attention Layer. It can dynamically focus on specific modalities or time points according to the importance of each modality at different time points, and can more effectively capture relevant information across modalities. However, self-attention has a quadratic time and space complexity to the sequence length and cannot be effectively extended to long sequences. In order to improve the efficiency of the lumped Transformer network, the learning-to-hash attention mechanism (LHA) is used here to replace self-attention. The core idea is to learn separate parameterized hash functions for queries and keys respectively, so that the sparse pattern in LHA is no longer limited to distance-based hash functions such as (Locality Sensitive Hashing) LSH or online k-means, and can adapt to the mismatch between query and key distributions. The definition of sparse attention is as follows:

[0096]

[0097] In the formula, is the hash function of the key and query, B is the hash bucket, (d h 、h K ,h Q , Q i ,K,V, d h Represents dimensions; query key (Q), key (K), and value (V); V represents the attention weight between the i-th query and the j-th key in the sparse attention mechanism; j Denotes the value of the j-th key vector K j The associated value vector. is satisfied The attention weights are defined by parameterizing the function LHA implements a learnable hash function The specific process is as follows:

[0098]

[0099] In the formula, H Q and H K is an arbitrary parameterized function. Finally, a single linear layer is applied to obtain the hidden state

[0100] Add&Norm layer. It adds the hidden state output from the attention layer Addition is performed to keep the original input information in the network to avoid information loss; the normalization step normalizes the result of the addition operation to help the network converge more stably during training and improve the generalization performance of the model.

[0101] Step 25 Classifier. It is the last layer of the lumped Transformer network. First, the input data is encoded using each time series Transformer to generate feature representations. These feature representations contain the contextual information of the input data; then, the classifier maps the features output by the addition & normalization layer to the probability distribution of the category through the feed-forward multi-layer perceptron (MLP) and the softmax activation function, as shown below:

[0102]

[0103] In the formula, x is the flattened data. Among them, W1 is the weight matrix and b1 is the bias vector. The purpose of the GeLU function is to introduce nonlinearity to help the network learn more complex patterns. After the GeLU activation function, the result is multiplied by another weight matrix W2 and added with the bias phasor b2. It is a Gaussian error linear unit that retains the linear part and the nonlinear part and has the property of continuous differentiability. Compared with ReLU, GeLU has a non-zero gradient in the negative value interval, which helps to better transfer the gradient during training and reduce the "dead neuron" problem; It is the hidden layer feature and serves as the output of the encoder for subsequent classification (the hidden layer is included in the Softmax layer, which is a regular component of the Softmax layer).

[0104] Step 3. Steps of 3D chaotic composite mapping

[0105] Many intelligent optimization algorithms have adopted random number strategies to some extent. It is worth noting that the essence of intelligence lies in randomness, and the degree of randomness determines the level of intelligence. The Lyapunov exponent of the compound chaotic map may decrease or even lose its chaotic characteristics. The mechanism of this step is that when the classification accuracy of the classifier does not reach the preset threshold, the chaotic composite mapping set WOA readjusts the hyperparameters of the Transformer network according to the fitness function. Therefore, in order to obtain a composite mapping with excellent chaotic characteristics, the present invention proposes a modified 3D Logistic-Sine composite chaotic map, whose mathematical expression is as follows:

[0106]

[0107] Where k is the chaos control parameter, k∈(0,4)x n ,y n 、z n ∈[0,1],x n ,y n , Z n They represent the three-dimensional coordinate values ​​of the pigeon's flying position, mod1 represents the remainder of 1, %1, then the next position update of the pigeon is as follows:

[0108]

[0109] Next, the chaotic orbit and Lyapunov exponent of the modified 3D Logistic-Sine composite mapping are plotted to verify the chaotic characteristics of the mapping. Figure 5 The chaotic orbits in the space are evenly spread across the entire space. Figure 6 The Lyapunov exponent in also obtains a large value. Therefore, these two indicators show that the composite chaotic map has excellent chaotic characteristics.

[0110] Step 4. Chaos-Whale Optimization Algorithm

[0111] The chaotic characteristics of chaotic mapping help to prevent WOA from falling into the local optimal solution, thereby improving the efficiency of searching for the global optimal solution and accelerating the convergence speed. The use of 3D Logistic-Sine composite mapping can improve the search efficiency of WOA and accelerate its convergence speed. The specific implementation steps are as follows:

[0112] Step 4.1 Parameter initialization

[0113] (1) Initialization task dimension D; WOA population size N D ; Maximum number of iterations T; Boundary constraints for different tasks B D ∈[B1,B2] D ;

[0114] (2) Initialize the control parameter c of the modified 3D Logistic-Sine composite mapping; variables x1 = rand(·), x2 = rand(·), x3 = rand(·), where rand(·) is a uniformly distributed random function and rand(·)∈(0,1);

[0115] (3) Initialize the initial position of each WOA in different populations As shown below:

[0116]

[0117]

[0118]

[0119] Where CRand is the spatial distance between the trajectory point of the chaotic composite mapping and the origin (0,0,0), and CRand∈[0,1]. Using CRand to initialize the position of WOA can increase the diversity of the population;

[0120] (4) Initialize the random vector r1, r2, p of WOA, and r1, r2, p = CRand;

[0121] Step 4.2 Hunting process

[0122] (1) Surrounding the prey

[0123] In the standard WOA, the parameter vector A determines the search capability of the WOA. Although CRand can enhance the nonlinear complexity of A, the control parameter a in A is linear, which affects the search capability of the WOA to a certain extent. Therefore, this paper adopts the nonlinear parameter a as shown below:

[0124]

[0125] Where t is the current iteration number; Tmax is the maximum iteration number; is a random disturbance, and

[0126] In addition, as the whale surrounds its prey, the coefficient vectors A and C of the position update control the whale's swimming pattern. In order to enhance their randomness, this section adopts the strategy of chaotic composite mapping as follows:

[0127]

[0128] In the formula, (respectively represent the three-dimensional coordinate information of the position of the i-th pigeon in the pigeon flock) are variables of the modified 3DLogistic-Sine composite mapping, and they all belong to [0,1].

[0129] (2) Hunting stage

[0130] At this stage, WOA used a spiral update position mechanism, which achieved good results in the application. However, according to the hunting behavior of whales, extending the spiral update mechanism from 2D to 3D is more in line with the hunting behavior of whales. Therefore, a 3D logarithmic spiral update rule is proposed, and the process is as follows.

[0131]

[0132] Where b is a constant used to define the logarithmic spiral shape; l∈[-1,1] is a random number, and the 3D spiral curve is as follows Figure 7 As shown. Figure 7 In the figure, the spatial point M(0,0,0) spirals up to the spatial point M(X spiral ,Y spiral ,Z spiral ), and the distance between two points in space is:

[0133]

[0134] Then, the spiral update mechanism of the standard WOA is improved using formula (16) as follows:

[0135]

[0136] The parameter D represents the task dimension, and the random numbers l, X(t+1), X * (t) represent the current optimal value and the previous optimal value of the pigeon respectively.

[0137] (3) Searching for prey

[0138] When |A|≥1, a whale randomly selected by the whale population updates its position. This strategy avoids falling into the local optimal solution to a certain extent. However, if the early search deviates from the target value, the later search will inevitably fall into the local optimal solution. In order to further reduce the probability of WOA falling into the local optimal solution, the Cauchy mutation strategy is introduced at this stage, and the mathematical model is shown below.

[0139]

[0140] D Rand The random distance representing the pigeon's position, X CRand( t) The spatial distance between the pigeon’s current position and the origin M(0,0,0), Represents the Hadamard operation.

[0141] Where Cauchy is the Cauchy operator, and the probability density function expression of the one-dimensional standard Cauchy distribution is as follows:

[0142]

[0143] From formula (19), we can see that the Cauchy operator has a long "tail". The distribution with longer ends can make individuals have a higher probability of jumping to a better position, so that individuals can escape from the local optimal solution. At the same time, the smaller peak at the center point indicates that the Cauchy operator spends less time searching the domain space. Therefore, the introduction of the Cauchy operator to mutate the individual position to generate a new solution can increase the probability of escaping the local optimal solution to a certain extent and accelerate the convergence speed of WOA.

[0144] Step 4.3 Fitness function

[0145] When using chaotic WOA to optimize the hyperparameters of the E-Transformer network, it is necessary to construct the corresponding fitness function as shown below:

[0146]

[0147] In the formula, Fitness is the fitness function of the E-Transformer network; loss is the loss function during the training process of the Transformer network; fitness is the fitness function of each independent Transformer network in the E-Transformer network, i=1...5.

[0148] Step 5. Train and test the E-Transformer network based on chaos-whale optimization.

[0149] The specific implementation process of the present invention is:

[0150] (1) Collect motor fault data of 5 different modes. Then, set the length of data of different dimensions in each mode to 1024, for a total of 200 groups. Then, randomly break up the data sets of different modes, with 60% of the data as training set, 20% of the data as validation set, and 20% as test set.

[0151] (2) Initialize the hyperparameters of each independent Transformer network in the E-Transformer network uniformly, that is, the embedding dimension is 8, the number of attention heads is 4, the hidden layer dimension is 32, the number of encoder layers is 2, the learning rate is 0.0001, the Dropout is 0.2, the Batch size is 16, and the execution environment is GPU;

[0152] (3) Initialize the Chaos-Whale Optimization Algorithm, with chaos control parameter c = 3.99, variables x1 = rand(·), x2 = rand(·), x3 = rand(·), and rand(·)∈(0,1); initialize the total number of WOA tasks E = 5, the dimension of a single task D = 7; the population size N D =[40,20,100,20,40,20,60]; Maximum number of iterations T = 200; Boundary constraints B for different tasks D ∈[B1,B2] D , where [B1,B2]1=[2,256], [B1,B2]2=[2,16], [B1,B2]3=[8,1024], [B1,B2]4= [1,20], [B1,B2]5=[1e-5,1e-3], [B1,B2]6=[0.1,0.5], [B1,B2]7=[1,256]; l=rand(.)∈[-1,1], β=[0.1,0.1,0.2,0.2,0.4], b=1;

[0153] (4). According to formula (20), the loss function of each independent Transformer network in the E-Transformer network is used as the corresponding fitness function, and the minimum value of the sum of these fitness functions is used as the fitness function of the E-Transformer network. Then, 120 sets of data in each mode of the 5 different modal data sets are used as training sets, and 40 sets of data are used as validation sets to input into the 5 corresponding networks in the E-Transformer network for training and testing. Finally, the embedding dimension, number of attention heads, hidden layer dimension, number of encoder layers, learning rate, Dropout, and Batch value of the E-Transformer network corresponding to the minimum value are updated accordingly;

[0154] (5) 60 groups from each of the five different modal data sets were input into the updated E-Transformer network as prediction sets for fault prediction and verification, with a diagnostic accuracy of 92.38%.

[0155] In large-scale industrial production environments, due to harsh working conditions and large load changes, asynchronous motors often face the risk of rotor breakage, local short circuits between turns of stator winding coils, and cracks, peeling, pitting and other faults in bearings. These faults not only affect the performance of the motor, but may even cause the entire production line to stagnate in severe cases. In order to diagnose various motor faults more efficiently and accurately, this paper proposes an innovative motor fault diagnosis method. This method combines the chaotic whale optimization algorithm (WOA) with the Transformer network, based on multimodal time series data, to provide more reliable fault warning and diagnosis for industrial production. Through chaotic WOA optimization, the Transformer network can better adapt to the complexity and uncertainty of motor fault data, thereby more accurately identifying and locating faults. The present invention not only improves the accuracy of diagnosis, but also provides strong technical support for motor maintenance in industrial production. The research content of the present invention is summarized as follows.

[0156] (1) In the field of deep learning, Transformer networks have become popular for their outstanding sequence modeling capabilities. Especially when processing time series data, this network structure shows its unique advantages. The time series-based lumped Transformer network proposed in this paper, also known as the parallel Transformer network, is a network architecture designed for efficient processing of time series data. The core idea is to combine multiple Transformer models in parallel to form a more powerful and flexible model. Each Transformer model can independently process the input time series data and capture the long-term dependencies and temporal patterns therein. Through parallel computing, this network structure can significantly improve the speed and efficiency of data processing, enabling the model to learn and infer faster;

[0157] (2) In the field of motor fault diagnosis, data diversity and richness are crucial for building efficient and accurate diagnostic models. To achieve this goal, we use a variety of sensors to fully capture the operating status of the motor. Specifically, we use professional current sensors to collect the three-phase current signals of the motor, which can reflect the electrical characteristics and changes inside the motor. At the same time, we also deploy high-precision vibration sensors to capture the three-dimensional vibration signals generated by the motor during operation. These signals often contain the working status of mechanical components and potential fault information. In addition, sound signals also provide us with valuable diagnostic clues. Therefore, we collect two-dimensional sound signals of the motor, which can reveal the acoustic characteristics of the motor during operation, thereby helping us identify certain specific fault modes. In addition, to further enrich the dataset, we also collect one-dimensional environmental sound signals. Although these signals are less related to the direct operating status of the motor, they may contain important information about external factors that affect the operation of the motor.

[0158] (3) In the field of deep learning and artificial intelligence, the choice of optimization algorithm is crucial to the performance of the model. As an emerging heuristic optimization algorithm, the whale optimization algorithm has shown unique advantages in solving complex optimization problems. However, the global search capability and convergence speed of the whale optimization algorithm still need to be improved. To overcome this challenge, we innovatively introduced a modified 3D Logistic-Sine composite mapping. This mapping provides a more complex and efficient search space for the whale optimization algorithm by combining the nonlinear characteristics of the Logistic mapping and the periodicity of the Sine mapping. It helps the algorithm to achieve a better balance between exploration and development, thereby improving the accuracy and efficiency of global search. At the same time, by adjusting the parameters of the mapping, the convergence speed of the algorithm can also be controlled, so that it can find the optimal solution faster while maintaining accuracy. By utilizing the proposed chaotic WOA optimization algorithm, we are able to automatically search and determine the optimal hyperparameter configuration of the Transformer network, thereby achieving higher classification performance.

Claims

1. An intelligent fault detection method for asynchronous induction motors based on multimodal time series, characterized in that: The following steps are involved: Step 1: The steps of the multimodal data acquisition module, synchronously collect different types of data, Step 2: Merge the data into a lumped Transformer network consisting of multiple parallel sets of time series; the specific steps are as follows: Step 21: Token embedding; time series Where M and N represent the length and dimension of the time series, C represents the number of channels, and R represents the set of real numbers; then, it is flattened into Cut into a block sequence along the time direction [x 1 ,x 2 ,…,x M ]and Since, the block sequence is unfolded and mapped to D dimensions through a trainable linear projection; Step 22: Classify tokens; add a learnable embedding to the chunk sequence [x 1 ,x 2 ,…,x M ], we get a new embedded block sequence z0=[x class ;x 1 E;x 2 E;…;x N E], where x class a classification token representing a sequence of blocks; Step 23: Position embedding: Using a learnable 1-D position embedding method, z0 is expanded to obtain an embedding vector sequence z0 = [x class ;x 1 E;x 2 E;…;x N E]+E positon ,in, The resulting embedding vector sequence is used as the input of the encoder; Step 24 encoder: use Transformer encoder, take the embedding vector sequence z0 as input; each Transformer block contains multiple self-attention layers, addition & normalization layers and feed-forward neural network layers; The feedforward neural network layer is represented by the following formula: z l ′=LayerNorm(MSA(z l-1 )+z l-1 ) l=1,2,…,N (1) z l =LayerNorm(FFN(z l ′ -1 )+z l ′ -1 ) l=1,2,…,N (2) Where MSA(·) is a multi-self-attention layer; FFN(·) is a feed-forward neural network layer, and the hidden state of each encoder output is It is defined as: In the formula, β n＝1,2,,N is a hyperparameter; Self-attention layer; sparse attention is defined as follows: In the formula, is the hash function of the key and query, B is the hash bucket, d h Indicates dimension; Q, K, and V represent query, key, and value respectively; V represents the attention weight between the i-th query and the j-th key in the sparse attention mechanism; j Denotes the value of the j-th key vector K j the associated value vector; is satisfied The attention weights are defined by parameterizing the function Implementing a hash function The specific process is as follows: In the formula, H Q and H K is an arbitrary parameterized function; a single linear layer is applied to obtain the hidden state H; Addition & Normalization Layer: The hidden state output by the self-attention layer Add and keep the original input information in the network; Step 25: Classifier; use each time series Transformer to encode the input data to generate feature representation; the classifier maps the features output by the addition & normalization layer to the probability distribution of the category through the feedforward multilayer perceptron and the softmax activation function, as shown below: In the formula, x is the flattened data; is the hidden layer feature; where W1 is the weight matrix, b1 is the bias vector; weight matrix W2, bias vector b2; Step 3. Steps of 3D chaotic composite mapping; Step 4 Chaos-Whale Optimization Algorithm, the implementation steps are as follows: Step 4.1 Parameter initialization (1) Initialization task dimension D; WOA population size N D ; Maximum number of iterations T; Boundary constraints for different tasks B D ∈[B1,B2] D ; (2) Initialize the control parameter c of the modified 3D Logistic-Sine composite mapping; variables x1 = rand(·), x2 = rand(·), x3 = rand(·), where rand(·) is a uniformly distributed random function and rand(·)∈(0,1); (3) Initialize the initial position of each WOA in different populations As shown below: Where CRand is the spatial distance between the trajectory point of the chaotic composite mapping and the origin (0,0,0), and CRand∈[0,1]; (4) Initialize the random vector r1, r2, p of WOA, and r1, r2, p = CRand; Step 4.2 Hunting process (1) Surrounding the prey The nonlinear parameter a is used as follows: Where, t is the current iteration number; Tmax is the maximum iteration number; is a random disturbance, and The strategy of chaotic composite mapping is as follows: In the formula, is the variable of the modified 3D Logistic-Sine composite mapping; (2) Hunting stage The logarithmic spiral update rule for 3D is as follows: Where b is a constant that defines the logarithmic spiral shape; l∈[-1,1] is a random number, and the spatial point M(0,0,0) spirals up to the spatial point M(X spiral ,Y spiral ,Z spiral ), and the distance between two points in space is: The spiral update mechanism of the standard WOA is improved using formula (16) as follows: The parameter D represents the task dimension, and the random numbers l, X(t+1), X * (t), respectively represent the current optimal value and the previous optimal value of the pigeon; (3) Searching for prey Using the Cauchy mutation strategy, the mathematical model is as follows: D Rand The random distance representing the pigeon's position, X CRand (t) The spatial distance between the pigeon’s current position and the origin M(0,0,0), represents the Hadamard operation; Cauchy is the Cauchy operator; Step 4.3 Fitness function Construct the corresponding fitness function as follows: Where, Fitness is the fitness function of the E-Transformer network; loss is the loss function during the Transformer network training process; fitness is the fitness function of each independent Transformer network in the E-Transformer network, i=1...5; Step 5. Train and test the E-Transformer network based on chaos-whale optimization.

2. The method for intelligent fault detection of asynchronous induction motor based on multimodal time series according to claim 1 is characterized in that: The mathematical expression of 3D chaotic composite mapping in step 3 is as follows: Where k is the chaos control parameter, k∈(0,4)x n ,y n 、z n ∈[0,1],x n ,y n , Z n They represent the three-dimensional coordinate values ​​of the pigeon's flying position, mod1 represents the remainder of 1, %1, then the next position update of the pigeon is as follows:

3. The method for intelligent fault detection of asynchronous induction motor based on multimodal time series according to claim 1 or 2, characterized in that: The data collected in step 1 include environmental sound, equipment sound, front-end vibration, rear-end vibration or data of different modes of current.

4. The method for intelligent fault detection of asynchronous induction motor based on multimodal time series according to claim 1 or 2, characterized in that: In step 1, the data set is randomly broken up, with 60% of the data used as a training set, 20% of the data as a validation set, and 20% as a test set.

5. The method for intelligent fault detection of asynchronous induction motor based on multimodal time series according to claim 1 or 2, characterized in that: The probability density function expression of the one-dimensional standard Cauchy distribution of the Cauchy operator is as follows:

Images