A motor fault diagnosis method based on binary pattern and PBA-ANN

Abstract

The application discloses a motor fault diagnosis method based on binary pattern and PBA-ANN, comprising: collecting motor fault data, establishing two kinds of motor fault sound data sets, i.e. a near-fault sound signal data set and a far-fault sound signal data set; using discrete wavelet transform (DWT) and one-dimensional binary pattern (1D-BP) to complete feature extraction of the motor fault sound signal, using a nearest component analysis (NCA) method to select the most representative motor fault features, so as to obtain typical features capable of effectively representing motor faults; using a PBA algorithm to optimize the weight of an ANN, establishing a PBA-ANN classification model, and applying the model to motor fault diagnosis and prediction. Compared with the prior art, the application completes diagnosis of the working state and faults of five induction motors, including health, one rotor bar fracture fault, two rotor bar fracture faults, three rotor bar fracture faults and bearing faults. Results show that the application has good diagnosis performance and can significantly improve the prediction accuracy of motor faults.

Description

Technical Field

[0001] This invention relates to the field of motor fault diagnosis technology, and specifically to a motor fault diagnosis method based on binary mode and PBA-ANN. Background Technology

[0002] Induction motors are frequently used in industrial applications. However, improper operation or prolonged running can lead to malfunctions. Common motor failures generally fall into two categories: mechanical and electrical. Mechanical failures are typically categorized as coil displacement, bearing failure, and dynamic / static eccentricity. Electrical failures include insulation failures, stator failures, rotor bar breakage, and magnetic circuit failures. These failures can directly cause industrial production to stall; therefore, early motor fault detection plays a crucial role in promoting and supporting industrial production.

[0003] Among the aforementioned faults, stator faults, rotor faults, and bearing faults are the most common. For the detection of these fault types, vibration, current, voltage, sound, and thermal imaging methods are generally used. In addition, machine learning-based methods have also been used to determine the types of various faults. For example, some researchers have proposed methods for detecting rotor and stator faults in motors, using current and vibration sensors to detect bearing and eccentricity faults. Other researchers have proposed stator and rotor fault detection methods based on current and voltage, using current signals to detect bearing faults in induction motors. Furthermore, methods using acoustic and vibration signals to detect motor faults have also emerged. For example, by using a 0.8kW induction motor to obtain speed, acoustic, current, voltage, and vibration signals, rotor and bearing faults can be identified. Regarding fault classification algorithms, neural networks are currently used in fault diagnosis. For example, some researchers have proposed a method for detecting rotor and bearing faults in three-phase induction motors based on LSTM neural networks. By collecting motor state sample data from six different locations, vibration sensors are used to detect the motor's state, and a Long Short-Term Memory (LSTM) deep classification method is used to complete the classification.

[0004] While there are many methods for diagnosing motor faults, overall, these methods have limited accuracy and are complex to operate. On the other hand, deep learning algorithms, an emerging branch of artificial intelligence, play a crucial role in fault diagnosis. However, inappropriate deep learning algorithms only increase diagnostic complexity and limit accuracy. Therefore, it is always necessary to select more suitable deep learning algorithms to effectively classify different types of motor faults. Furthermore, the selection of feature parameters is one of the main concerns in deep learning applications. Ineffective selection of feature parameters can significantly impact diagnostic results and increase the complexity of training the deep learning algorithm. Summary of the Invention

[0005] Purpose of the invention: To address the limitations of traditional motor fault diagnosis methods in terms of limited accuracy and complex operation, this invention proposes a motor fault diagnosis method based on binary mode and PBA-ANN. This method can accurately diagnose the status of five induction motors, including one healthy motor, one rotor bar fracture fault, two rotor bar fracture faults, three rotor bar fracture faults, and one bearing fault, thus enabling effective identification of motor faults.

[0006] Technical solution: This invention discloses a motor fault diagnosis method based on binary mode and PBA-ANN, including the following steps:

[0007] Step 1: Collect motor fault data and establish two types of motor fault sound datasets: near fault sound signal dataset and far fault sound signal dataset.

[0008] Step 2: Based on DWT and 1D-BP technology, multi-level one-dimensional discrete wavelet transform (1D-DWT) technology is used to decompose and denoise the motor fault sound signal to obtain the original feature signal of the motor fault; the 1D-BP method is used to extract the features of the motor fault signal, and nearest neighbor component analysis (NCA) is used to effectively select the motor fault features.

[0009] Step 3: Perform motor fault diagnosis and classification based on PBA-ANN deep learning. Optimize the weights of the ANN using the PBA algorithm, establish a PBA-ANN classification model, and apply it to the diagnosis of motor faults.

[0010] Furthermore, in step 1, motor fault data is collected, and the specific steps are as follows:

[0011] Step 1a: Identify and generate the fault types, including one healthy motor and four faulty motors, for a total of four fault types: one broken rotor bar fault, two broken rotor bar faults, three broken rotor bar faults, and bearing faults.

[0012] Step 1b: Collect motor fault data, establish and store fault sound dataset.

[0013] Record motor fault sound data. Record the sound signals of the motor under five different fault conditions to complete the acquisition of experimental data in two cases: one is the sound acquisition of near faults and the other is the sound acquisition of far faults. Finally, sample 44,000 experimental samples per second to establish two motor fault sound datasets, namely the near fault sound signal dataset and the far fault sound signal dataset. Divide the sound data into 1-second samples as sound samples, and store and process the sound data.

[0014] Furthermore, in step 2, the process of using multi-level one-dimensional discrete wavelet transform (1D-DWT) technology to decompose and denoise the motor fault sound signal to obtain the original characteristic signal of the motor fault is as follows:

[0015] The frequency coefficients of motor fault sound data are obtained using multi-level one-dimensional discrete wavelet transform (1D-DWT) technology, followed by signal decomposition. Five levels of 1D-DWT transformation are performed using Symlets 6 filters in wavelet transform, resulting in five low-pass filters at each level. The mathematical definition of multi-level discrete wavelet transform is as follows:

[0016] [Low1,High1]=dwt(Vol,sym6)

[0017] Low i High i ] = dwt(Lo i-1 ,sym6),i=2,3,4

[0018] dwt(·) is a 1D-DWT function, and Symlets 6 is the filter used; thus, the original characteristic signal of the motor fault is obtained through the above low-pass filter subband and fault sound signal.

[0019] Furthermore, in step 2, the 1D-BP method is used to extract features from the motor fault signal. In 1D-BP, since the signal is one-dimensional, its neighborhood relationship is extracted by using overlapping blocks of length 9. 1D-BP consists of several steps, including block processing, generating binary features, and constructing the BP signal. The specific steps are as follows:

[0020] (1) Collect and load the fault sound signal of the three-phase induction motor;

[0021] (2) Perform block processing, dividing the fault sound signal into blocks of length 9. The specific calculation formula is as follows:

[0022] block=Vol(i:i+8);i=2,…,length-8

[0023] Where length is the sampling length of the signal.

[0024] (3) Calculate the BP bit. Use the signum function to calculate the binary BP bit, as follows:

[0025] bits k =signum(block) i ,block5);i=1,2,…,9;i≠5;k=1,2,…,8

[0026]

[0027] Among them, bits k It is the extracted k-th binary value, and x and y are the input signals of the signum function. It can be seen that each value of the block is compared with the 5th value (center) of the block to extract the relationship between blocks.

[0028] The low-pass filter coefficients generated by the BP(·) function above are used to extract feature signals from the original signal, as follows:

[0029] f1 = BP(Vol)

[0030] f2 = BP(Low1)

[0031] f3 = BP(Low2)

[0032] f4 = BP(Low3)

[0033] f5 = BP(Low4)

[0034] f6 = BP(Low5)

[0035] Where f1, f2, f3, f4, f5, and f6 represent features of size 256 at each level;

[0036] (4) Calculate the decimal feature value. Calculate the decimal value corresponding to the extracted binary feature, as follows:

[0037]

[0038] Where decval is the corresponding decimal eigenvalue;

[0039] (5) Feature cascade integration: The features extracted above are cascaded and integrated. The cascade process is completed according to the following formula.

[0040] ff = conc(f1,f2,f3,f4,f5,f6)

[0041] Where ff is the final feature vector with a size of 1536, and conc(·) is the cascade function;

[0042] (6) Normalization processing: Normalize the data in each column of ff as follows:

[0043]

[0044] Among them, f n These are the normalized features, ff min and f max These are the minimum and maximum values ​​of ff.

[0045] Furthermore, in step 2, nearest neighbor component analysis (NCA) is used to select motor fault features. Nearest neighbors are randomly selected, and the transformation matrix in the Mahalanobis distance is obtained by optimizing the leave-one-out cross-validation results, as detailed below:

[0046] (1) Calculate the Mahalanobis distance between two samples.

[0047] For the n input feature samples obtained above, x1, x2, ..., x n ,x i ∈R D i = 1, 2, ..., n, each with class labels c1, c2, ..., c n To find a distance that optimizes nearest neighbor classification, the Mahalanobis distance transformation matrix Q is set to be a symmetric positive semi-definite matrix, i.e., Q = A. T A, then, the Mahalanobis distance between two samples is defined as:

[0048]

[0049] Where i = 1, 2, ..., n;

[0050] (2) The classification error is calculated using the leave-one-out method. A differentiable softmax function is introduced, as shown in the following equation:

[0051]

[0052] p ij The definition means that the sample point x i Randomly select a nearest neighbor, and finally select another sample point x. j As its nearest neighbor, and inheriting its class tag c j The probability of sample x, therefore, i The probability of correct classification is:

[0053]

[0054] Among them, c i ={j|c i =c j};

[0055] (3) Maximize the objective function to complete the dimensionality reduction and selection of feature data. The objective function is to maximize the number of correctly classified points, and is defined as follows:

[0056]

[0057] This definition is a continuously differentiable matrix function. The optimization task of the algorithm is accomplished by maximizing this objective function. A is obtained using the conjugate gradient method, and the conjugate gradient is calculated as follows:

[0058]

[0059] Where, x ij =x i -x j ;

[0060] When A is a non-square matrix of d×D, reduce the sample size to R. d This allows for dimensionality reduction and sorting of the feature data.

[0061] (4) Select 128 of the most typical features. Using the NCA method, select 128 of the most typical features as follows:

[0062] f s (i)=f n (indice(i)), i = 1, 2, ..., 128

[0063] Among them, f s The selected feature is called `indice`, which is the sorted feature obtained using the NCA method.

[0064] Furthermore, in step 3, the weights of the ANN are optimized using the PBA algorithm to establish a PBA-ANN classification model. The specific steps are as follows:

[0065] Step 3a: Update the weights and biases of each layer of the model using the following equations:

[0066]

[0067]

[0068] Where W, B, l, λ, x, n, m, t, and O represent weights, biases, number of layers, regularization parameter, learning rate, total number of training samples, momentum, update step, and cost function, respectively; the output unit G and hidden unit H are mathematically represented by the following equations:

[0069] G = ΔB l +∑HΔW l

[0070] H = ΔB l +∑(D i W ij

[0071] Among them, W ij This represents the weights between the input and the hidden layer;

[0072] Step 3b: Optimize the weights of the neural network ANN using the beetle algorithm PBA.

[0073] Furthermore, step 3b, which involves optimizing the weights of the neural network ANN using the beetle algorithm (PBA), specifically includes the following steps:

[0074] (1) Complete random sampling of the sample data, and use random sampling technique (RST) to ensure that all search spaces can be correctly represented by N. pop One sample;

[0075] (2) Initialization: The entire algorithm is in the 0th generation of optimization. The optimized individuals of PBA are arbitrarily placed within the hypervolume equivalent related to the global search space to search for the optimal position vector. These individuals are pioneer individuals; each pioneer individual in the initial population. It can be arbitrarily placed into the global search space, and its specific definition is as follows:

[0076]

[0077] L = [L1, L2, ..., L D ],U=[U1,U2,…,U D ]

[0078] Where U and L represent the upper and lower bound variables of the optimization variables, and D is the data dimension;

[0079] When all individuals identify the hypervolume location (solution vector), N pop Each solution will be compared one by one, and the best individual will be obtained. The first population will be generated at this hypervolume space location. Each pioneer individual of the initial population is used to represent the weight value between the input layer and the hidden layer. Thus, the weight parameters and cost function are initialized.

[0080] (3) New Hyperspace Selection Mode

[0081] In this stage, each newly emerging individual will search for new solutions within the search space to create a total population belonging to the individual itself; among them, there are five hypervolume selection modes: neighbor search, global search, medium-sized search, large-scale search, and memory-considered search;

[0082] Using the above search method, the optimal individual position is established, and the new precursor particle is arbitrarily placed in the search space obtained through random sampling technique (RST), as follows:

[0083]

[0084]

[0085] Where i represents the j-th th Individual solution vector l j The elements of , where e represents the generating (tracing) algebra, It is the i-th th An element of a birth position or solution vector. and It is the eth th The i-th search th The upper and lower limits of the dimension, f pat The model parameters are represented; the above equation is iterated, and during each iteration, the weight and cost functions are analyzed by evaluating the fitness function and compared with the fitness function of the previous iteration to detect the optimal value, thereby re-initializing the weight and cost parameters;

[0086] (4) Update population location

[0087] The population's position is updated by updating the positions of all individuals in the population. Except for the information stored in memory, all other information is deleted. Specifically, all previous populations disappear, and the newly discovered positions become the birthplace of the new generation of the total population. Based on the previous rules, new populations are generated by this hypervolume pattern, and this process is repeated until the PBA reaches the stopping condition.

[0088] Furthermore, using the random sampling technique RST, the uniform distribution function of each variable is divided into multiple segments with equal boundary probabilities. These segments are randomly selected for each variable, and the sample is defined by randomly mixing different variables, i.e., by dividing the range of each D variables into N segments. pop The sample is constructed by dividing the space into non-overlapping equal segments, and then samples of dimension D are created by randomly assigning values ​​to all parameters; therefore, the D-dimensional space is divided into The data is divided into segments, and a single value is randomly selected from each segment to generate a sample of size D;

[0089] For D independent uniformly distributed variables (x)i,j ∈[l i ,u i ],i=1,2,…,D,j==1,2,…,N pop Each of the D variables is divided into N... pop There are equal intervals, thus generating There are units, where each row and each column contains only one sample, and satisfies the following expression:

[0090]

[0091] Among them, rx i ∈[l i ,u i ] is a random number, F x It is x i,j The cumulative distribution function. In N pop Given the number of segments and D variables, the maximum number of combinations of the sampled data is:

[0092]

[0093] Typically, variable x i Defined within a given search space, defined by global boundaries L and U, specifically, and It is the low and high boundary of the ith dimension in the gth step, and the maximum distance of the samples. (Also known as tracking distance) is calculated by the following expression:

[0094]

[0095] Therefore, the sampling technique RST for the gth tracking step can be described as follows:

[0096] x j =RST(l (g) ,u (g) ,D,N pop )

[0097] During the sampling process, it is necessary to [l( g) ,u (g) Sampling is performed within the range of x, and within the global search space. i,j ∈[L i U i ], and any point outside this range will be corrected.

[0098] Furthermore, the population position is updated when optimizing the weight values ​​of the neural network ANN using the beetle algorithm PBA, and the selection of the hypervolume mode is achieved according to the following formula:

[0099]

[0100] Among them, f ls f represents the large-scale factor. ms f represents the medium size factor. nb Indicates adjacent factors.

[0101] Beneficial effects:

[0102] This invention proposes a motor fault diagnosis method based on binary pattern recognition and PBA-ANN. It collects two types of fault sound datasets and extracts features from these sound signals. First, a multi-level feature generation method is proposed, combining Discrete Wavelet Transform (DWT) and one-dimensional binary pattern recognition (1D-BP) to extract features from the motor fault sound signals. Then, the Non-Comparative Analysis (NCA) method is used to select the most representative motor fault features. Second, a motor fault diagnosis method based on PBA-ANN deep learning is proposed. The PBA algorithm is used to optimize the weights of the ANN, thereby establishing a PBA-ANN model, which is then applied to a motor fault diagnosis and prediction system, achieving good motor fault diagnosis results. This invention addresses four types of motor faults: one broken rotor bar, two broken rotor bars, three broken rotor bars, and bearing faults. In motor fault classification using short-range fault sound detection, the proposed method achieves an average classification accuracy (UAP) of 97.5. In motor fault classification using long-range fault sound detection, the proposed method achieves an average classification accuracy (UAP) of 94.5. This demonstrates that the proposed method achieves a motor fault diagnosis accuracy exceeding 94%, exhibiting excellent fault diagnosis performance. Furthermore, this sound-based fault diagnosis method is highly applicable, easy to implement, and has strong potential for widespread adoption. Attached Figure Description

[0103] Figure 1 This describes the implementation process of a motor fault diagnosis method based on binary mode and PBA-ANN.

[0104] Figure 2 This is the block diagram of the multi-level 1D-DWT principle used;

[0105] Figure 3 This is an example of 1D-BP;

[0106] Figure 4 It is a complete motor fault feature extraction process based on binary mode and PBA-ANN;

[0107] Figure 5 It is a PBA-ANN classifier network structure;

[0108] Figure 6 It is a complete motor fault diagnosis process based on binary mode and PBA-ANN. Detailed Implementation

[0109] To better explain and facilitate understanding of the present invention, the technical solution of the present invention is described in detail below. The following embodiments are illustrative of the present invention, but the present invention is not limited to the following embodiments.

[0110] To address the limitations of conventional motor fault diagnosis methods in terms of limited accuracy and operational complexity, this invention proposes a motor fault diagnosis method based on binary mode and PBA-ANN. This method is highly applicable and easy to implement. First, acoustic processing is used to perform multi-level feature extraction based on DWT and 1D-BP. Second, a classification method based on PBA-ANN is proposed, utilizing the PBA algorithm to optimize the weights of the ANN, thereby establishing a PBA-ANN model, which is then applied to a motor fault diagnosis and prediction system.

[0111] Figure 1 The diagram illustrates the implementation process of the motor fault diagnosis method based on binary mode and PBA-ANN in an embodiment of the present invention.

[0112] First, a multi-level feature extraction method is proposed, which uses Discrete Wavelet Transform (DWT) and multi-level one-dimensional discrete wavelet transform (1D-DWT) to extract features from motor fault sound signals. Then, the Neighborhood Component Analysis (NCA) method is used to select meaningful features, a total of 128 features are selected, and finally these features are used as input to the classifier.

[0113] Combination Figure 1 A motor fault diagnosis method based on binary mode and PBA-ANN includes the following steps:

[0114] Step 1: Collect motor fault data and establish two types of motor fault sound datasets: a near-fault sound signal dataset and a far-fault sound signal dataset. The specific processing steps are as follows:

[0115] Step (1)a: Select the motor test object, determine and generate the fault type.

[0116] A squirrel-cage three-phase induction motor was used as the experimental object, with an active power of 0.75 kW and a nominal voltage of 380 V. The characteristics of the induction motor used in the experiment are shown in Table 1.

[0117] Table 1 Characteristics of Induction Motors

[0118] parameter Value Active power 0.75 / kW nominal voltage 380 / V rotational speed 1380 / rpm nominal current 2.2 / A Extreme number 4 Number of rotor bars 22 Nominal frequency 50 / HZ

[0119] The experiment included one healthy motor and four faulty motors, with four types of faults: one broken rotor bar, two broken rotor bars, three broken rotor bars, and bearing failure. To create the faults, holes were drilled in the motors to induce one, two, or three broken rotor bars. Additionally, holes were drilled in the bearings to induce bearing failure.

[0120] Step (1)b: Collect motor fault data, establish and store fault sound dataset.

[0121] During implementation, fault detection was performed on five induction motors, obtaining sound data of the motors in different operating states, including one healthy motor, one rotor bar fracture fault, two rotor bar fracture faults, three rotor bar fracture faults, and one bearing fault. During recording, short-range and long-range sound datasets of the motors were created, and the sound data was divided into 1-second samples (44,000). The sound data was then transmitted to a MATLAB program. In the MATLAB environment, feature extraction was performed using DWT and 1D-BP methods, obtaining 1536 features. Next, the NCA method was used for feature selection, obtaining the 128 most meaningful optimal features. Finally, a classification algorithm was used to identify the different operating states of the motors, thus validating the proposed motor fault diagnosis method.

[0122] To collect sound data for different fault types, fault sounds from the motor were recorded using a mobile phone's recording app. The recording format was m4a, the encoder bit rate was 320kbps, and the sampling rate was 44kHz. Motor sound signals under five different conditions were recorded, completing two types of experimental data acquisition. The first was short-range fault sound data acquisition, where the mobile phone was fixed 10cm away from the experimental motor, recording for 330 seconds. The second was long-range fault sound data acquisition, where the microphone was placed more than 100cm away from the experimental motor, recording for 270 seconds. Finally, according to the settings in the recording app, 44,000 experimental samples were sampled per second.

[0123] Step 2: Extract multi-stage motor fault features based on DWT and 1D-BP, and use nearest neighbor component analysis (NCA) to effectively select motor fault features. The specific steps are as follows:

[0124] Step (2)a: Decompose and reduce noise of motor fault sound signal using multi-level one-dimensional discrete wavelet transform (1D-DWT) technology.

[0125] First, the fault sound signal of a three-phase induction motor is acquired and used as the object of feature extraction. Then, it is preprocessed, primarily to obtain the frequency coefficient data of the fault sound signal. At this stage, a multi-level one-dimensional discrete wavelet transform (1D-DWT) technique is used for the decomposition and noise reduction of the fault sound signal. The frequency coefficients of the fault sound data are obtained using 1D-DWT, followed by signal decomposition. This patent uses a total of five levels of 1D-DWT and calculates five low-pass filter sub-bands.

[0126] The block diagram of the multi-level 1D-DWT principle used is as follows: Figure 2 As shown.

[0127] In the implementation, five low-pass filters were obtained by applying 1D-DWT at five levels, among which the Symlets 6 filter from wavelet transform was used. The mathematical definition of multi-level discrete wavelet transform is as follows:

[0128] [Lo1,Hi1]=dwt(Vol, sym6) (1)

[0129] [Lo i Hi i ] = dwt(Lo i-1 , sym6), i=2,3,4 (2)

[0130] Where dwt(·) is the 1D-DWT function, and Symlets 6 is the filter used. The original characteristic signal of the motor fault is obtained through the above low-pass filter subband and the audio signal.

[0131] Step (2)b: Use the 1D-BP method to extract features from the motor fault signal.

[0132] In the second stage of feature extraction, the 1D-BP method is used. 1D-BP is a one-dimensional local binary pattern that mainly uses windowing to extract feature signals of different magnitudes from different signals. Its main advantages are its simple mathematical structure, ease of encoding, fast processing speed, and ability to extract features independent of signal magnitude.

[0133] In 1D-BP, since the signal is one-dimensional, its neighborhood relationships are extracted using overlapping blocks of length 9. 1D-BP consists of several steps, including block processing, generating binary features, and constructing the BP signal, as detailed below:

[0134] (1) Collect and load the fault sound signal of the three-phase induction motor.

[0135] (2) Perform block processing

[0136] The fault sound signal is divided into blocks of length 9, and the specific calculation formula is as follows:

[0137] block=Vol(i:i+8); i=2,…,length-8 (3)

[0138] Where length is the sampling length of the signal.

[0139] (3) Calculate the BP bit

[0140] The signum function is used to calculate the binary BP bits, as follows:

[0141] bits k =signum(block) i ,block5); i=1,2,…,9; i≠5;k=1,2,…,8 (4)

[0142]

[0143] Among them, bits k It is the extracted k-th binary value, and x and y are the input signals of the signum function.

[0144] Therefore, each value in a block is compared with the fifth value (center) of the block to extract the relationship between blocks. The low-pass filter coefficients are generated using the BP(·) function above, and feature signals are extracted from the original signal, as follows:

[0145] f1 = BP(Vol) (6)

[0146] f2 = BP(Low1) (7)

[0147] f3 = BP(Low2) (8)

[0148] f4 = BP(Low3) (9)

[0149] f5 = BP(Low4) (10)

[0150] f6 = BP(Low5) (11)

[0151] Where f1, f2, f3, f4, f5, and f6 represent features of size 256 at each level.

[0152] (4) Calculate the characteristic value of the decimal number.

[0153] The decimal values ​​corresponding to the above binary characteristics are calculated as follows:

[0154]

[0155] Here, decval is the corresponding decimal characteristic value.

[0156] (5) Feature Cascade Integration

[0157] The features extracted in the previous step are concatenated and integrated. The concatenation process is completed according to the following formula.

[0158] ff=conc(f1,f2,f3,f4,f5,f6) (13)

[0159] Here, ff is the final feature vector with a size of 1536, and conc() is the cascade function.

[0160] (6) Normalization

[0161] Normalize each column of data in ff as follows:

[0162]

[0163] Among them, f n These are the normalized features, ff min and f max These are the minimum and maximum values ​​of ff.

[0164] The decimal feature code calculated by the above formula is 8 bits. Therefore, the 1D-BP of each parameter has 256 features.

[0165] Examples of 1D-BP are as follows Figure 3 As shown.

[0166] In this study, a five-layer wavelet transform was used for feature extraction, resulting in six corresponding layer parameters. In each layer, 256 motor fault features were extracted by using a low-pass filter sub-band. These features were then fused to obtain 256*6=1536 feature sequences.

[0167] Step (2)c: Use the NCA method to select motor fault characteristics.

[0168] In the feature selection phase, the NCA method was used to select 128 of the most salient features from 1536 features. NCA stands for Neighborhood Component Analysis, a distance-based feature weighting method. This method selects features based on a fitness function using Manhattan distance, calculating the correlation between features and the target. During selection, initial weights are first randomly assigned. Then, based on the Manhattan distance function, gradient descent is used to update the weights, thereby generating positive weights for each feature.

[0169] NCA is a simple and effective distance metric learning algorithm. This algorithm randomly selects nearest neighbors and obtains the transformation matrix in Mahalanobis distance by optimizing the leave-one-out cross-validation results. The specific steps are as follows:

[0170] (1) Calculate the Mahalanobis distance between two samples.

[0171] For the n input feature samples obtained above, x1, x2, ..., x n ,x i ∈R D The samples, i = 1, 2, ..., n, have class labels c1, c2, ..., cn respectively. n To find a distance that optimizes nearest neighbor classification, the Mahalanobis distance transformation matrix Q is set to be a symmetric positive semi-definite matrix, i.e., Q = A. T A, then, the Mahalanobis distance between two samples is defined as:

[0172]

[0173] Where i = 1, 2, ..., n.

[0174] (2) Use the leave-one-out method to calculate the classification error.

[0175] Next, we use the leave-one-out method to calculate the classification error, introducing a differentiable softmax function, as shown in the following equation:

[0176]

[0177] p ij The meaning of the definition is that the sample point x i Randomly selecting a nearest neighbor will eventually lead to the selection of another sample point x. j As its nearest neighbor and inheriting its class tag c j The probability of sample x, therefore, i The probability of correct classification is:

[0178]

[0179] Among them, c i ={j|c i =c j}

[0180] (3) Maximize the objective function to complete the dimensionality reduction and selection of feature data.

[0181] The objective function is to maximize the number of correctly classified points, and is defined as follows:

[0182]

[0183] This definition is a continuously differentiable matrix function. The optimization task of the algorithm is accomplished by maximizing this objective function. This is an unconstrained optimization problem, and A is obtained using the conjugate gradient method. The conjugate gradient is calculated as follows:

[0184]

[0185] Where, x ij =x i -x j .

[0186] When A is a non-square matrix of d×D, the sample size can be reduced to R. d This allows for dimensionality reduction and sorting of the feature data.

[0187] (4) Select 128 of the most typical features

[0188] Finally, the NCA method was used to select 128 of the most typical features, as follows:

[0189] f s (i)=f n (indice(i)),i=1,2,…,128 (20)

[0190] Among them, f s The selected feature is index, which is the ranked feature implemented using the NCA method.

[0191] The complete process for extracting motor fault features based on binary mode and PBA-ANN is as follows: Figure 4 .

[0192] Step 3: A motor fault diagnosis method based on PBA-ANN deep learning was implemented. The weights of the ANN were optimized through the PBA algorithm, a PBA-ANN classification model was established, and it was applied to motor fault diagnosis.

[0193] Artificial neural networks (ANNs) typically contain only one hidden layer (HL), which consumes a lot of training time, increases generalization error, results in low accuracy, and leads to overfitting. Furthermore, the weights in traditional neural networks are randomly initialized, which can also cause the classifier to produce classification results with a high error rate. Therefore, this patent proposes using five hidden layers to overcome the training and accuracy problems. Moreover, the pity beetle algorithm (PBA) is used to optimize the network weight values ​​between the input and hidden layers, and between the hidden and output layers, thereby initializing the weight values ​​between each layer and establishing a PBA-ANN model. The PBA-ANN structure is as follows: Figure 3 As shown, where {S1,S2,…,S} n The data is the sensed data, which serves as the input to the classifier.

[0194] The PBA-ANN classifier network structure is as follows: Figure 5 .

[0195] The proposed classifier consists of three layers: an input layer, hidden layers, and an output layer. The number of hidden layers is determined based on classification accuracy. Since the system uses a large amount of motor fault sound data, five hidden layers are used to improve classification performance. The PBA algorithm is used to optimize the initial weights of each layer, then the fitness value for each weight is calculated, and finally, the optimized weight values ​​are obtained based on the optimal fitness value. The final output decision of the classifier is based on the weights and biases of the previous layers.

[0196] The specific steps are as follows:

[0197] Step (3)a: Update the weights and biases of each layer of the model.

[0198] In the proposed classification method, the weights and biases of each layer of the model are updated using equations (21) and (22), respectively:

[0199]

[0200]

[0201] Where W, B, l, λ, x, n, m, t, and O represent weights, biases, number of layers, regularization parameters, learning rate, total number of training samples, momentum, update steps, and cost function, respectively. The output unit G and the hidden unit H are mathematically represented by equations (23) and (24), as shown below:

[0202] G = ΔB l +∑HΔW l (twenty three)

[0203] H = ΔB l +∑(D i W ij (twenty four)

[0204] Among them, W ij This represents the weights between the input and the hidden layer.

[0205] Step (3)b: Optimize the neural network weights using the PBA algorithm.

[0206] Specifically, it includes the following steps:

[0207] (1) Complete random sampling (RST) of the sample data.

[0208] First, a random sampling technique (RST) is proposed to ensure that N can be correctly represented in all search spaces. popThe sample is defined by the RST sampling technique, where the uniform distribution function of each variable is divided into multiple segments with equal boundary probabilities. These segments are randomly selected for each variable, and the sample is defined by randomly mixing different variables, i.e., by dividing the range of each D variables into N segments. pop Samples are constructed by dividing the space into non-overlapping, equal segments, and then samples of dimension D are created by randomly assigning values ​​to all parameters. Therefore, the D-dimensional space is divided into... The data is divided into segments, and a single value is randomly selected from each segment to generate a set of samples of size D.

[0209] For D independent uniformly distributed variables (x) i,j ∈[l i ,u i ],i=1,2,…,D,j==1,2,…,N pop Each of the D variables is divided into N... pop There are equal intervals, thus generating There are units, where each row and each column contains only one sample, and satisfies the following expression:

[0210]

[0211] Among them, rx i ∈[l i ,u i ] is a random number, F x It is x i,j The cumulative distribution function. In N pop Given the number of segments and D variables, the maximum number of combinations of the sampled data is:

[0212]

[0213] Typically, variable x i Defined within a given search space, defined by global boundaries L and U, specifically, and It is the low and high boundary of the ith dimension in the gth step, and the maximum distance of the samples. (Also known as tracking distance) is calculated by the following expression:

[0214]

[0215] Therefore, the sampling technique (RST) for the gth tracking step can be described as follows:

[0216] x j =RST(l (g) ,u (g) ,D,N pop (28)

[0217] During the sampling process, it is necessary to [l] (g) ,u (g) Sampling is performed within the range of x, and within the global search space. i,j ∈[L i U i ], and any point outside this range will be corrected.

[0218] (2) Initialization

[0219] At this point, the entire algorithm is in generation 0 of optimization. The optimized individuals of PBA are arbitrarily placed within the hypervolume equivalent related to the global search space to search for the optimal position vector. These individuals are pioneer individuals. Each pioneer individual in the initial population It can be arbitrarily placed in the global search space, and its specific definition is as follows.

[0220]

[0221] L = [L1, L2, ..., L D ],U=[U1,U2,…,U D (30)

[0222] Where U and L represent the upper and lower bound variables of the optimization variables, and D is the data dimension.

[0223] When all individuals identify the hypervolume location (solution vector), N pop Each solution will be compared one by one, and the best individual will be obtained. The first population will then be generated at this hypervolume location. Each pioneer individual in the initial population is used to represent the weight value between the input layer and the hidden layer, thereby initializing the weight parameters and cost function.

[0224] (3) New Hyperspace Selection Mode

[0225] In this stage, each newly emerging individual will search for new solutions within the search space to create a population belonging to that individual. This involves five hypervolume selection modes: neighbor search, global search, medium-sized search, large-scale search, and memory-considered search.

[0226] Using the above search method, the optimal individual position is established, and the new precursor particle is arbitrarily placed in the search space obtained through random sampling (RST) technique, as follows:

[0227]

[0228]

[0229] Where i represents the j-th thIndividual solution vector l j The elements of , where e represents the generating (tracing) algebra, It is the i-th th An element of a birth position or solution vector. and It is the eth th The i-th search th The upper and lower limits of the dimension, f pat The model parameters are represented by the equations above. The process iterates according to these equations. During each iteration, the weight and cost functions are analyzed by evaluating the fitness function and compared with the fitness function from the previous iteration to detect the optimal value, thereby re-initializing the weight and cost parameters.

[0230] (4) Update population location

[0231] In this step, the population's position is updated by updating the positions of all individuals within the population. All information except for that stored in memory is deleted. Specifically, all previous populations disappear, and the newly discovered locations become the birthplace of the new generation of the population. Based on the previous rules, a new population is generated from this hypervolume pattern. This process is repeated until the PBA reaches the stopping condition. Specifically, the hypervolume pattern is selected according to the following formula:

[0232]

[0233] Among them, f ls f represents the large-scale factor. ms f represents the medium size factor. nb Indicates adjacent factors.

[0234] Using the PBA algorithm described above, the optimal weight values ​​for an neural network (ANN) are obtained. Because the PBA algorithm is more robust and efficient than other algorithms, it is more effective in optimizing ANN network weights. In the specific execution process, initialization is performed first, followed by the execution of a hypervolume space selection mode. During this process, fitness values ​​need to be calculated based on the classifier's accuracy. Finally, the population positions are updated, thereby obtaining the optimal cost and weights. The individuals in this algorithm represent the weight values ​​between the input layer and hidden layers of the neural network.

[0235] A complete motor fault diagnosis process based on binary mode and PBA-ANN is as follows: Figure 5 .

[0236] Step 4: Based on the above technologies, build an experimental platform to implement the classification of motor faults, mainly completing the following test experiments:

[0237] (1) Experimental preparation

[0238] In the implementation process, the experimental subjects used were squirrel-cage three-phase induction motors. Fault detection was performed on five induction motors, and different motor operating sound data were obtained, including healthy motors, motors with one broken rotor bar, motors with two broken rotor bars, motors with three broken rotor bars, and motors with bearing faults. During the recording process, two cases were created: short-range sound data and long-range sound data for the faulty motors. The sound data was divided into 1-second samples (44,000) as sound samples, and the sound data was transmitted to a MATLAB program. In the MATLAB environment, feature extraction was performed using the DWT and 1D-BP methods to obtain 1,536 features. Next, the NCA method was used for feature selection to obtain the 128 most meaningful optimal features. Finally, a classification algorithm was used to classify the motor operating states under different conditions, thus verifying the proposed motor fault detection method.

[0239] During the experiment, the subjects included one healthy motor and four faulty motors, with a total of four fault types: one broken rotor bar fault, two broken rotor bar faults, three broken rotor bar faults, and bearing faults. To create the faults, holes were drilled in the motors to induce one, two, or three broken rotor crossbars. Additionally, holes were drilled in the bearings to induce bearing faults.

[0240] To collect sound data for different fault types, a mobile phone's recording program was used to record motor fault sounds. The recording format was m4a, the encoder bit rate was 320kbps, and the sampling rate was selected at 44kHz. Sound signals from the motor under five different fault conditions were recorded, completing two experimental data acquisition scenarios: First, near-field fault sound acquisition, where the mobile phone was fixed 10cm away from the experimental motor, recording for 330 seconds; second, remote fault sound acquisition, where the microphone was placed more than 100cm away from the experimental motor, recording for 270 seconds. Finally, according to the settings in the recording program, 44,000 experimental samples were sampled per second.

[0241] (2) Performance Analysis

[0242] In practice, to verify the fault diagnosis effect of the proposed classification algorithm, linear support vector machine (SVM), BP neural network, and the DWT-LBP classification algorithm proposed in this patent were designed to classify the above motor faults. For each experimental case, each algorithm was run 1000 times, and the following statistical parameters were mainly examined: accuracy (CACC), unweighted average accuracy (UAP), F-measure (F1), and calculated geometric mean (GM), specifically defined as in equations (34)-(38):

[0243]

[0244]

[0245]

[0246]

[0247]

[0248] Among them, TP c ,TN c ,FP c ,FN c These represent the number of correct positive and negative numbers, and the number of incorrect positive and negative numbers, respectively. NC is the number of categories.

[0249] (3) Experimental conclusions

[0250] Finally, the above methods were used to classify the state of the faulty motor, and the classification results after 1000 algorithm iterations were obtained, as shown in Tables 2 and 3. Table 2 shows the motor fault classification results under the short-range fault sound detection scenario, and Table 3 shows the motor fault classification results under the long-range fault sound detection scenario.

[0251] Table 2. Motor Fault Classification Results (Short-Range Fault Sound Detection)

[0252]

[0253] Table 3. Motor Fault Classification Results (Remote Fault Sound Detection)

[0254]

[0255]

[0256] As can be seen from the table above, in motor fault classification using short-range fault sound detection, the UAP of the DWT-LBP method is 97.5, and in motor fault classification using long-range fault sound detection, the UAP of the DWT-LBP method is 94.5. The classification accuracy of the proposed fault classification method is higher than 94% in both cases. These results demonstrate the effectiveness of the proposed method. Therefore, this sound-based fault diagnosis method has good applicability and is easy to implement.

[0257] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A binary-mode-based and The method for diagnosing motor faults is characterized by, Includes the following steps: Step 1: Collect motor fault data and establish two types of motor fault sound datasets: near fault sound signal dataset and far fault sound signal dataset. Step 2: Based on DWT and 1D-BP technology, multi-level one-dimensional discrete wavelet transform (1D-DWT) technology is used to decompose and denoise the motor fault sound signal to obtain the original feature signal of the motor fault. The 1D-BP method was used to extract features from motor fault signals, and the nearest neighbor component analysis (NCA) method was used to effectively select motor fault features. Step 3: Based on Deep learning is used for motor fault diagnosis and classification. Algorithm optimization Weights, establish A classification model was developed and applied to the diagnosis of motor faults. Step 3a: Update the weights and biases of each layer of the model using the following equations: ； ； in, Indicates weight, Indicates bias. Indicates the number of floors. Represents the regularization parameter. Indicates the learning rate. This represents the total number of training samples. Indicates momentum. Indicates the update step, Represents the cost function, output unit and hidden units Mathematically, this can be represented by the following equation: ； ； in, This represents the weights between the input and the hidden layer; Step 3b: Using the beetle algorithm Optimize the weights of an ANN (Advanced Neural Network); The beetle algorithm is used in step 3b. Optimizing the weights of an ANN (Advanced Neural Network) involves the following steps: (1) Complete the random sampling of sample data by using random sampling techniques. To ensure that all search spaces are represented correctly One sample; (2) Initialization: The entire algorithm is in generation 0 of optimization. The optimized individuals are arbitrarily placed within a hypervolume equivalent related to the global search space to search for the optimal position vector. These individuals are pioneer individuals, each pioneer individual in the initial population. It can be arbitrarily placed into the global search space, and its specific definition is as follows: ； ； in, and This represents the upper and lower bound variables for optimization. For data dimensions; When all individuals identify the hypervolume location Each solution will be compared one by one, and the best individual will be obtained. The first population will be generated at the hypervolume space location. Each pioneer individual in the initial population is used to represent the weight value between the input layer and the hidden layer. Thus, the weight parameters and cost function are initialized. (3) New Hyperspace Selection Mode In this stage, each newly emerging individual will search for new solutions within the search space to create a total population belonging to the individual itself; among them, there are five hypervolume selection modes: neighbor search, global search, medium-sized search, large-scale search, and memory-considered search; Using the above search methods, the optimal individual position is established, and new precursor particles are arbitrarily placed using random sampling techniques. The search space obtained is as follows: ； ； in, Indicates the first Individual solution vector The elemental composition, Represents generating algebra, It is the first An element of a birth position or solution vector. and It is the first The second search The upper and lower limits of the dimension, The model parameters are represented; the above equation is iterated, and during each iteration, the weight and cost functions are analyzed by evaluating the fitness function and compared with the fitness function of the previous iteration to detect the optimal value, thereby re-initializing the weight and cost parameters; (4) Update population location The population's position is updated by updating the position of all individuals in the population. Except for the information stored in memory, all other information is deleted. Specifically, all previous populations disappear, and the newly discovered position becomes the birthplace of the new generation of the total population. Based on the previous rules, a new population is generated by this hypervolume pattern, and this process is repeated until the PBA reaches the stopping condition. Using random sampling techniques The uniform distribution function of each variable is divided into multiple segments with equal boundary probabilities. These segments are randomly selected for each variable, and the sample is defined by randomly mixing different variables, i.e., by... The range of each variable is divided into The sample is constructed by dividing the sample into non-overlapping equal segments, and then a dimension of is created by randomly assigning values ​​to all parameters. The sample; therefore, 4D space is divided into Segment, and randomly select a single value from each unit to generate a set of size. The sample; In response to A number of independent uniformly distributed variables ,Every The variables were divided into There are equal intervals, thus generating There are units, where each row and each column contains only one sample, and satisfies the following expression: ； in, It is a random number. yes The cumulative distribution function, in Duan He Given variables, the maximum number of combinations of the sampled data is: ； Typically, variables Defined within a given search space, bounded by a global boundary. and Definition, specifically, and It is the first Step 1 The low and high boundaries of each dimension, and the maximum distance of the samples. Also known as the tracking distance, it is calculated by the following expression: ； Therefore, the first Sampling techniques for each tracking step It can be described as: ； During the sampling process, it is necessary to Sampling is performed within the range and is located within the global search space, i.e. Any point outside this range will be corrected; Using the beetle algorithm When optimizing the weights of an ANN (Advanced Neural Network), the population position is updated, and the selection of the hypervolume mode is achieved according to the following formula: ； in, Indicates a large-scale factor. Indicates the medium size factor. Indicates adjacent factors.

2. The binary mode based method according to claim 1 and The method for diagnosing motor faults is characterized by, Step 1 involves collecting motor fault data, and the specific steps are as follows: Step 1 a: Identify and generate the fault types, including one healthy motor and four faulty motors, for a total of 4 fault types: one broken rotor bar fault, two broken rotor bar faults, three broken rotor bar faults, and bearing faults. Step 1b: Collect motor fault data, establish and store fault sound dataset. Record motor fault sound data, record the sound signals of the motor under five different fault conditions, and complete the collection of experimental data under two conditions; One is the acquisition of near-range fault sound; Secondly, remote fault sound acquisition; finally, 44,000 experimental samples are sampled per second to establish two types of motor fault sound datasets, namely near fault sound signal dataset and far fault sound signal dataset. The sound data is divided into 1-second samples as sound samples, and the sound data is stored and processed.

3. The binary mode based method according to claim 1 and The method for diagnosing motor faults is characterized by, In step 2, the process of using multi-level one-dimensional discrete wavelet transform (1D-DWT) technology to decompose and denoise the motor fault sound signal to obtain the original characteristic signal of the motor fault is as follows: The frequency coefficients of motor fault sound data are obtained using multi-level one-dimensional discrete wavelet transform (1D-DWT) technology, followed by signal decomposition. Five levels of 1D-DWT transformation are performed using Symlets 6 filters in wavelet transform, resulting in five low-pass filters at each level. The mathematical definition of multi-level discrete wavelet transform is as follows: ； ； It is a 1D-DWT function, and Symlets 6 is the filter used; thus, the original characteristic signal of the motor fault is obtained through the above low-pass filter subband and fault sound signal.

4. The binary mode based method according to claim 1 and The method for diagnosing motor faults is characterized by, In step 2, the 1D-BP method is used to extract features from the motor fault signal. In 1D-BP, since the signal is one-dimensional, its neighborhood relationship is extracted by using overlapping blocks of length 9. 1D-BP consists of several steps, including block processing, generating binary features, and constructing the BP signal. The specific steps are as follows: (1) Collect and load the fault sound signal of the three-phase induction motor; (2) Perform block processing, dividing the fault sound signal into blocks of length 9. The specific calculation formula is as follows: ； in, The sampling length of the signal; (3) Calculate the BP bit. Use the signum function to calculate the binary BP bit, as follows: ； ； in, It is the extracted first A binary value, and yes The input signal of the function shows that each value of the block is compared with the fifth value of the block to extract the relationship between the blocks; The low-pass filter coefficients generated by the BP() function above are used to extract feature signals from the original signal, as follows: ； ； ； ； ； ； in, , , , , , This represents a feature of size 256 at each level; (4) Calculate the decimal feature value. Calculate the decimal value corresponding to the extracted binary feature, as follows: ； in, It is the corresponding decimal characteristic value; (5) Feature cascade integration: The features extracted above are cascaded and integrated. The cascade process is completed according to the following formula; ； in, This is the final feature vector, with a size of 1536. It is a cascade function; (6) Normalization processing, for Each column of data is normalized as follows: ； in, These are the normalized features. and yes The minimum and maximum values.

5. The binary mode based method according to claim 1 and The method for diagnosing motor faults is characterized by, In step 2, nearest neighbor component analysis (NCA) is used to select motor fault features. Nearest neighbors are randomly selected, and the transformation matrix in the Mahalanobis distance is obtained by optimizing the leave-one-out cross-validation results. The details are as follows: (1) Calculate the Mahalanobis distance between two samples. Regarding the above information One input feature sample, Each has a class tag To find a distance that optimizes nearest neighbor classification, a Mahalanobis distance transformation matrix is ​​defined. It is a symmetric positive semi-definite matrix, that is Therefore, the Mahalanobis distance between two samples is defined as: ； in, ; (2) Use the leave-one-out method to calculate the classification error, and introduce differentiability. The function is as follows: ； The definition means sample points Randomly select a nearest neighbor, and finally select another sample point. As its neighbor, and inheriting its class tag The probability of the sample, therefore, The probability of correct classification is: ； in, ; (3) Maximize the objective function to complete the dimensionality reduction and selection of feature data. The objective function is to maximize the number of correctly classified points, and is defined as follows: ； This definition is a continuously differentiable matrix function. The optimization task of the algorithm is accomplished by maximizing this objective function, which is obtained using the conjugate gradient method. The conjugate gradient is calculated as follows: ； in, ; when yes The non-square matrix reduces the sample size to This allows for dimensionality reduction and sorting of the feature data. (4) Select 128 of the most typical features. Using the NCA method, select 128 of the most typical features as follows: ； in, It is the selected feature. It is a sorted feature implemented using the NCA method.

Images