Direct current motor fault diagnosis method based on improved local mean decomposition and composite multi-scale bubble entropy fusion

Abstract

The invention discloses a DC motor fault diagnosis method based on improved local mean decomposition and composite multi-scale bubble entropy fusion, and relates to the technical field of motor fault diagnosis. In order to solve the technical defect of insufficient fault feature extraction of the existing motor fault diagnosis technology in the prior art, the technical scheme provided by the invention comprises the following steps: collecting a sound signal in a running state of a direct current motor as an original signal; decomposing the original signal by adopting an improved local mean decomposition method to obtain a plurality of product components, and screening out effective feature components with high correlation with the original signal; performing composite multi-scale bubble entropy calculation on the effective feature component, and extracting a multi-dimensional entropy feature vector representing signal complexity; and inputting the multi-dimensional entropy feature vector into a particle swarm optimization extreme learning machine model for classification so as to obtain a fault diagnosis result of the DC motor. The method can be widely applied to real-time monitoring and intelligent fault identification of the running state of the direct-current motor in industrial production.

Description

Technical Field

[0001] It relates to the technical field of motor fault diagnosis, and specifically relates to a method for diagnosing DC motor faults. Background Art

[0002] As an important power device, DC motors are widely used in many key fields such as electric vehicles, intelligent manufacturing, and aerospace. With the increasingly complex operating environment of the equipment, motors are prone to typical faults such as rotor shaft bending, blade fracture, and bearing damage due to factors such as long-term operation causing mechanical wear, uneven rotor force, and overload. If these faults cannot be detected and diagnosed in time, it will cause equipment shutdown and production interruption, and even cause safety accidents in severe cases. Therefore, accurate and reliable fault diagnosis of DC motors has important practical significance.

[0003] Currently, the research in the field of DC motor fault diagnosis mainly focuses on feature extraction and recognition technologies based on vibration signals, acoustic signals, and current signals. Traditional diagnostic methods, such as those based on Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD), have made certain progress in the identification of typical motor faults. For example, Sun Hao et al. studied the method for diagnosing the rotor bar breakage fault of a cage-type motor based on vibration signals and current signals, pointing out that traditional methods are affected by noise interference in practical applications, which affects the diagnostic accuracy. Xia Zhiling et al. proposed a method for diagnosing the rotor bar breakage fault of an asynchronous motor using Variational Mode Decomposition (VMD), demonstrating its certain advantages in a strong noise environment; however, VMD still requires manual experience adjustment in parameter selection, resulting in limited applicability of the method. In addition, Wu Liyuan et al. used an improved Nonlinear Mode Decomposition algorithm (NME-EMD) and combined energy entropy to analyze the early rotor bar breakage fault characteristics. Although the diagnostic accuracy was improved, it also exposed the problem that the mode mixing problem is still difficult to completely avoid.

[0004] On the other hand, relevant progress has also been made in the field of blade fault and bearing damage diagnosis. For example, Zhang Junhua et al. used the starting signal to diagnose the crack and fracture state of the fan blade, and achieved high accuracy based on time-frequency analysis and pattern recognition technology; Jia Lin et al. optimized the VMD parameters using the particle swarm optimization (PSO) algorithm, thereby improving the discrimination of faults in the inner and outer rings of the bearing. However, these methods are mostly based on signal decomposition and feature extraction at a single scale, lacking a comprehensive analysis of the signal complexity at different scales, resulting in insufficient generalization ability in practical applications. In addition, traditional signal entropy analysis methods, such as multiscale entropy (MSE) or single-scale bubble entropy (BE), often lose important local feature information due to the coarse-graining process when analyzing non-linear and non-stationary signals.

[0005] Therefore, the following problems need to be solved urgently in the existing technology:

[0006] (1) Traditional signal decomposition methods are vulnerable to mode mixing, resulting in insufficient extraction of fault features and reducing the accuracy of diagnosis;

[0007] (2) Single-scale entropy analysis methods are difficult to comprehensively and effectively reflect the complexity characteristics of signals at multiple scales, restricting the effective extraction of deep information of fault signals;

[0008] (3) Existing fault diagnosis models lack automatic optimization of model parameters, resulting in large fluctuations in diagnosis performance and poor generalization ability.

[0009] Therefore, there is an urgent need to propose a DC motor fault diagnosis method that can effectively overcome the above deficiencies and improve the diagnosis accuracy and robustness. Summary of the Invention

[0010] To solve the technical defects existing in the prior art, namely, insufficient extraction of fault features, restriction of the effective extraction of deep information of fault signals, and large fluctuations in diagnosis performance in the existing motor fault diagnosis technology, the technical solution provided by the present invention is as follows:

[0011] A DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy, including:

[0012] The step of collecting the sound signal of the DC motor in the running state as the original signal;

[0013] The step of decomposing the original signal using an improved local mean decomposition method to obtain a number of product components and screening out effective feature components with high correlation with the original signal;

[0014] Steps of performing composite multi-scale bubble entropy calculation on the effective feature components and extracting a multi-dimensional entropy feature vector characterizing the signal complexity;

[0015] Steps of inputting the multi-dimensional entropy feature vector into a particle swarm optimization-based extreme learning machine model for classification to obtain the fault diagnosis result of the DC motor.

[0016] Furthermore, a preferred embodiment is provided. Sound signals during the operation of the motor are collected in real time by an acoustic sensor. The distance between the acoustic sensor and the motor housing is 10 cm, and two sensors are arranged orthogonally at 90°.

[0017] Furthermore, a preferred embodiment is provided. An integral sliding window strategy is used to calculate the local mean of the adjacent extreme point intervals, and a local mean function with high smoothness is obtained through moving average filtering to suppress mode mixing and baseline noise interference.

[0018] Furthermore, a preferred embodiment is provided. The method for screening effective feature components is: calculating the correlation coefficient between each product component and the original signal, and removing the components with a correlation coefficient lower than 0.3.

[0019] Furthermore, a preferred embodiment is provided. The scale factor of the composite multi-scale bubble entropy calculation is 10, the embedding dimension is 2, the time delay is 1, and the similarity tolerance is 0.15 times the standard deviation of the effective feature components.

[0020] Furthermore, a preferred embodiment is provided. In the particle swarm optimization-based extreme learning machine model, the number of particles in the particle swarm optimization algorithm is 30, the maximum number of iterations is 100, and the number of hidden layer nodes is 100.

[0021] Based on the same inventive concept, the present invention also provides a DC motor fault diagnosis device based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy, including:

[0022] A module for collecting sound signals during the operation state of the DC motor as the original signal;

[0023] A module for decomposing the original signal using an improved local mean decomposition method to obtain a number of product components and screening out effective feature components with high correlation with the original signal;

[0024] A module for performing composite multi-scale bubble entropy calculation on the effective feature components and extracting a multi-dimensional entropy feature vector characterizing the signal complexity;

[0025] A module for inputting the multi-dimensional entropy feature vector into a particle swarm optimization-based extreme learning machine model for classification to obtain the fault diagnosis result of the DC motor.

[0026] Based on the same inventive concept, the present invention also provides a computer storage medium for storing a computing program, and when the computer program is read by a computer, the computer executes the method described above.

[0027] Based on the same inventive concept, the present invention also provides a computer, including a processor and a storage medium, and when the processor reads the computer program stored in the storage medium, the computer executes the method described above.

[0028] Based on the same inventive concept, the present invention also provides a computer program product, which is a computer program, and when the computer program is executed, the method described above is implemented.

[0029] Compared with the prior art, the beneficial effects of the technical solution provided by the present invention are as follows:

[0030] This solution uses an improved local mean decomposition (LIMD) method, adopts an integral sliding window strategy and adaptive boundary processing, can effectively suppress the endpoint effect and baseline noise, and significantly reduce the mode mixing phenomenon in the signal decomposition process, so as to extract more accurate fault-related feature information. Compared with the traditional local mean decomposition (LMD) method, the LIMD method of this solution can obtain clearer time-frequency spectrum features, ensure that the effective signal is not interfered by noise during the fault diagnosis process, and fundamentally improve the accuracy of fault feature extraction.

[0031] This solution introduces composite multi-scale bubble entropy (CMBE) to analyze the complexity of the decomposed signal. By improving the generation method of the coarse-grained sequence, it effectively solves the problems of information loss and inaccurate entropy estimation existing in the traditional single-scale or simple multi-scale entropy method under long-scale factors, and improves the comprehensiveness and stability of the signal complexity features. Compared with the existing single-scale entropy analysis method, the CMBE of this solution can reliably reflect the small differences between different fault signals under multi-scale conditions, and improve the ability to distinguish fault types.

[0032] This solution applies the particle swarm optimization algorithm (PSO) to the parameter optimization process of the extreme learning machine (ELM), realizes the adaptive optimization of the hidden layer connection weights and bias vectors, and overcomes the problem of model performance fluctuation caused by the random initialization of the traditional ELM model parameters. The experimental results show that the ELM model optimized by PSO has higher recognition accuracy and generalization ability. Compared with the unoptimized ELM model, the accuracy rate is increased by 2.8125 percentage points, further improving the robustness and stability of fault diagnosis.

[0033] In summary, the proposed LIMD-CMBE-PSO-ELM fusion model in this solution is superior to existing signal processing and diagnosis methods in terms of the accuracy of DC motor multi-class fault diagnosis, the robustness of feature extraction, and the generalization performance. The overall diagnosis accuracy rate reaches 99.6875%, providing an effective and advanced solution for the intelligent fault diagnosis of DC motors.

[0034] It can be widely applied to the real-time monitoring of the operating state of DC motors and the intelligent fault identification in industrial production. Description of the Drawings

[0035] Figure 1 It is the flow chart of the DC motor fault diagnosis method;

[0036] Figure 2 It is the LIMD decomposition diagram of the normal state signal. Among them, (a) is the time-domain diagram of the LIMD decomposition of the normal state signal, and (b) is the time-frequency spectrum diagram of the LIMD decomposition of the normal state signal;

[0037] Figure 3 It is the LIMD decomposition diagram of the rotor shaft bending signal. Among them, (a) is the time-domain diagram of the LIMD decomposition of the rotor shaft bending signal, and (b) is the time-frequency spectrum diagram of the LIMD decomposition of the rotor shaft bending signal;

[0038] Figure 4 It is the LIMD decomposition diagram of the blade fracture signal. Among them, (a) is the time-domain diagram of the LIMD decomposition of the blade fracture signal, and (b) is the time-frequency spectrum diagram of the LIMD decomposition of the blade fracture signal;

[0039] Figure 5 It is the LIMD decomposition diagram of the bearing fault signal. Among them, (a) is the time-domain diagram of the LIMD decomposition of the bearing fault signal, and (b) is the time-frequency spectrum diagram of the LIMD decomposition of the bearing fault signal;

[0040] Figure 6 It is the LMD decomposition diagram of the bearing fault signal. Among them, (a) is the time-domain diagram of the LMD decomposition of the bearing fault signal, and (b) is the time-frequency spectrum diagram of the LMD decomposition of the bearing fault signal;

[0041] Figure 7 It is the schematic diagram of the classification result of the LMD-ELM method;

[0042] Figure 8 It is the schematic diagram of the classification result of the LMD-PSO-ELM method;

[0043] Figure 9 It is the schematic diagram of the classification result of the LIMD-ELM method;

[0044] Figure 10 It is the schematic diagram of the classification result of the LIMD-PSO-ELM method;

[0045] Figure 11 It is the fitness curve graph of the LMD - PSO - ELM method;

[0046] Figure 12 It is the fitness curve graph of the LIMD - PSO - ELM method. Specific implementation manners

[0047] To make the advantages and beneficial effects of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention will be further described in detail with reference to the accompanying drawings. Specifically:

[0048] Embodiment 1. This embodiment provides a DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi - scale bubble entropy, including:

[0049] The step of collecting the sound signal of the DC motor under the running state as the original signal;

[0050] The step of decomposing the original signal by using the improved local mean decomposition method to obtain a number of product components and screening out the effective characteristic components with high correlation with the original signal;

[0051] The step of calculating the composite multi - scale bubble entropy of the effective characteristic components to extract the multi - dimensional entropy feature vector characterizing the signal complexity;

[0052] The step of inputting the multi - dimensional entropy feature vector into the particle swarm optimization - based extreme learning machine model for classification to obtain the fault diagnosis result of the DC motor.

[0053] An acoustic sensor is used to collect the sound signal during the operation of the motor in real time. The distance between the acoustic sensor and the motor housing is 10 cm, and two sensors are arranged orthogonally at 90°.

[0054] An integral sliding window strategy is used to calculate the local mean of the adjacent extreme point intervals, and a local mean function with high smoothness is obtained through moving average filtering to suppress mode mixing and baseline noise interference.

[0055] The method for screening effective characteristic components is: calculating the correlation coefficient between each product component and the original signal, and removing the components with a correlation coefficient lower than 0.3.

[0056] The scale factor of the composite multi - scale bubble entropy calculation is 10, the embedding dimension is 2, the time delay is 1, and the similarity tolerance is 0.15 times the standard deviation of the effective characteristic components.

[0057] In the particle swarm optimization - based extreme learning machine model, the number of particles of the particle swarm optimization algorithm is 30, the maximum number of iterations is 100, and the number of hidden layer nodes is 100.

[0058] Embodiment 2. This embodiment further describes in detail the above-provided technical solution. Specifically:

[0059] The following is the content written for the specific embodiment in the patent application document:

[0060] This solution provides a DC motor fault diagnosis method based on the fusion of improved local mean decomposition (LIMD) and composite multi-scale bubble entropy (CMBE). Specifically, this method includes the following steps:

[0061] First, use an acoustic sensor to collect the sound signal generated during the operation of the DC motor in real time. In this embodiment, a Siemens LMS data acquisition system is selected, and two GPRAS acoustic sensors with a sensitivity of 50 mV / Pa are used for data acquisition. The sensors are arranged at a position 10 cm away from the motor housing and are orthogonally arranged at 90° to avoid interference effects. The sampling frequency is set to 12.8 kHz to obtain the original sound signals in the normal operating state of the motor and various fault states including rotor shaft bending, blade fracture, and bearing damage. The length of each group of signals is 50,000 sampling points.

[0062] Second, perform improved local mean decomposition (LIMD) on the collected sound signals. Specifically, the integral sliding window method is used to determine the local mean points of the original signal, and the integral average value of the signal within the interval between every two adjacent extreme points is calculated to form a local mean function with high smoothness. Through this method, the endpoint effect, mode mixing, and baseline noise interference existing in the traditional local mean decomposition process are effectively reduced, and several product function (PF) components highly correlated with the original signal are obtained. Subsequently, calculate the correlation coefficient between each PF component and the original signal, and remove the redundant components with a correlation coefficient lower than the preset threshold (set to 0.3 in this embodiment), and retain the effective feature components with strong correlation (such as PF1 and PF2 components) for further analysis.

[0063] Third, perform composite multi-scale bubble entropy (CMBE) feature extraction on the effective feature components obtained in the above steps. The specific process is as follows: First, perform multi-scale coarse-graining processing on the above effective feature components respectively, divide the original time series into subsequences with different scale factors, and use a sliding window to generate overlapping subsequences to ensure that the details of the original signal can be fully retained after coarse-graining. Then calculate the corresponding bubble entropy value at each scale, and further construct a multi-dimensional entropy feature vector that can reflect the complexity characteristics of the signal (in this embodiment, the scale factor is set to 10 to form a 10-dimensional entropy feature vector). The entropy feature vector extracted by this method can effectively quantify the non-linear complexity information of the signal at different scales, which helps to improve the classification accuracy of the fault state.

[0064] Finally, the obtained multi-dimensional entropy feature vector is input into a diagnosis model of an extreme learning machine (ELM) based on particle swarm optimization (PSO) for classification. Specifically, first, the input weights and biases of the hidden layer nodes of the extreme learning machine are randomly initialized, and the above parameters of the ELM model are globally optimized through the particle swarm optimization algorithm to obtain the optimal hidden layer parameters that match the distribution of fault feature data. Secondly, the output weights of the ELM model are quickly solved by the least squares method to establish a mapping relationship, so as to realize the accurate identification and diagnosis of the normal state and various fault states (such as rotor shaft bending, blade fracture, bearing fault) of the DC motor. When using this solution for diagnosis, a classification accuracy of 99.6875% is achieved on the experimental data set, which is significantly higher than that of the traditional LMD-ELM and the unoptimized LIMD-ELM models, showing obvious superiority and reliability.

[0065] The following further elaborates on the specific implementation manners in the above steps:

[0066] (1) Specific implementation manner of data acquisition

[0067] In this embodiment, the experimental objects are three common types of micro DC motors, namely 555, 775, and 895. The operating states of the motors include normal operation, rotor shaft bending, blade fracture, and bearing fault states. 400 groups of sound signals are collected for each state, totaling 1600 groups of sample data. During the data acquisition process, the sensitivity of the acoustic sensor is set to 50 mV / Pa, the sampling frequency is fixed at 12.8 kHz, and the length of each group of signal data is 50,000 points, ensuring that the data can truly reflect the operating state of the motor.

[0068] (2) Specific implementation manner of improved local mean decomposition (LIMD)

[0069] Taking the collected original sound signal as the input signal, first, the extreme point detection algorithm (setting the extreme tolerance threshold to ±5%) is used to extract local extreme points, and then the integral sliding window mean is calculated within each adjacent extreme point interval to obtain the local statistical characteristics of the signal. Further, through moving average filtering processing, a smooth local mean function is obtained, and based on this, the original signal is gradually decomposed to obtain multiple product function (PF) components with clear physical meanings. Subsequently, the correlation coefficient is calculated between each PF component and the original signal, and the PF components with a correlation coefficient greater than 0.3 are selected as effective feature components for subsequent feature extraction.

[0070] (3) Specific implementation manner of composite multi-scale bubble entropy (CMBE) feature extraction

[0071] For the effective feature components screened above, set the scale factor to 10, the embedding dimension to 2, the time delay to 1, and the similarity tolerance to 0.15 times the standard deviation of the original signal. Perform coarse-grained sequence processing and use a sliding window to generate overlapping subsequences to ensure the integrity of signal feature details. At each scale factor, use bubble entropy to analyze the complexity of the coarse-grained subsequences and construct a 10-dimensional entropy feature vector that can comprehensively describe the complexity characteristics of the signal as the input features of the classifier.

[0072] (4) Specific implementation of particle swarm optimization extreme learning machine (PSO-ELM)

[0073] First, input the entropy feature vector into the ELM model, and randomly generate the input weight matrix and bias vector of the hidden layer nodes. Use the particle swarm optimization algorithm, with the root mean square error of the diagnostic model on the training set as the fitness function, and globally optimize the ELM hidden layer parameters by iteratively updating the particle positions and velocities. After sufficient iteration, obtain the optimal input weight matrix and bias parameters. Finally, determine the output weight matrix based on the least squares method to construct a high-precision fault diagnosis model, and finally realize the accurate identification and classification of different states of the DC motor.

[0074] In summary, the technical solution provided by this embodiment effectively suppresses modal aliasing and noise interference through the improved local mean decomposition (LIMD) method, fully excavates the signal complexity information through the composite multi-scale bubble entropy (CMBE) method, and combines the extreme learning machine optimized by particle swarm optimization (PSO-ELM) to achieve precise optimization of the diagnostic model. Finally, it greatly improves the accuracy and robustness of DC motor fault diagnosis, can fully achieve the technical effects proposed by this solution, and has complete feasibility.

[0075] Embodiment 3. Combination Figures 1-12 To illustrate this embodiment, this embodiment further describes the above-provided technical solution in detail through specific examples. Specifically:

[0076] Local mean decomposition

[0077] Local mean decomposition (LMD) is an adaptive signal decoupling method, and its core goal is to decompose non-stationary and multi-component complex signals into a series of single-component amplitude-modulated and frequency-modulated signals (product function components PF) and trend term residuals with clear physical meanings. Each product function (PF) component is composed of the product of an envelope function and a pure frequency-modulated signal. Among them, the envelope function represents the instantaneous amplitude of this component, and the instantaneous frequency can be realized by performing a first-order derivative operation on the phase function of the pure frequency-modulated signal. The LMD algorithm gradually extracts PF components through an iterative screening process, and its specific decomposition steps are as follows:

[0078] (1) Let the original sound signal be x(n), where n = 0, 1, …, N and N is the length of the signal. Find all the local maximum points and local minimum points of x(n). The position corresponding to the i-th maximum or minimum point is n i , and calculate n i and the average value m i+1 of the adjacent extreme points n i :

[0079]

[0080] Extend all the average values m i linearly between the extreme points n i and n i+1 , and use the moving average method to smooth the extended line to obtain the local mean function m 11 (n). The first subscript of m 11 (n) refers to the first local mean function, and the second subscript refers to the first iteration, and so on.

[0081] (2) Calculate half of the absolute value of the difference between n i and the adjacent extreme point n i+1 to obtain the local amplitude a i :

[0082]

[0083] Extend all the local amplitudes a i linearly between the extreme points n i and n i+1 , and use the moving average method to smooth the extended line to obtain the envelope estimation function a 11 (n). The first subscript of a 11 (n) refers to the first envelope estimation function, and the second subscript refers to the first iteration, and so on.

[0084] (3) Separate the local mean function m 11 from the original sound signal x(n) to obtain the zero-mean function h 11 :

[0085] h 11 (n) = x(n) - m 11 (n) (3)

[0086] (4) Demodulate h 11 (n), which is achieved by dividing h 11 (n) by the envelope estimation function a 11 (n), to obtain the demodulated function s 11 :

[0087]

[0088] At this time, it is necessary to determine whether s 11 (n) is a pure frequency modulation function (the amplitude of a pure frequency modulation function is always 1, and -1 ≤ s 11 (n) ≤ 1). According to the method in step (2), find the envelope estimation function a 11 (n) of s 12 (n). If a 12 (n) = 1, it means that s 11 (n) is an ideal pure frequency modulation function; if a 12 (n) ≠ 1, then s 11 (n) needs to be used as the initial signal, and repeat steps (1) to (4) until a pure frequency modulation function s 1o (n) is obtained. When the iteration terminates, there is:

[0089]

[0090] Among them

[0091]

[0092] (5) Multiply all the envelope estimation functions calculated in step (2) during the iteration process to obtain the envelope signal a1(n):

[0093]

[0094] (6) Multiply the envelope signal a1(n) generated in step (5) by the pure frequency modulation function s 1o (n) generated in step (4). The obtained product is the first PF component of the LMD decomposition, denoted as PF1(n):

[0095] PF1(n) = a 1o (n)s 1o (n) (8)

[0096] (7) Finally, subtract the PF1(n) component from the original sound signal x(n) to obtain a new signal u1(n). Take this signal as the original signal, repeat steps (1) to (6) and perform k cycles until u k (n) is a monotonic function:

[0097]

[0098] After multiple cycles of iterative decomposition, the original sound signal x(n) is finally decomposed into the sum of k PF components and a residue u k (n), that is:

[0099]

[0100] Improved Local Mean Decomposition

[0101] Although the Local Mean Decomposition (LMD) optimizes the symmetry of extreme points and the efficiency of mode separation based on the Empirical Mode Decomposition (EMD) framework, its essence still belongs to a data-driven adaptive decomposition method. It is difficult to completely eliminate the mode mixing phenomenon, resulting in band overlap and energy leakage in the decomposed components, which in turn affects the accurate extraction of fault features and the reliability of diagnostic results. To break through this bottleneck, this paper proposes a local integral mean optimization strategy to suppress the interference of mode mixing on signal decomposition by reconstructing the calculation processes of the local mean function and envelope estimation. Specifically, the traditional local mean idea is abandoned, and a method based on local integral mean is introduced. The core steps are as follows:

[0102] (1) Find all the maximum and minimum points in the original sound signal x(n), and arrange the extreme points in ascending order to form (n k , x k )(n k , x k ), where n k is the position index of the extreme point, and x k is the signal amplitude corresponding to the extreme point.

[0103] (2) For each pair of adjacent extreme points (n k , x k ) and (n k + 1, x k - 1), calculate the local integral mean within their interval where is the midpoint between adjacent extreme points, and is the integral mean of the signal between adjacent extreme points:

[0104]

[0105] (3) Use the moving average algorithm to smooth all the local integral means to obtain the local mean function m 11 (n).

[0106] (4) The other steps are the same as steps (2) - (7) of the local mean decomposition.

[0107] The sound signals of four states of a DC motor collected are decomposed based on improved local mean decomposition to obtain five product function components. Then, the correlation coefficients of these five product function components are calculated, and further, the characteristic components strongly correlated with the original signal are screened out. Table 1 shows the calculation results of the correlation coefficients. Analysis shows that the correlation coefficients of the PF1 component and the PF2 component are both greater than 0.3. These components contain the main characteristics of the original signal. The PF3 - PF5 components mainly contain environmental noise and high - frequency interference components, which are removed to reduce the feature redundancy. Finally, the PF1 and PF2 components are selected as effective feature carriers to replace the original signal and input into the composite multi - scale bubble entropy (CMBE) model for feature extraction.

[0108] Table 1 Calculation Results of Correlation Coefficients

[0109]

[0110] Extraction of Composite Multi - Scale Bubble Entropy Feature Vectors

[0111] Composite multi - scale bubble entropy (CMBE) is an improved multi - scale entropy analysis method, aiming to solve the problem of the decline in entropy estimation accuracy caused by the shortening of the signal coarse - grained sequence under long scale factors in traditional multi - scale bubble entropy (MBE).

[0112] The specific steps of the composite multi - scale entropy are as follows:

[0113] (1) Select the PF1 component and the PF2 component as effective feature carriers to replace the original signal. At this time, the signal is the signal after improved local mean decomposition, denoted as f(n), where n = 0, 1, …, N, and N is the length of the signal. Conduct coarse - graining calculation. For each given scale factor s (representing the time - window length), divide the sequence after improved local mean decomposition into non - overlapping windows, and calculate the mean of each segment:

[0114]

[0115] In the formula, y (s) (e) is the e - th coarse - grained subsequence when the scale factor is s, f(g) is the g - th sequence of the sequence after improved local mean decomposition, e ranges from 1 to is the number of coarse - grained sequences.

[0116] Since the sequence length is shortened to resulting in a decline in entropy estimation accuracy, through composite processing, overlapping subsequences are generated using a sliding window. Overlapping subsequences are generated with a step size of Δt = 1 to generate s groups of overlapping subsequences, ensuring that the total length of the coarse - grained sequence remains n - s + 1. The mean of its subsequences is:

[0117]

[0118] (2) For each coarse-grained subsequence y (s) (e) Calculate the bubble entropy. Set the embedding dimension m and the time delay u, and construct the phase space vector:

[0119]

[0120] where F(n) is the phase space vector at point n, and {y(n), y(n + u), …, y(n + (m - 1)u)} are m consecutive samples extracted from the coarse-grained sequence y, with n = 1, 2, …, l - (m - 1)u, where l is the subsequence length,

[0121] Arrange each embedding vector F(n) in ascending order, and record the number of element exchanges j(n). Statistically analyze the probability distribution of the number of exchanges:

[0122]

[0123] where P(j) is the probability of the number of exchanges, and the value of the time delay u is 1.

[0124] Calculate the Rényi entropy of order 2 based on the probability distribution:

[0125]

[0126] where H swaps is the Rényi entropy, p j is the probability of the jth event, and lg is the natural logarithm.

[0127] By adjusting the embedding dimension m, calculate the normalized bubble entropy value:

[0128]

[0129] where and represent the Rényi entropy values calculated at embedding dimensions m + 1 and m respectively, and BE is the normalized bubble entropy value.

[0130] (3) Calculation of the composite multi-scale bubble entropy. For the scale factor s, aggregate the mean values of the bubble entropy values of all coarse-grained subsequences:

[0131]

[0132] where CMBE(s) is the composite multi-scale bubble entropy value at the scale factor s, n is the total length of the signal after improved local mean decomposition, y (s) (e) is the e-th coarse-grained subsequence at the scale factor s, BE(y (s)(e) is the bubble entropy value of the e-th coarse-grained subsequence, and n - s + 1 is the total number of coarse-grained subsequences at scale s.

[0133] The main parameters affecting the calculation accuracy of the composite multi-scale bubble entropy are the scale factor s, embedding dimension m, time delay u, and similarity tolerance r. When extracting the composite multi-scale bubble entropy feature vector, the scale factor s is set to 10, the embedding dimension m is set to 2, the time delay u is set to 1, and the similarity tolerance r is set to 0.15×σ, where σ is the standard deviation of the original signal. Based on the composite multi-scale bubble entropy (CMBE), a multi-dimensional feature vector T characterizing the complexity of the acoustic signal is constructed. Its dimension is determined by the scale factor s = 10, i.e., T = [CMBE1, CMBE2,..., CMBE 10 .

[0134] PSO-ELM model

[0135] The Extreme Learning Machine (ELM) was proposed by scholars such as Huang and is an efficient supervised learning framework based on the Single-hidden Layer Feedforward Neural Network (SLFN). Its core innovation lies in: randomly fixing the hidden layer parameters (input weight matrix W and bias vector B) and directly solving the output weights by the least squares method, thus significantly reducing the number of training parameters, improving the learning speed, and enhancing the model generalization ability. However, due to the random generation characteristics of W and B, it cannot adaptively match the data distribution characteristics, which may lead to model performance fluctuations and limited accuracy. To overcome the above defects, the Particle Swarm Optimization (PSO) algorithm is introduced to globally optimize the hidden layer parameters of ELM, and a hybrid diagnostic model PSO-ELM is constructed. PSO is based on the swarm intelligence mechanism and can efficiently converge by jointly searching for the individual historical optimal and the global optimal solutions of the group, with both multi-objective optimization ability and efficient convergence characteristics. By dynamically optimizing W and B through PSO, the feature mapping ability and diagnostic accuracy of ELM can be effectively improved while maintaining its fast training advantage. The specific steps are as follows:

[0136] (1) The core innovation of the Extreme Learning Machine (ELM) lies in the random fixation of the hidden layer parameters and the interpretation and solution of the output weights.

[0137] Random fixation of hidden layer parameters: The input weight matrix W ∈ R d×L and the bias vector B ∈ R L are randomly initialized and then fixed (d is the input feature dimension, and L is the number of hidden layer nodes).

[0138] Output weight analytical solution: The output weight β ∈ R is directly calculated through the Moore-Penrose generalized inverse L×z , and the mathematical expression is:

[0139]

[0140] where H ∈ R v×L is the hidden layer output matrix, whose dimension is v × L, where v is the number of training samples and L is the number of hidden layer nodes; Z ∈ R v×z is the target matrix, whose dimension is v × z, where z is the number of output classes. λ is the regularization coefficient, E is the identity matrix, whose dimension is the same as that of H T H, is the Moore-Penrose generalized inverse, which is used to calculate the inverse matrix of H T H plus the regularization term, and H T is the transpose of H, with dimension L × v.

[0141] (2) Particle Swarm Optimization (PSO) fusion strategy: To overcome the parameter sensitivity of ELM, the Particle Swarm Optimization algorithm is introduced to construct a hybrid diagnostic model PSO-ELM. The specific steps are as follows:

[0142] First, the input weight matrix W and the bias vector B in the hidden layer parameters are concatenated into a particle position vector:

[0143] P ii =[vec(W (ii) ),B (ii) ∈ R (d+1)L (20)

[0144] where P ii is the position vector of the ii-th particle, and vec(W (ii) ) represents the column vectorization operation of the input weight matrix W (ii) , B (ii) represents the bias vector of the ii-th particle, and R (d+1)L represents the (d + 1)L-dimensional space over the real number field.

[0145] Then, the fitness function is designed, and the weighted root mean square error (RMSE) is used as the target fitness function.

[0146]

[0147] where v is the number of training samples; z is the number of output classes; tjjll is the true label of the ll-th dimension of the jj-th sample; β mmll is the connection weight from the mm-th node in the hidden layer to the ll-th node in the output layer; g(·) is the hidden layer activation function (Sigmoid function); is the dot product of the weight vector from the input layer to the m-th hidden layer node and the input sample x jj ; b mm is the bias value of the m-th hidden layer node.

[0148] (3) Perform particle update and convergence control.

[0149] First, perform velocity update:

[0150]

[0151] In the formula, v (t+1)ii is the velocity of the i-th particle at time t+1, w is the inertia weight (w = 0.7), c1 and c2 are learning factors (both c1 and c2 are 2), random numbers r1 and r2 are randomly generated in the range [0,1], p best,ii is the individual optimal position of the i-th particle, g best is the global optimal position, p (t)ii is the position of the i-th particle at time t.

[0152] Then, perform position update:

[0153]

[0154] In the formula, p (t+1)ii is the position of the i-th particle at time t+1, p (t)ii is the position of the i-th particle at time t.

[0155] Finally, when the termination condition is reached, when RMSE < 10 -3 or the maximum number of iterations T max = 100 is reached, output the global optimal solution P*, that is, output the global optimal solution g best , * indicates the optimal solution.

[0156] (4) Substitute the optimized {W*, B*} into the ELM framework, analytically calculate the output weight β*, and form the PSO-ELM classifier.

[0157] The flow chart of the PSO-ELM model is as Figure 1 shown.

[0158] Experimental Results and Analysis

[0159] Based on the improved local mean decomposition (LIMD) method, the acoustic signals of normal state, rotor shaft bending, blade fracture, and bearing failure are decomposed, and 5 product function (PF) components with clear physical meanings are extracted respectively under each fault state. The decomposition results (including time-domain PF components and time-frequency spectra) are as Figures 2 to 5 shown.

[0160] To verify the optimization effect of the improved Local Integral Mean Decomposition (LIMD), taking the bearing fault signal as an example, the decomposition performances of LIMD and the classical LMD are compared and analyzed. Figure 6 The time-domain diagram and time-frequency diagram of the bearing fault signal decomposition based on LMD are shown. Combining Figure 5 with the LIMD decomposition results, the following conclusions can be drawn: By comparing the time-domain diagram and time-frequency diagram of the bearing fault signal decomposition of LIMD and LMD, it can be found that the extraction of high-frequency components of LIMD decomposition in PF1 and PF2 is more uniform, reducing the mode mixing problem. At the same time, LIMD decomposition is less sensitive to noise and the decomposition results are more stable. In contrast, although LMD decomposition can gradually decompose the high-frequency to low-frequency components in the signal, there are certain mode mixing problems in the extraction of high-frequency components and it is more sensitive to noise. Therefore, LIMD is superior to LMD in terms of signal decomposition effect.

[0161] Four types of DC motor acoustic signals are collected in the experiment: normal state, rotor shaft bending, blade fracture, and bearing fault. Each type of state contains 400 groups of samples (sampling frequency is 12,800 Hz, the duration of each type of state is 2,000 s, and the data length of each group of samples is 50,000), totaling 1,600 groups of data. They are divided into a training set (1,280 groups) and a test set (320 groups) according to a 4:1 ratio, and the label encoding is c = [1, 2, 3, 4]. Configuration of the Extreme Learning Machine (ELM): The number of hidden layer nodes L = 100, and the activation function is Sigmoid; the input layer dimension d = 10, and the output layer dimension z = 4. Particle Swarm Optimization (PSO) parameter settings: The number of particles is 30, the maximum number of iterations T max is 100, the inertia weight w is 0.7, the learning factors c1 and c2 are both 2, the random numbers r1 and r2 are randomly generated within the range of [0, 1], and the weight coefficient α = 0.7. First, the signals are decomposed by the traditional LMD decomposition method and the improved LIMD decomposition method respectively. Subsequently, the Pearson correlation coefficients between each Product Function (PF) component and the original signal are calculated, and the components with correlation coefficients lower than 0.3 are removed. After reconstructing the retained PF components, the composite multi-scale bubble entropy value of the reconstructed signal is calculated, and the identification of four types of acoustic signal data is realized by using the Extreme Learning Machine (ELM) classifier and the Extreme Learning Machine (ELM) after Particle Swarm Optimization (PSO). The classification results based on the LMD-ELM method, LMD-PSO-ELM method, LIMD-ELM method, and LIMD-PSO-ELM method are respectively as Figure 7 、 Figure 8 、 Figure 9 、 Figure 10As shown. By analyzing these four figures, we can know that the accuracy of the test set of the LMD-ELM method is 50.625%. Feature noise causes the acoustic signals of DC motors in normal state, DC motors with blade fractures, and DC motors with bearing faults to be confused with each other. While the LMD-PSO-ELM method increases the accuracy to 79.0625%, which is sufficient to show that PSO optimization can significantly suppress parameter sensitivity and improve the accuracy of fault diagnosis. The accuracy of the test set of the LIMD-ELM method is 96.875%, indicating that LIMD decomposition enhances feature separability and has a stronger ability to improve the accuracy of fault diagnosis than PSO optimization. The LIMD-PSO-ELM method has the highest accuracy of 99.6875%, which verifies that the PSO algorithm and the LIMD decomposition algorithm can be optimized synergistically to improve the accuracy of fault diagnosis.

[0162] Figure 11 , Figure 12 are respectively the fitness curve graphs of particle swarm optimization in the LMD-PSO-ELM method and the fitness curve graphs of particle swarm optimization in the LIMD-PSO-ELM method. By analyzing these two figures, we can know that in Figure 11 , when the number of iterations is 0 - 20 times, the accuracy rapidly increases from 0.5 to 0.7, indicating that the PSO algorithm quickly explores the region of better solutions in the early stage; when the number of iterations is 20 - 70 times, the accuracy slowly rises to about 0.75, reflecting the dynamic balance between global search and local optimization of the particle swarm, and there may be feature interference caused by mode mixing. When the number of iterations is 70 - 100 times, the diagnosis starts. The experimental results show that the accuracy of this method increased by 4% in a certain period among 320 groups of test samples, and then the accuracy remained unchanged, indicating that the PSO algorithm gradually converges and finally stabilizes at the optimal solution. The reason for the low accuracy may be insufficient particle diversity or too high inertia weight, which affects the fine search ability in the later stage. In Figure 12 , the accuracy starts at about 97%, verifying that the improved LIMD method directly provides highly discriminative features through noise suppression and mode separation. In the initial stage of iteration, the accuracy increases from 97% to 99%, and the accuracy improvement is relatively fast. While in the middle stage of iteration, the accuracy has been maintained at about 99.5%, indicating that the PSO algorithm has approached the global optimum. At the 90th iteration, the accuracy increased by about 0.3% again, indicating that the PSO algorithm has found the global optimum solution, and the accuracy at this time reaches the highest, which is 96.875%. The LIMD-PSO-ELM method only needs to iterate 10 times to approach the peak performance, indicating the high efficiency of the LIMD-PSO-ELM method. Figure 11 The initial correct rate of Figure 1050.625% of it indicates that the LIMD decomposition algorithm provides feature inputs with a higher signal-to-noise ratio through integral sliding windows and noise suppression.

[0163] The technical solutions provided by the present invention are further described in detail through several specific embodiments to highlight the advantages and benefits of the technical solutions provided by the present invention. However, the several specific embodiments described above are not used as a limitation to the present invention. Any reasonable modifications and improvements to the present invention, combinations of implementation manners, equivalent replacements, etc. within the spirit and principle scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy, characterized in that, Including: The step of collecting the sound signal of the DC motor in the running state as the original signal; The step of decomposing the original signal by using the improved local mean decomposition method to obtain a number of product components and screening out the effective feature components with high correlation with the original signal; The step of calculating the composite multi-scale bubble entropy of the effective feature components and extracting the multi-dimensional entropy feature vector characterizing the signal complexity; The step of inputting the multi-dimensional entropy feature vector into the extreme learning machine model optimized by the particle swarm optimization for classification to obtain the fault diagnosis result of the DC motor.

2. The DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy according to claim 1, wherein, The sound signal during the operation of the motor is collected in real time by using an acoustic sensor, the distance between the acoustic sensor and the motor housing is 10 cm, and the two sensors are arranged orthogonally at 90°.

3. The DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy according to claim 1, characterized in that, The integral sliding window strategy is used to calculate the local mean of the adjacent extreme point intervals, and the local mean function with high smoothness is obtained through the moving average filtering process to suppress the mode mixing and baseline noise interference.

4. The DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy according to claim 1, characterized in that, The method for screening the effective feature components is: calculating the correlation coefficient between each product component and the original signal, and removing the components with the correlation coefficient lower than 0.

3.

5. The DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy according to claim 1, characterized in that, The scale factor of the composite multi-scale bubble entropy calculation is 10, the embedding dimension is 2, the time delay is 1, and the similarity tolerance is 0.15 times the standard deviation of the effective feature components.

6. The DC motor fault diagnosis method based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy according to claim 1, characterized in that, In the extreme learning machine model optimized by the particle swarm optimization, the number of particles of the particle swarm optimization algorithm is 30, the maximum number of iterations is 100, and the number of hidden layer nodes is 100.

7. A DC motor fault diagnosis device based on the fusion of improved local mean decomposition and composite multi-scale bubble entropy, characterized in that, Including: The module for collecting the sound signal of the DC motor in the running state as the original signal; The module for decomposing the original signal by using the improved local mean decomposition method to obtain a number of product components and screening out the effective feature components with high correlation with the original signal; The module for calculating the composite multi-scale bubble entropy of the effective feature components and extracting the multi-dimensional entropy feature vector characterizing the signal complexity; The module for inputting the multi-dimensional entropy feature vector into the extreme learning machine model optimized by the particle swarm optimization for classification to obtain the fault diagnosis result of the DC motor.

8. A computer storage medium for storing a computing program, characterized in that, When the computer program is read by a computer, the computer executes the method described in claim 1.

9. A computer, comprising a processor and a storage medium, characterized in that, When the processor reads the computer program stored in the storage medium, the computer executes the method described in claim 1.

10. A computer program product, as a computer program, characterized in that, When the computer program is executed, the method described in claim 1 is implemented.

Images