Sound and vibration fusion fault diagnosis method for planetary gear of rocker arm of coal mining machine

Abstract

The invention discloses a coal mining machine rocker arm planetary gear sound and vibration fusion fault diagnosis method, which comprises the following steps of: firstly, adopting an internal sound and external vibration layout, installing a sound sensor on an internal planet carrier of a rocker arm, installing a vibration sensor on an external machine shell, and synchronously acquiring sound and vibration data under five working conditions of normal, abrasion, pitting corrosion, cracks and tooth breakage; then respectively performing characteristic mode decomposition on the vibration signals, performing variational mode decomposition on the sound signals, and extracting effective characteristics; visual fusion of sound and vibration characteristics is realized based on a point symmetry technology SDP, and the SDP constructs a new path of multi-dimensional signal fusion analysis by converting a time sequence sound and vibration signal into a polar coordinate image; and finally, end-to-end fault diagnosis is realized based on the Korotkoff enhanced multi-domain focused residual network model EWKAR. According to the method, hardware optimization, signal fusion and deep learning are combined, so that the signal quality is guaranteed, the anti-interference capability is enhanced, early weak faults can be accurately identified, the risk of missed diagnosis and misdiagnosis is reduced, and the method adapts to underground complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of coal mining machine technology, and specifically to a method for diagnosing acoustic and vibration fusion faults in the planetary gears of a coal mining machine rocker arm. Background Technology

[0002] Coal mining machines are core equipment in coal mine production. The rocker arm, as a key component, operates under harsh conditions of high load, strong vibration, and high dust levels. This makes it prone to tooth wear, pitting, cracking, and tooth breakage. If these faults are not diagnosed promptly, they can lead to decreased transmission efficiency, equipment downtime, and even underground safety accidents, severely impacting coal mine production efficiency and operational safety. Existing fault diagnosis technologies for the planetary gears of coal mining machine rocker arms have the following problems:

[0003] (1) Traditional methods often use a single vibration signal for detection, which is easily affected by complex working conditions in the mine (such as coal and rock impact, mechanical coupling vibration), making it difficult to capture early fault characteristics. On the other hand, single sound signal monitoring is affected by environmental noise aliasing, resulting in low fault identification. Neither method can fully reflect the equipment status when used alone.

[0004] (2) The existing sensor installation locations lack specificity. Vibration sensors are mostly installed in non-critical areas, sound sensors are easily affected by external noise, and the signal transmission of sensors on rotating parts is easily affected by entanglement and attenuation, resulting in a low proportion of effective information in the collected signals.

[0005] (3) Existing diagnostic algorithms have two prominent defects: insufficient comprehensiveness and effectiveness of feature extraction; some algorithms rely too much on complex models, which not only puts high demands on hardware computing power and makes it difficult to adapt to the high-precision fault diagnosis needs of downhole equipment, but also greatly increases the difficulty and cost of engineering deployment.

[0006] In summary, existing technologies struggle to balance diagnostic accuracy, anti-interference capabilities, and real-time performance, necessitating a sound-vibration fusion fault diagnosis solution with a rational hardware layout, efficient signal processing, and simple diagnostic algorithms. Summary of the Invention

[0007] The purpose of this invention is to propose a method for diagnosing the acoustic-vibration fusion fault of the planetary gear in the rocker arm of a coal mining machine, so as to solve the problems existing in the prior art.

[0008] The technical solution adopted in this invention is: a method for diagnosing acoustic and vibration fusion faults in the planetary gears of a coal mining machine rocker arm, comprising the following steps:

[0009] Step S1: Install a sound sensor on the planetary carrier inside the rocker arm to collect gear meshing sound signals. The sound signals are output to a signal acquisition unit installed outside the rocker arm. Install a vibration sensor on the outer housing of the rocker arm to collect structural vibration signals. The vibration signals are output to the signal acquisition unit.

[0010] Step S2: The signal acquisition unit collects acoustic and vibration synchronization data of the planetary gear under five working conditions to form a dataset; the five working conditions are normal working condition, wear working condition, pitting working condition, crack working condition and tooth breakage working condition.

[0011] Step S3: The vibration signal is decomposed using eigenmode decomposition to obtain the intrinsic mode components of the vibration signal.

[0012] Step S4: The sound signal is decomposed by variational mode to obtain the eigenmode components of the sound signal.

[0013] Step S5: Dynamically fuse features using the point-symmetric technique SDP (Symmetrized Dot Pattern) and visualize them to generate a fused feature map; specifically:

[0014] The first four intrinsic mode components of each of the decomposed sound and vibration signals are taken. These eight intrinsic mode components are then subjected to polar coordinate transformation, and signal mapping is performed using the polar coordinate system to form an RGB fusion feature map for the adaptation network input. The core formula is:

[0015]

[0016]

[0017]

[0018] In the formula, This represents the maximum amplitude of the signal. The minimum amplitude of the signal. For signal sample points, For time parameters, The rotation angle is... As the amplification factor, This is the clockwise deflection angle relative to the baseline;

[0019] Step S6: Fault diagnosis is achieved through the Enhanced Wavelet-Kolmogorov-Arnold Residual Network (EWKAR) model, specifically as follows:

[0020] Step S61: Establish an EWKAR network model, which includes an efficient multi-scale attention mechanism module, wavelet convolutional layers, a ResNet residual network, and a KAN network module.

[0021] Step S62, Model Training: The generated fused feature maps are divided into training set, validation set and test set, and stratified sampling is used to ensure class balance; the optimizer is AdamW (Adam with Weight Decay), and the loss function is cross-entropy loss;

[0022] Step S63: After initializing the network parameters, the training set is input into the network in batches, and the prediction results are obtained through forward propagation. The parameters are updated through backpropagation. The training process executes a preset iteration cycle until the model performance converges.

[0023] As a further improvement of the present invention, in step S3, the characteristic mode decomposition takes maximizing the correlation kurtosis as the optimization objective, and extracts fault features by adaptively updating the filter coefficients; wherein, the signal period parameter needs to be determined based on autocorrelation spectrum analysis; the autocorrelation spectrum can highlight the periodic components of the signal, and the interval at which a local maximum first appears after crossing zero is the signal characteristic period, providing a basis for filter updating; the specific steps are as follows:

[0024] Step S31: Import the vibration signal data to be analyzed and set the parameters, including the number of target decomposition modes. FIR filter length parameter and the number of frequency domain segments ;

[0025] Step S32: Construct an initial filter bank based on the Hanning window function, with the number of filters being... Initialize the iteration counter i=1;

[0026] Step S33: Perform convolution operation Obtain each modal component, where ;

[0027] Step S34: Using the original input signal x and modal components and the autocorrelation spectrum corresponding to Local maxima The period corresponding to the zero crossing point This serves as the basis for updating the filter coefficients;

[0028] Step S35: Verify the current number of iterations. If the preset threshold has not been reached, return to step S33 to continue the optimization operation.

[0029] Step S36: Construct the modal correlation coefficient matrix Identify the two most correlated modes and calculate their correlation kurtosis index, remove redundant modes with low kurtosis values, and update the mode counter. -1;

[0030] Step S37: Check the current number of modes Does it meet the set value? If satisfied If yes, proceed to the next step; otherwise, return to S32 to reinitialize.

[0031] Step S38, Retained The modal components are the final decomposition result.

[0032] As a further improvement of the present invention, in step S4, the sound signal is subjected to variational mode decomposition, specifically as follows:

[0033] Step S41: Assume there exists a signal to be decomposed:

[0034]

[0035] In the formula, This represents the decomposed intrinsic mode components. Indicates the number of decompositions. for amplitude, for The phase angle, each eigenmode component has a corresponding center frequency. ;

[0036] Step S42: Establish the following constrained variational optimization model:

[0037]

[0038] In the formula, Represents modal components, Represents the center frequency of each component. Represents the original signal. It is a pulse function;

[0039] Step S43: To transform the constrained variational optimization problem in step S42 into an unconstrained optimization problem, a quadratic penalty mechanism and the Lagrange multiplier method are adopted: the quadratic penalty term suppresses the deviation between the original signal and the sum of modal components, and the Lagrange multipliers ensure that the constraints are satisfied. The final constructed Lagrange extension function has the following specific form:

[0040] In the formula, This is a secondary penalty item. For Lagrange multipliers;

[0041] Step S44: Solve using the alternating direction multiplier method, updating alternately. and This achieves iterative approximation of the optimal solution, and the specific solution formula is as follows:

[0042]

[0043] In the formula, , , They are respectively , , The Fourier transform, here This represents the number of iterations.

[0044] Step S45: After updating the parameters of all intrinsic mode components, update the center frequency parameter. Optimization and adjustments will be made:

[0045]

[0046] In the formula, This refers to noise tolerance parameters;

[0047] Step S46: Set the judgment precision The iteration process terminates when the following convergence criteria are met; otherwise, it returns to step S44 to continue optimization:

[0048] Step S47: Optimize the results obtained after iteration. By performing inverse Fourier transform processing, K eigenmode components with specific frequency band characteristics can be obtained.

[0049] As a further improvement of the present invention, in step S61, the method for constructing the efficient multi-scale attention mechanism is as follows:

[0050] Step 1, Input Processing: Divide the C×H×W feature map into G groups along the channel dimension;

[0051] Step 2, Branch Design: Features are extracted through three independent branches, two of which are 1×1 convolutional branches that capture the global context, and one 3×3 convolutional branch that models local multi-scale features;

[0052] Step 3, Processing flow of each branch: The 1×1 branch is processed by global average pooling, concatenation, convolution, sigmoid activation and reweight feature selection; the 3×3 branch is processed by convolution, pooling and softmax normalization, and then the result is double-producted with the 1×1 branch result to generate a feature map.

[0053] Step 4, Output Integration: Aggregate the two feature maps, perform Sigmoid activation and Reweight selection, and output a feature map that retains its original size (C×H×W).

[0054] As a further improvement of the present invention, in step S61, the method for constructing the wavelet convolutional layer is as follows:

[0055] Step 1: Establish the mathematical expression for wavelet transform:

[0056]

[0057] Step 2: Use Haar wavelets to perform multi-scale frequency domain decomposition and reconstruction of the feature map to extract effective features for fault diagnosis. The corresponding decomposition filter can be expressed as:

[0058] Step 3: After performing convolution operations using the Haar wavelet kernel, a downsampling operation with a stride of 2 is performed to obtain four feature sub-maps:

[0059] Step 4: Perform small-scale convolution operations on the feature maps representing different frequencies respectively:

[0060] in, For each level of convolution kernel;

[0061] Step 5: After the convolution operation is completed, the system reconstructs the feature map through inverse wavelet transform; this reconstruction process can be formally described as follows:

[0062] As a further improvement of the present invention, in step S61, the method for constructing the ResNet residual network is as follows:

[0063] ResNet34 is used as the deep feature extraction backbone network, and the residual block adopts a dual-path structure design, which includes two key components: identity connection and feature transformation.

[0064] In the identity connection path, the input feature x is propagated forward directly without processing. When the number of input and output channels is inconsistent, dimension alignment is achieved through 1×1 convolution.

[0065] The feature transformation path consists of two convolutional units, each containing a convolution operation and a ReLU activation function, used to extract a deep representation of the input features;

[0066] The outputs of the two paths are merged by adding them element-wise, and the output expression is:

[0067]

[0068] in, Represents output, This represents a nonlinear transformation after two convolutional layers. This is achieved through direct mapping of the input.

[0069] As a further improvement of the present invention, in step S61, the method for constructing the KAN network module is as follows:

[0070] Step 1: Based on the Kolmogorov-Arnold decomposition theorem, establish the multivariate continuous function.

[0071] In the formula, Input variables Obtained through univariate transformation , For nonlinear integration of the transformation results;

[0072] Step 2, the calculation process of KAN, is described as follows:

[0073] In the formula, It is an optimizable nonlinear transformation;

[0074] Step 3: The KAN model uses a function matrix instead of the weight matrix in traditional neural networks. Each element in the matrix... All are optimizable nonlinear mapping functions; the nonlinear mapping functions are responsible for implementing the neurons of the previous layer. To the next layer of neurons The feature transformation, specifically the calculation process, can be represented as follows:

[0075] As a further improvement of the present invention, in step S62, the expression for the cross-entropy loss function is:

[0076]

[0077] Compared with the prior art, the present invention has the following technical advantages:

[0078] (1) The present invention adopts a hardware layout and synchronous acquisition scheme of internal sound and external vibration. The sound sensor is placed inside the rocker arm and the vibration sensor is placed outside the rocker arm. The sensor installation position is in line with the characteristics of the signal source: the internal sound sensor can reduce environmental noise interference, and the external vibration sensor can avoid the influence of internal oil and temperature. At the same time, it is equipped with a double conductive slip ring structure to ensure stable signal transmission.

[0079] (2) Based on the acoustic-vibration joint diagnostic technology, this invention integrates multi-dimensional information of sound and vibration signals to make up for the perception limitations of a single signal and enhance anti-interference ability; combined with deep learning technology, it effectively characterizes complex fault features and can significantly improve the diagnostic accuracy and reliability of the coal mining machine rocker arm under complex working conditions.

[0080] (3) In conjunction with the EWKAR network model, this invention can realize end-to-end fault diagnosis of the rocker arm planetary gear of the coal mining machine. The fault diagnosis accuracy rate can reach more than 99%, which can accurately identify early weak faults and greatly reduce the risk of missed diagnosis and misdiagnosis. Attached Figure Description

[0081] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] Figure 1 This is an overall flowchart of the coal mining machine rocker arm planetary gear acoustic and vibration fusion fault diagnosis method of the present invention.

[0083] Figure 2 This is a diagram showing the positions of the sound sensor and vibration sensor on the rocker arm in an embodiment of the present invention.

[0084] Figure 3 This is a schematic diagram of the dual conductive slip ring structure in an embodiment of the present invention.

[0085] Figure 4 These are the SDP diagrams corresponding to five different operating conditions in the embodiments of the present invention.

[0086] Figure 5 This is the EWKAR network model in the embodiments of the present invention.

[0087] Figure 6 This is a diagram showing the confusion matrix results of the single vibration signal comparison method in the verification experiment.

[0088] Figure 7 This is a visualization result of the dimensionality reduction method for comparing single vibration signals in the verification experiment.

[0089] Figure 8 This is a confusion matrix result diagram of the multimodal data fusion comparison method in the verification experiment.

[0090] Figure 9 This is a visualization of the dimensionality reduction results of the multimodal data fusion comparison method in the verification experiment.

[0091] Explanation of reference numerals in the attached diagram: 1-Planetary carrier, 2-Output shaft, 3-Inner conductive slip ring, 31-Inner conductive slip ring stator, 32-Inner conductive slip ring mover, 4-Outer conductive slip ring, 41-Outer conductive slip ring stator, 42-Outer conductive slip ring mover, 5-Sound sensor, 6-Sun gear shaft, 7-Outer gear ring, 8-Planetary gear bearing, 9-Vibration sensor, 10-Wire, 11-End cap. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0093] Please see Figures 1-3 The method for diagnosing acoustic-vibration fusion faults of the planetary gears of the coal mining machine rocker arm according to the present invention includes the following steps:

[0094] Step S1: Install the sound sensor 5 on the planetary carrier inside the rocker arm to collect gear meshing sound signals. The sound signals are output to the signal acquisition unit located outside the rocker arm. Install the vibration sensor 9 on the outer housing of the rocker arm to collect structural vibration signals. The vibration signals are output to the signal acquisition unit. Specifically, the sound sensor 5 is attached to the planetary gear intervals of the planetary carrier 1 inside the rocker arm via a magnetic socket, allowing for close-range capture of gear meshing sound signals. The sensor probe is wrapped with a sound-permeable diaphragm and protected against oil and impact. The vibration sensor 9 is bolted to the input shaft end of the outer housing of the rocker arm to collect structural vibration signals. During installation, ensure a tight fit with the housing, and grind the mounting surface flat to reduce signal attenuation. The vibration sensor is a triaxial accelerometer. The sun gear is mounted on the sun gear shaft 6, the planetary gears are mounted on the planetary gear bearings 8, and the external gear ring 7 is fixedly connected to the rocker arm housing.

[0095] In this embodiment, a dual conductive slip ring structure is used to achieve signal transmission between the rotating and fixed components. Specifically, the signal collected by the sound sensor 5 is transmitted to a signal acquisition card outside the rocker arm. In this embodiment, the dual conductive slip ring structure includes an inner conductive slip ring 3 and an outer conductive slip ring 4. The inner conductive slip ring 3 includes an inner conductive slip ring stator 31 and an inner conductive slip ring mover 32, and the outer conductive slip ring 4 includes an outer conductive slip ring stator 41 and an outer conductive slip ring mover 42. The outer conductive slip ring 4 is mounted on the bottom of the planetary carrier 1 and on the output shaft 2. The output shaft 2 has a hollow structure. The outer conductive slip ring stator 41 is fixedly sleeved on the output shaft 2. The outer conductive slip ring mover 42 is in contact with the outer conductive slip ring stator 41. The wire 10 on the sound sensor 5 is electrically connected to the outer conductive slip ring stator 41. The sound sensor 5 and the outer conductive slip ring stator 41 rotate synchronously with the planetary carrier 1. An inner conductive slip ring 3 is installed inside the output shaft 2 and end cover 11. The inner conductive slip ring mover 32 is fixedly installed inside the output shaft 2, and the inner conductive slip ring stator 31 is installed at the lower end of the output shaft 2 via a flange. The inner conductive slip ring mover 32 contacts the inner conductive slip ring stator 31. Inside the output shaft 2, one end of a wire 10 is connected to the outer conductive slip ring stator 41, and the other end is connected to the inner conductive slip ring mover 32. The inner conductive slip ring mover 31 and the outer conductive slip ring stator 41 rotate synchronously with the output shaft 2. A wire 10 is connected to the inner conductive slip ring stator 31, and this wire 10 is connected to the signal acquisition unit. Through the above-described double conductive slip ring structure and its cooperation with the hollow output shaft, the signal from the sound sensor 5 can be transmitted to the signal acquisition unit, i.e., the acquisition card. The acquisition card used is the Altair PCIe893x. The X, Y, and Z axis signals of the vibration sensor 9 and the signal of the sound sensor 5 are respectively connected to the four channels of the acquisition card. The sampling frequency is set to 10kHz and the sampling bit depth is 16 bits to ensure synchronous acquisition of sound and vibration signals and control the timing error within 1ms.

[0096] Step S2: The signal acquisition unit collects acoustic and vibration synchronization data of the planetary gear under five operating conditions to form a dataset; the five operating conditions are normal operating condition, wear operating condition, pitting operating condition, crack operating condition, and tooth breakage operating condition. In this embodiment, 1×10⁻⁶ data are continuously collected for each operating condition. 6 A total of 10 data points were collected to obtain acoustic and vibration synchronization datasets for five different operating conditions.

[0097] Step S3: The vibration signal is decomposed using eigenmodes to obtain the intrinsic mode components of the vibration signal.

[0098] The basic principle of Feature Mode Decomposition (EMD) is as follows: EMD employs a non-recursive decomposition strategy, initializes the FIR filter bank using the Hanning window function, and utilizes correlation kurtosis as the optimization objective to achieve adaptive adjustment of the filter, thereby effectively extracting different feature modes. In the initialization phase, an improved maximum correlation kurtosis deconvolution algorithm is used to determine the optimal FIR filter parameters. To achieve Hanning window initialization, the original signal spectrum is uniformly divided into K segments, and the upper and lower cutoff frequencies of each band are clearly defined. The upper and lower cutoff frequencies of the segments are specified as follows:

[0099] In the formula, This represents the sampling frequency of the original signal; based on the MATLAB platform, a set of coverage... K uniformly distributed FIR filters in the frequency band are used, and the fault characteristics are accurately decomposed through an adaptive filter update and mode selection mechanism.

[0100] In the initial processing stage of the algorithm, the finite-length unit impulse response filter only has basic noise suppression capabilities and is insufficient to completely eliminate interference factors in the signal. Considering the characteristics of rotating machinery fault diagnosis, the signal to be analyzed mainly contains the following typical components: periodic transient impacts, non-stationary interference, harmonic oscillations, and environmental noise. Given that the relevant kurtosis index can comprehensively reflect the periodic and transient characteristics of the signal, it is established as the core optimization criterion for feature extraction. For discrete-time series... The feature pattern decomposition is described as a constrained optimization problem, the specific expression of which is:

[0101] In the formula, Representing the Decomposition structure, For length is The A finite impulse response filter, The period of the input signal, The displacement series; the decomposition mode can be expressed in matrix form:

[0102]

[0103]

[0104] The relevant kurtosis indices of each decomposition mode are expressed mathematically as follows:

[0105] In the formula, This is a weighted control matrix.

[0106] Substituting the mode decomposition relation into the equation, we can derive the expression based on the filter coefficients:

[0107]

[0108] in and Let represent the weighted autocorrelation matrix and the standard autocorrelation matrix, respectively. The optimization solution to the above equation can be transformed into solving the following generalized eigenvalue problem:

[0109]

[0110] In the formula, It is the largest eigenvalue.

[0111] During the iterative calculation process, the first... The filter coefficients are updated based on the principle of maximizing the correlation kurtosis of the output signal. It is important to emphasize that the input signal period... The accuracy of the period estimation directly affects the optimization effect of filter parameters. However, in practical engineering applications, due to uncertainties in operating conditions and speed fluctuations, accurately obtaining the fault characteristic period is quite difficult. To solve this technical problem, the IMCKD (Eigenmode Decomposition) algorithm proposes a period estimation method based on autocorrelation analysis. The core idea of ​​this method is that the autocorrelation spectrum of the signal exhibits obvious peak characteristics at the fault characteristic period. Indicates signal x The autocorrelation function, its autocorrelation function It can be represented as:

[0112]

[0113] During signal analysis, Points in The feature points at the zero-crossing point are used to determine the period parameter. Specifically, the period estimate of each filtered signal is selected as the local extremum point of the autocorrelation spectrum that appears for the first time after crossing the zero point.

[0114] Performing a full update on all initialized filters may result in multiple modes containing similar components. To eliminate mode aliasing and redundancy, the system first identifies the mode with the highest correlation coefficient, and then compares two candidate modes. and of Define the fault information content and retain modalities that are more valuable for diagnosis.

[0115]

[0116] The following are the specific steps for mode decomposition of vibration signals:

[0117] Step S31: Import the vibration signal data to be analyzed and set the parameters, including the number of target decomposition modes. FIR filter length parameter and the number of frequency domain segments ;

[0118] Step S32: Construct an initial filter bank based on the Hanning window function, with the number of filters being... Initialize the iteration counter i=1;

[0119] Step S33: Perform convolution operation Obtain each modal component, where ;

[0120] Step S34: Using the original input signal x and modal components and the autocorrelation spectrum corresponding to Local maxima The period corresponding to the zero crossing point This serves as the basis for updating the filter coefficients;

[0121] Step S35: Verify the current number of iterations. If the preset threshold has not been reached, return to step S33 to continue the optimization operation.

[0122] Step S36: Construct the modal correlation coefficient matrix Identify the two most correlated modes and calculate their correlation kurtosis index, remove redundant modes with low kurtosis values, and update the mode counter. -1;

[0123] Step S37: Check the current number of modes Does it meet the set value? If satisfied If yes, proceed to the next step; otherwise, return to S32 to reinitialize.

[0124] Step S38, Retained The modal components are the final decomposition result.

[0125] Step S4: Perform variational mode decomposition on the sound signal to obtain the intrinsic mode components of the sound signal.

[0126] Suppose there exists a signal to be decomposed:

[0127]

[0128] In the formula, This represents the IMF components (intrinsic mode components) after decomposition. Indicates the number of decompositions. for amplitude, for The phase angle, each IMF component has a corresponding center frequency. .

[0129] The following constrained variational optimization model can be established:

[0130]

[0131] In the formula, Represents modal components, Represents the center frequency of each component. Represents the original signal. It is a pulse function;

[0132] By employing a quadratic penalty mechanism and the Lagrange multiplier method, the original constrained optimization problem is transformed into an unconstrained optimization problem. The specific form of the constructed Lagrange expansion function is as follows:

[0133] In the formula, This is a secondary penalty term with a value of 2000. It is a Lagrange multiplier.

[0134] The solution is obtained by alternating direction multipliers, which involves updating the solution alternately. , Achieve iterative approximation of the optimal solution. The specific solution process includes the following key steps:

[0135]

[0136] In the formula, , , They are respectively , , Fourier transform, This represents the number of iterations.

[0137] After updating the parameters of all intrinsic modal components, the center frequency parameter needs to be updated. Optimization and adjustments will be made:

[0138]

[0139] In the formula, This is the noise tolerance parameter.

[0140] Set the judgment precision The iterative process terminates when the following convergence criteria are met:

[0141] Step S5: Feature Fusion Visualization

[0142] Please see Figure 3 The eight IMF components are mapped to a 224×224 RGB fusion feature map. The core formula is:

[0143]

[0144]

[0145]

[0146] In the above formula, The maximum amplitude of the signal, For the minimum amplitude of the signal, For signal sample points, For time parameters, For rotation angle, As the amplification factor, It represents the clockwise deflection angle relative to the baseline.

[0147] Step S6, Fault Diagnosis

[0148] A multi-dimensional feature fusion space is constructed using the Symmetrized Dot Pattern (SDP) technique, and fault diagnosis is achieved through the Enhanced Multi-domain Focused Residual Network (EWKAR) model with Kolmogorov-Arnold Representation. During the polar coordinate mapping process, the transformation effect of SDP is mainly controlled by three key parameters: the rotation angle of the symmetry plane, the sampling time interval t, and the angle scaling factor ξ. Different frequency components exhibit significant differences in the transformed image, specifically manifested in the obvious differences in the curvature of the petal structure and the scatter distribution characteristics. The specific implementation process of the EWKAR network model is as follows:

[0149] 1. Establish the EWKAR network model

[0150] Please see Figure 4 The EWKAR network model consists of four modules: an efficient multi-scale attention mechanism module, wavelet convolutional layers, a ResNet residual network, and a KAN network module.

[0151] Efficient Multi-Scale Attention Mechanism (EMA): The EMA module uses a multi-branch parallel architecture to process the input feature map (size C×H×W). First, the feature map is divided into G groups along the channel dimension, and feature extraction is performed through three independent branches: two 1×1 convolutional branches are responsible for capturing global context information, and another 3×3 convolutional branch focuses on local multi-scale feature modeling. In the 1×1 branch processing flow, global average pooling is performed along the x and y axes respectively to generate channel-encoded features. After concatenation, these features are converted into two one-dimensional vectors through a 1×1 convolutional layer. A sigmoid activation function is used for non-linear transformation, and the output is selected by the Reweight module. The 3×3 branch extracts multi-scale spatial features through a 3×3 convolutional kernel. After global average pooling encoding, the features are normalized using the Softmax function. The output of the 3×3 branch is multiplied by the normalized 1×1 branch result to generate the first spatial attention map. After encoding and activation operations are performed on the 3×3 branch again, the output is multiplied twice by the 1×1 branch features to generate the second feature map. The two feature maps are aggregated and a sigmoid activation function is applied. After feature selection by the Reweight module, the output maintains the original feature map size (C×H×W).

[0152] Wavelet convolutional layer: Wavelet transform, as a multi-resolution analysis tool, can achieve frequency domain decomposition of signals. Its mathematical expression is as follows:

[0153]

[0154] Taking the classic Haar wavelet as an example, its corresponding decomposition filter can be expressed as:

[0155] After performing convolution operations using the Haar wavelet kernel, a downsampling operation with a stride of 2 is performed to obtain four feature sub-maps:

[0156] Perform small-scale convolution operations on feature maps representing different frequencies:

[0157] in, These are the convolution kernels for each level.

[0158] After the convolution operation is completed, the system reconstructs the feature map using inverse wavelet transform (IWT). This reconstruction process can be formally described as follows:

[0159] ResNet Residual Network: Employing ResNet34 as the deep feature extraction backbone, the residual block uses a dual-path structure, containing two key components: identity connections and feature transformation. In the identity connection path, the input feature x propagates forward directly without processing. When the number of input and output channels is inconsistent, dimensionality alignment is achieved through 1×1 convolutions. The feature transformation path consists of two convolutional units, each containing a convolution operation and a ReLU activation function, used to extract a deep representation of the input features. The outputs of the two paths are fused element-wise, with the output expression as follows:

[0160]

[0161] in, Represents output, This represents a nonlinear transformation after two convolutional layers. This is achieved through direct mapping of the input.

[0162] KAN Network Module: The KAN network is based on the Kolmogorov-Arnold decomposition theorem, which establishes the multivariate continuous function... It can be represented as:

[0163] In the KAN model, input variables Obtained through univariate transformation ,and It is a nonlinear integration of these transformation results.

[0164] The calculation process of KAN can be described as follows:

[0165] in, It is an optimizable nonlinear transformation.

[0166] The KAN model uses a function matrix instead of the weight matrix in traditional neural networks, where each element of the matrix... All are optimizable nonlinear mapping functions. These functions are responsible for implementing the previous layer of neurons. To the next layer of neurons Feature transformation. The specific calculation process can be represented as follows:

[0167] 2. Model Training

[0168] 25,000 fused feature maps were divided into training, validation, and test sets in a 7:2:1 ratio, with stratified sampling to ensure class balance. The AdamW optimizer was used, with a learning rate of 0.001, weight decay of 0.0001, exponential decay scheduling (gamma = 0.8), a batch size of 64, and 20 epochs. The loss function used was cross-entropy loss.

[0169]

[0170] After initializing the network parameters, the training set is input into the network in batches, and the prediction results are obtained through forward propagation. The parameters are updated through backpropagation. The network is evaluated with a validation set every epoch. If the accuracy does not improve after 3 consecutive epochs, the network is stopped early. The final performance is verified with a test set after training.

[0171] To verify the superiority of the EWKAR network model, this invention designed a series of comparative experiments for fault diagnosis of planetary gears in coal mining machine rocker arms, evaluating the performance of different input data and models in fault detection. The experiments used single vibration signals, single sound signals, and multimodal fusion of vibration and sound signals as input data, constructing various fault diagnosis methods for comparison. For single signals, the experiments employed two traditional deep learning methods: 1DCNN and 2DCNN based on time-frequency maps, while WDCNN (Deep Convolutional Neural Network with wide kernels in the first layer) and traditional support vector machines were selected for comparison. For multimodal data, multi-scale fusion convolution (FMSCNN) and SPAF (sound and vibration signals decomposed by MVMD (Multivariate Variational Mode Decomposition), then plotted as SDP images using SDP technology and input into the CNN network for diagnosis) were used. These were compared and validated with the model proposed in this invention. The experiments were conducted under the same data preprocessing, training strategies, and evaluation metrics, with test accuracy as the main performance evaluation indicator. The accuracy of different types of test sets is shown in Table 1.

[0172] Table 1

[0173] Input data Fault diagnosis methods Test accuracy / % Vibration signal 1DCNN 92.2 sound signal 1DCNN 69.2 Vibration signal Time-frequency graph - 2DCNN 92.32 sound signal Time-frequency graph - 2DCNN 89.70 Vibration signal WDCNN 93.34 sound signal WDCNN 84.53 Vibration signal SVM 76.6 sound signal SVM 74.2 Vibration signal + sound signal FMSCNN (Multi-Scale Fusion Convolution) 89.2 Vibration signal + sound signal SPAF (MVMD-SDP-CNN) 93.8 Vibration signal + sound signal This method 99.24

[0174] The experimental results show that, in the model with a single vibration signal input, the confusion matrix results and dimensionality reduction visualization results obtained by the comparison method are as follows: Figure 6 and Figure 7 As shown, 1DCNN and time-frequency plot-2DCNN achieved accuracies of 92.2% and 92.32%, respectively. The WDCNN-based model achieved an even higher accuracy of 93.34% on vibration signals, while the traditional SVM method also achieved 76.6% accuracy on vibration signals. This demonstrates the efficient feature extraction capability of traditional methods for vibration signals.

[0175] For a single audio signal input, the confusion matrix results and dimensionality reduction visualization results of the comparison method are shown in the figure. -26 and Figure As shown in Figure 27, 1DCNN and 2DCNN based on time-frequency maps achieved accuracies of 69.2% and 89.70%, respectively, indicating that converting the sound signal into a time-frequency map allows for better utilization of deep networks for feature representation. WDCNN, however, achieved an accuracy of 85.53% on the sound signal, lower than the time-frequency map-based 2DCNN. The SVM method's accuracy dropped to 74.2% for sound signal input, demonstrating the limitations of traditional machine learning methods in terms of poor robustness and complexity.

[0176] Regarding the fusion of multimodal data, its comparison method distinguishes between confusion matrix results and dimensionality reduction visualization results, such as... Figure 8 and Figure 9 As shown, the FMSCNN and SPAF methods achieved accuracies of 92.8% and 93.8%, respectively. However, the accuracy of FMSCNN decreased compared to WDCNN and 2DCDNN algorithms for vibration signals, indicating that single multi-scale fusion is not suitable for fusing sound and vibration signals in this method. This is because a large amount of noise in the sound signal is fused into the network without processing, interfering with the classification boundary of vibration features and thus reducing diagnostic accuracy. The method proposed in this invention significantly improves accuracy to 99.24%. This demonstrates that by fusing vibration and sound signals and employing a more efficient feature fusion and discrimination mechanism, the model of this invention can more comprehensively capture and integrate multi-source fault features, thereby achieving higher accuracy in fault diagnosis tasks.

[0177] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes that can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention are all within the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing acoustic-vibration fusion faults in the planetary gears of a coal mining machine rocker arm, characterized in that, Includes the following steps: Step S1: Install a sound sensor on the planetary carrier inside the rocker arm to collect gear meshing sound signals. The sound signals are output to a signal acquisition unit installed outside the rocker arm. Install a vibration sensor on the outer housing of the rocker arm to collect structural vibration signals. The vibration signals are output to the signal acquisition unit. Step S2: The signal acquisition unit collects acoustic and vibration synchronization data of the planetary gear under five working conditions to form a dataset; the five working conditions are normal working condition, wear working condition, pitting working condition, crack working condition and tooth breakage working condition. Step S3: The vibration signal is decomposed using eigenmode decomposition to obtain the intrinsic mode components of the vibration signal. Step S4: The sound signal is decomposed by variational mode to obtain the eigenmode components of the sound signal. Step S5: Dynamically fuse features using the point-symmetric technique SDP and visualize them to generate a fused feature map; specifically: The first four intrinsic mode components of each of the decomposed sound and vibration signals are taken. These eight intrinsic mode components are then subjected to polar coordinate transformation, and signal mapping is performed using the polar coordinate system to form an RGB fusion feature map for the adaptation network input. The core formula is: ； ； ； In the formula, This represents the maximum amplitude of the signal. The minimum amplitude of the signal. For signal sample points, For time parameters, The rotation angle is... As the amplification factor, This is the clockwise deflection angle relative to the baseline; Step S6: Fault diagnosis is achieved through the EWKAR (Enhanced Multi-Domain Focusing Residual Network) model, specifically as follows: Step S61: Establish an EWKAR network model, which includes an efficient multi-scale attention mechanism module, wavelet convolutional layers, a ResNet residual network, and a KAN network module. Step S62, Model Training: The generated fused feature maps are divided into training set, validation set and test set, and stratified sampling is used to ensure class balance; the optimizer is AdamW, and the loss function is cross-entropy loss; Step S63: After initializing the network parameters, the training set is input into the network in batches, the prediction results are obtained through forward propagation, and the parameters are updated through back propagation. The training process executes a preset iteration cycle until the model performance converges.

2. The method for diagnosing acoustic-vibration fusion faults of the planetary gears in the rocker arm of a coal mining machine according to claim 1, characterized in that, In step S3, the eigenmode decomposition aims to maximize the correlation kurtosis and extracts fault features by adaptively updating the filter coefficients. The signal periodicity parameter needs to be determined based on autocorrelation spectrum analysis. The autocorrelation spectrum highlights the periodic components of the signal; the interval between the first local maxima after the zero-crossing point is the signal's characteristic period, providing a basis for filter updates. The specific steps of eigenmode decomposition are as follows: Step S31: Import the vibration signal data to be analyzed and set the parameters, including the number of target decomposition modes. FIR filter length parameter and the number of frequency domain segments ; Step S32: Construct an initial filter bank based on the Hanning window function, with the number of filters being... Initialize the iteration counter i=1; Step S33: Perform convolution operation Obtain each modal component, where ; Step S34: Using the original input signal x and modal components and the autocorrelation spectrum corresponding to Local maxima The period corresponding to the zero crossing point This serves as the basis for updating the filter coefficients; Step S35: Verify the current number of iterations. If the preset threshold has not been reached, return to step S33 to continue the optimization operation. Step S36: Construct the modal correlation coefficient matrix Identify the two most correlated modes and calculate their correlation kurtosis index, remove redundant modes with low kurtosis values, and update the mode counter. -1; Step S37: Check the current number of modes Does it meet the set value? If satisfied If yes, proceed to the next step; otherwise, return to S32 to reinitialize. Step S38, Retain The modal components are the final decomposition result.

3. The method for diagnosing acoustic-vibration fusion faults of the planetary gears in the rocker arm of a coal mining machine according to claim 1, characterized in that, In step S4, the sound signal is subjected to variational mode decomposition, specifically: Step S41: Assume there exists a signal to be decomposed: ； In the formula, This represents the decomposed intrinsic mode components. Indicates the number of decompositions. for amplitude, for The phase angle, each eigenmode component has a corresponding center frequency. ; Step S42: Establish the following constrained variational optimization model: ； In the formula, Represents modal components, Represents the center frequency of each component. Represents the original signal. It is a pulse function; Step S43: To transform the constrained variational optimization problem in step S42 into an unconstrained optimization problem, a quadratic penalty mechanism and the Lagrange multiplier method are adopted: the quadratic penalty term suppresses the deviation between the original signal and the sum of modal components, and the Lagrange multipliers ensure that the constraints are satisfied. The final constructed Lagrange extension function has the following specific form: ; In the formula, This is a secondary penalty item. For Lagrange multipliers; Step S44: Solve using the alternating direction multiplier method, updating alternately. and This achieves iterative approximation of the optimal solution, and the specific solution formula is as follows: ； In the formula, , , They are respectively , , The Fourier transform, here This represents the number of iterations. Step S45: After updating the parameters of all intrinsic mode components, update the center frequency parameter. Optimization and adjustments will be made: ； In the formula, This refers to noise tolerance parameters; Step S46: Set the judgment precision The iteration process terminates when the following convergence criteria are met; otherwise, it returns to step S44 to continue optimization: ; Step S47: Optimize the results obtained after iteration. By performing inverse Fourier transform processing, K eigenmode components with specific frequency band characteristics can be obtained.

4. The method for diagnosing acoustic-vibration fusion faults of the planetary gears in the rocker arm of a coal mining machine according to claim 1, characterized in that, In step S61, the method for constructing the efficient multi-scale attention mechanism is as follows: Step 1, Input Processing: Divide the C×H×W feature map into G groups along the channel dimension; Step 2, Branch Design: Features are extracted through three independent branches, two of which are 1×1 convolutional branches that capture the global context, and one 3×3 convolutional branch that models local multi-scale features; Step 3, Processing flow of each branch: The 1×1 branch is processed by global average pooling, concatenation, convolution, sigmoid activation and reweight feature selection; the 3×3 branch is processed by convolution, pooling and softmax normalization, and then the result is double-producted with the 1×1 branch result to generate a feature map. Step 4, Output Integration: Aggregate the two feature maps, perform Sigmoid activation and Reweight selection, and output a feature map that retains its original size (C×H×W).

5. The method for diagnosing acoustic-vibration fusion faults of the planetary gears in the rocker arm of a coal mining machine according to claim 1, characterized in that, In step S61, the wavelet convolutional layer is constructed as follows: Step 1: Establish the mathematical expression for wavelet transform: ； Step 2: Use Haar wavelets to perform multi-scale frequency domain decomposition and reconstruction of the feature map to extract effective features for fault diagnosis. The corresponding decomposition filter can be expressed as: ； Step 3: After performing convolution operations using the Haar wavelet kernel, a downsampling operation with a stride of 2 is performed to obtain four feature sub-maps: ; Step 4: Perform small-scale convolution operations on the feature maps representing different frequencies respectively: ； in, For each level of convolution kernel; Step 5: After the convolution operation is completed, the system reconstructs the feature map through inverse wavelet transform; this reconstruction process can be formally described as follows:

6. The method for diagnosing acoustic-vibration fusion faults of the planetary gears in the rocker arm of a coal mining machine according to claim 1, characterized in that, In step S61, the method for constructing the ResNet residual network is as follows: ResNet34 is used as the backbone network for deep feature extraction. The residual block adopts a dual-path structure design, which includes two key components: identity connection and feature transformation. In the identity connection path, the input feature x is propagated forward directly without processing. When the number of input and output channels is inconsistent, dimension alignment is achieved through 1×1 convolution. The feature transformation path consists of two convolutional units, each containing a convolution operation and a ReLU activation function, used to extract a deep representation of the input features; The outputs of the two paths are merged by adding them element-wise, and the output expression is: ； in, Represents output, This represents a nonlinear transformation after two convolutional layers. This is achieved through direct mapping of the input.

7. The method for diagnosing acoustic-vibration fusion faults of the planetary gears in the rocker arm of a coal mining machine according to claim 1, characterized in that, In step S61, the method for constructing the KAN network module is as follows: Step 1: Based on the Kolmogorov-Arnold decomposition theorem, establish the multivariate continuous function. ； In the formula, Input variables Obtained through univariate transformation , For nonlinear integration of the transformation results; Step 2, the calculation process of KAN, is described as follows: ； In the formula, It is an optimizable nonlinear transformation; Step 3: The KAN model uses a function matrix instead of the weight matrix in traditional neural networks. Each element in the matrix... All are optimizable nonlinear mapping functions; the nonlinear mapping functions are responsible for implementing the neurons of the previous layer. To the next layer of neurons The feature transformation, specifically the calculation process, can be represented as follows:

8. The method for diagnosing acoustic-vibration fusion faults of the planetary gears in the rocker arm of a coal mining machine according to claim 1, characterized in that, In step S62, the expression for the cross-entropy loss function is:

Images