Fault diagnosis method and system for two-dimensional imaging of sound signals of carrier roller of belt conveyor

Abstract

The invention provides a two-dimensional imaging fault diagnosis method and system for a carrier roller sound signal of a belt conveyor, and relates to the field of underground coal mine equipment.The method comprises the steps that noise reduction preprocessing is conducted on a sound vibration signal based on a wavelet threshold noise reduction model of a two-parameter three-section threshold function, and the generalization ability of the noise reduction model is improved; on the basis of an acoustic signal imaging method based on MTF image coding, time domain features of acoustic signals are fused into a signal two-dimensional imaging process through a mode of combining a Laplace algorithm and mean pixel enhancement, and the effectiveness of fault texture feature extraction is improved; a RegNet network model and a Biform attention mechanism are combined, a lightweight two-dimensional convolutional neural network model is built, and low-consumption, high-efficiency and high-precision classification of carrier roller fault types is realized.

Description

Technical Field

[0001] The invention relates to the technical field of underground coal mine equipment, and in particular to a fault diagnosis method and system for two-dimensional imaging of acoustic signals of belt conveyor rollers. Background Art

[0002] The intelligent construction of coal mines is the only way for the high-quality development of coal mines. During the transportation of coal mines, intelligent timing and quantitative inspection of belt conveyors is one of the key measures to ensure the safety of operations. Its stable and reliable operation is directly related to the efficiency of coal production safety. Traditional fault detection usually relies on manual inspections, which is time-consuming and labor-intensive, and is easily affected by human subjective factors, resulting in low detection efficiency. At the same time, traditional fault detection mainly diagnoses by analyzing the speed, torque, vibration and other signals of the rollers. If traditional vibration acceleration sensors are used to collect their vibration data, two problems will be faced: First, a large number of sensors need to be deployed, which will increase the difficulty and deployment cost of sensor maintenance; second, the amount of data generated by many sensors is huge, which poses new challenges to the computer hardware system in the process of processing these complex data.

[0003] Traditional contact vibration sensors can no longer meet the requirements of intelligent inspection. In addition, due to the serious noise interference in the underground environment, the effect of fault analysis of belt conveyor rollers by collecting acoustic signals is limited. Therefore, when conducting intelligent inspection of rollers, how to improve the noise reduction capability of data preprocessing during the inspection process and ensure the accuracy of the diagnosis results has become an important challenge at this stage. Summary of the invention

[0004] To solve the above problems, an embodiment of the present invention provides a fault diagnosis method for two-dimensional imaging of acoustic signals of belt conveyor rollers, the method comprising: obtaining acoustic vibration signals of belt conveyor rollers when they are working; performing denoising preprocessing on the acoustic vibration signals based on a wavelet threshold denoising model with a two-parameter three-segment threshold function; performing Markov image coding on the acoustic vibration signals after denoising preprocessing to obtain an encoded image, and performing image enhancement processing on the encoded image; the image enhancement processing adopts a Laplace algorithm and a mean pixel enhancement algorithm; the image after image enhancement processing is input into a pre-trained two-layer routing attention mechanism Biformer-RegNet roller fault diagnosis model to obtain a fault diagnosis result; the Biformer-RegNet roller fault diagnosis model introduces a Biformer attention mechanism in the Block module of the RegNet network, and the Biformer-RegNet roller fault diagnosis model is trained by a training set and a test set divided by the image after image enhancement processing.

[0005] The embodiment of the present invention performs denoising preprocessing on acoustic vibration signals based on a wavelet threshold denoising model with a dual-parameter three-segment threshold function, thereby improving the generalization ability of the denoising model; an acoustic signal imaging method based on MTF image coding, and on this basis, a Laplace algorithm and mean pixel enhancement are combined to integrate the time domain characteristics of the acoustic signal into the two-dimensional imaging process of the signal, thereby improving the effectiveness of fault texture feature extraction; a RegNet network model is combined with a Biformer attention mechanism to build a lightweight two-dimensional convolutional neural network model, thereby achieving low-cost, high-efficiency and high-precision classification of roller fault types.

[0006] Optionally, the performing image enhancement processing on the encoded image includes: performing second-order spatial differentiation on pixel values ​​of the encoded image according to a Laplace algorithm, or performing a convolution operation on the encoded image.

[0007] In the embodiment of the present invention, the image coding method is improved, image enhancement processing based on the Laplace algorithm is added, and the time domain characteristic parameters of the acoustic signal are integrated into the signal two-dimensional imaging process.

[0008] Optionally, the image enhancement processing performed on the encoded image includes: calculating the time domain characteristic parameters of the acoustic signal to obtain the time domain mean of the acoustic signal; using the time domain mean as a multiplication factor to expand the intensity of each channel of each pixel point to obtain the pixel value of each pixel point; normalizing the pixel value of each pixel point and then reallocating each channel value between (0, 255) to obtain a new pixel value and reconstructing the image according to the new pixel value.

[0009] In the embodiment of the present invention, image enhancement processing based on averaged pixel enhancement is added, and the time domain characteristic parameters of the acoustic signal are integrated into the signal two-dimensional imaging process.

[0010] Optionally, the Biformer-RegNet roller fault diagnosis model includes: an input layer CBR, a backbone layer Body and an output layer RegHead; the backbone layer Body includes a plurality of Restage modules connected in sequence, the Restage module includes at least one BotteleNeck_x submodule, and a Biformer submodule is added to the BotteleNeck_x submodule.

[0011] In the embodiment of the present invention, the RegNet network model and the Biformer attention mechanism are combined to improve the accuracy of model diagnosis and recognition.

[0012] Optionally, the BotteleNeck_x submodule includes: a 3×3 deep convolution submodule, a two-layer routing attention submodule, a normalization layer and a multi-layer perceptron.

[0013] In the embodiment of the present invention, a Biformer submodule is added to the BotteleNeck_x submodule, thereby improving the accuracy of model diagnosis and recognition.

[0014] Optionally, the output layer RegHead includes a 1×1 convolution, a GAP (Global average pooling) layer, a Dropout layer and a fully connected layer connected in sequence; the fully connected layer is used to finally output a working status classification label.

[0015] Optionally, the wavelet threshold denoising model based on a two-parameter three-segment threshold function performs denoising preprocessing on the acoustic vibration signal, including: performing multi-resolution decomposition on the acoustic vibration signal by wavelet decomposition, and performing order reduction decomposition on the signal by using a Mallat tower algorithm; selecting a threshold according to the noise level and signal characteristics, and performing zeroing or proportional reduction on the wavelet coefficients according to the threshold.

[0016] In the embodiment of the present invention, a new wavelet threshold denoising algorithm is adopted to highlight the useful components of the signal, optimize the selection of wavelet basis functions, thresholds and decomposition levels on the basis of retaining the advantages of traditional soft and hard threshold functions for denoising, and realize preprocessing of effective signal denoising under strong noise background.

[0017] An embodiment of the present invention provides a fault diagnosis system for two-dimensional imaging of acoustic signals of belt conveyor rollers, which is used to execute any of the above-mentioned fault diagnosis methods for two-dimensional imaging of acoustic signals of belt conveyor rollers.

[0018] Optionally, it includes: a login system module and a fault diagnosis module; the login system module is used to set usage permissions, and when the user name or password is entered incorrectly, the error information can be fed back to the login interface in time. After the user with usage permissions successfully logs in, he enters the fault diagnosis main interface; the fault diagnosis module is configured with a trained Biformer-RegNet network model, and when an image is imported into the fault diagnosis module, it can recognize and classify the image, and display the recognition result at a specified position on the interface.

[0019] The fault diagnosis system for two-dimensional imaging of belt conveyor roller sound signals provided by the embodiment of the present invention can achieve the same technical effect as the fault diagnosis method for two-dimensional imaging of belt conveyor roller sound signals described above. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0021] Figure 1 A fault diagnosis method framework for two-dimensional imaging of acoustic signals of belt conveyor rollers provided in an embodiment of the present invention;

[0022] Figure 2 It is a schematic flow chart of a fault diagnosis method for two-dimensional imaging of acoustic signals of belt conveyor rollers in an embodiment of the present invention;

[0023] Figure 3 Schematic diagram of MTF image coding of vibration signal in an embodiment of the present invention;

[0024] Figure 4 Schematic diagram of MTF image coding of acoustic signals in an embodiment of the present invention;

[0025] Figure 5 Schematic diagram of the process of the average pixel enhancement algorithm in the embodiment of the present invention;

[0026] Figure 6 The Biformer-RegNet model structure in the embodiment of the present invention;

[0027] Figure 7 It is a spatial structure diagram of the RegNet model in an embodiment of the present invention;

[0028] Figure 8 This is a schematic diagram of the structure of Block in the RegNet model in an embodiment of the present invention;

[0029] Fig. 9 Schematic diagram of the classification process of the RegNet model in an embodiment of the present invention;

[0030] Fig.10 This is a structural diagram of the Biformer network model in an embodiment of the present invention;

[0031] Fig.11 This is a structural diagram of the Biformer attention mechanism in an embodiment of the present invention;

[0032] Fig.12 1 is a functional structure diagram of a roller fault diagnosis system in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0034] In order to solve the above technical problems, the present invention provides a fault diagnosis method for belt conveyor roller sound signal two-dimensional imaging, which has the following characteristics:

[0035] (1) Aiming at the nonlinear and non-stationary characteristics of acoustic signals when the roller is working, a new wavelet threshold denoising algorithm is adopted to highlight the useful components of the signal. On the basis of retaining the advantages of traditional soft and hard threshold function denoising, the selection of wavelet basis function, threshold and decomposition layer is optimized, and effective signal denoising preprocessing is realized under strong noise background.

[0036] Specifically, according to the existing environment in coal mines, a wavelet threshold denoising model based on a dual-parameter three-stage threshold function is used to achieve denoising preprocessing of acoustic signals under complex environmental conditions.

[0037] (2) The image coding method of one-dimensional signals is introduced into the feature extraction of acoustic signals and vibration signals, and the Laplace algorithm and time domain feature fusion are used to perform image enhancement processing. From the perspective of the image domain, fault texture features are extracted in the Convolutional Neural Networks (CNN) model, and the images are used as training sets and test sets of the CNN model to verify the effectiveness and reliability of this image processing method.

[0038] Optionally, a two-dimensional image coding method based on Markov Transition Field (MTF) is used to achieve feature dimensionality upgrade of acoustic signals; on this basis, the Laplace algorithm and time domain feature fusion method are used to perform image enhancement processing, and the time domain features of the acoustic vibration signal are integrated into the signal two-dimensional image to improve the effectiveness of fault texture feature extraction. In this embodiment, the standard MTF image coding method is not used, but improved.

[0039] (3) Build a lightweight two-dimensional convolutional neural network model to achieve low-cost, efficient and high-precision classification of roller fault types.

[0040] Optionally, the RegNet network model and the Biformer attention mechanism are combined, and the acoustic signal encoding images of the roller under four working conditions are used as data sets to build and train a RegNet two-dimensional convolutional neural network model based on the Biformer attention mechanism, and then build a roller fault diagnosis system based on the trained model.

[0041] Figure 1 The present invention provides a two-dimensional imaging fault diagnosis method framework for belt conveyor roller acoustic signals. Figure 1 As shown, the method includes three parts: a denoising method based on a dual-parameter three-segment wavelet threshold function, a signal feature conversion method based on MTF image coding, and a roller fault diagnosis method based on the Biformer-RegNet model.

[0042] Among them, the denoising method based on the dual-parameter three-segment wavelet threshold function includes: analyzing the influencing factors of wavelet threshold denoising, theoretical verification of the dual-parameter three-segment wavelet threshold function and denoising simulation experiments, and finally verifying the denoising ability of the BT-WTD algorithm.

[0043] The signal feature conversion method based on MTF imaging coding includes: time-frequency domain feature extraction, acoustic signal imaging based on Markov transfer field, and image enhancement processing integrating time domain features. Finally, a training set is constructed to train the Biformer-RegNet model.

[0044] The roller fault diagnosis method based on the Biformer-RegNet model includes: roller fault diagnosis result analysis and model optimization based on the Biformer attention mechanism, and finally the design of the roller fault diagnosis system based on the RegNet model. The roller fault diagnosis system based on the RegNet model realizes lightweight fault diagnosis of belt conveyor rollers.

[0045] Figure 2 The present invention is a schematic flow chart of a fault diagnosis method for two-dimensional imaging of acoustic signals of belt conveyor rollers in an embodiment of the present invention. The method comprises the following steps:

[0046] S202, obtaining the acoustic vibration signal of the belt conveyor roller when it is working.

[0047] For fault diagnosis of belt conveyors, the acoustic and vibration signals of the rollers when they are working can be collected. The acoustic and vibration signals can include acoustic signals and vibration signals.

[0048] S204, performing denoising preprocessing on the acoustic vibration signal based on a wavelet threshold denoising model with a two-parameter three-segment threshold function.

[0049] Among them, the Biparameter and Trisegment of Wavelet Threshold Denoising (BT-WTD) model with a dual-parameter three-segment threshold function can flexibly process acoustic vibration signals under different noise levels by setting two thresholds and processing the wavelet decomposition signals within each threshold range in segments. The dual adjustment parameters can flexibly adjust the length of the denoising threshold interval and the asymptote of the threshold function in the non-threshold interval, thereby improving the generalization ability of the denoising model.

[0050] Specifically, the acoustic vibration signal is decomposed into multiple resolutions by wavelet decomposition, and the signal is decomposed into reduced order by Mallat's tower algorithm. Then, a threshold is selected according to the noise level and signal characteristics, and the wavelet coefficients are set to zero or scaled down according to the threshold.

[0051] Wavelet decomposition uses Mallat's tower algorithm to reduce the order of the signal. The wavelet decomposition process is to select a wavelet basis function to perform multi-scale wavelet decomposition on the signal to be processed. The decomposition at each scale can be expressed as:

[0052]

[0053] Where N is the number of sampling points; i is the number of decomposition layers; j is the scale metric space of the filter; h j is a low-pass filter, g j is a high-pass filter; c i is the low-frequency wavelet coefficient, b i are high frequency wavelet coefficients.

[0054] In each decomposition process, many wavelet coefficients can be obtained. After n layers of decomposition, the wavelet coefficients of the useful signal are slightly larger, that is, the low-frequency approximate component (CA), which is characterized by a large amplitude and a small fluctuation frequency; the wavelet coefficients obtained after noise decomposition are smaller, that is, the high-frequency detail component (CD), which is characterized by a smaller amplitude and a large fluctuation frequency. The noise energy is also mainly concentrated in the high-frequency detail component after wavelet decomposition. With the increase in the number of decomposition layers of the approximate coefficient (low-frequency approximate component CA), the weak noise signal in the previous layer will also be continuously dispersed into the CA of the next layer, further reducing the noise level in the signal. However, when the number of layers is too high, the useful signal will be mistakenly treated as a noise signal and processed, which will affect the reconstruction of the signal. In order to take both into account, 3 to 5 layers of decomposition layers are generally selected.

[0055] Threshold function denoising is a key step in the wavelet threshold denoising process. Its principle is to select a threshold value according to the noise level and signal characteristics, and to set the wavelet coefficients of different sizes to zero or scale them down to achieve the purpose of filtering the noise signal.

[0056] In wavelet threshold denoising, the threshold function reflects different ways of processing wavelet coefficients outside the threshold range. After the acoustic signal is decomposed by wavelet, the threshold function is needed to process the detail coefficients at different decomposition levels to obtain the wavelet estimation coefficients. Therefore, the design of the threshold function is the key to the wavelet threshold denoising method.

[0057] Aiming at the shortcomings of the classic soft threshold and hard threshold functions, this paper introduces exponential function and adjustment parameters α and β into the threshold function, and divides the processing interval into three sections: proportional reduction interval, transition reduction interval and zeroing interval in the form of double thresholds. A double-parameter three-segment wavelet denoising threshold function (The Bifactorial and Three-Segment function, TBTS) with the characteristics of continuity, constant deviation, symmetry and flexibility is proposed. The new threshold function expression is as follows:

[0058]

[0059] Among them, e is a natural constant, λ and λ 0 are all thresholds and satisfy λ 0 =βλ, α and β are flexibly adjustable parameters, α∈(0,+∞), β∈(0,1).

[0060] This threshold function achieves flexible adjustment of the threshold interval and the constant deviation by introducing two adjustment parameters: when the wavelet decomposition coefficient is greater than the threshold, it can quickly approach the hard threshold function curve, reduce the constant deviation and avoid signal distortion (proportional reduction interval); when the wavelet decomposition coefficient is between the two thresholds, it can avoid the pseudo-Gibbs problem (transition reduction interval); when the wavelet decomposition coefficient is less than the threshold, the wavelet coefficient is processed with zero noise components (zeroing interval), thereby retaining more useful signal features, which not only enhances the correlation between the wavelet coefficients of the soft threshold function, but also makes up for the defect of the hard threshold function being discontinuous at the threshold point.

[0061] In the improved threshold function, two flexibly adjustable parameters are introduced, namely the slope adjustment parameter α and the threshold range adjustment parameter β. When the threshold range adjustment parameter β is 0.95, as the value range of α increases from 0 to positive infinity, the asymptote of the improved threshold function approaches y=x, which is the same as the asymptote of the hard threshold function, greatly reducing the deviation of the wavelet coefficients outside the threshold range during the wavelet denoising process.

[0062] When α=2 is fixed and the threshold range adjustment parameter β changes from 0 to 1, the zeroing interval of the threshold function changes from small to large, and the improved threshold function's processing ability for wavelet coefficients within a given threshold range changes from strong to weak. Therefore, by adjusting the range of the adjustment parameter β, it is possible to flexibly process noise signals in various frequency bands, effectively retain the useful features of the original signal in the denoising process, and minimize the noise signal.

[0063] S206, performing Markov image coding on the acoustic vibration signal after the noise reduction preprocessing to obtain a coded image, and performing image enhancement processing on the coded image.

[0064] Markov transition field MTF is a time series image coding method based on Markov chain. This method regards the change of time series as a Markov process, that is, based on the known current signal characteristics, the change of future signal time domain waveform does not depend on the past characteristics. The principle of Markov transition field image representation is to retain the original time domain information of the signal by sequentially representing the Markov transition probability of acoustic and vibration signals in the time domain through the transfer matrix on the basis of the first-order Markov chain, and at the same time, the transition probability of the amplitude quantile interval is used to realize the analysis of the dynamic characteristics of the time series signal, and finally realize the conversion of one-dimensional time series signal into two-dimensional image representation.

[0065] When the one-dimensional acoustic signal or vibration signal time series is processed, the Markov matrix is ​​finally obtained. The matrix reflects the probability transfer relationship between each quantile interval. The expression is shown in the following formula:

[0066]

[0067] Among them, w ij (i,j∈{1,2,…,Q}), represents q j The elements in q are i The probability of an element in following is determined by q j The frequency of the neighboring data amplitude in is determined by the following formula:

[0068] w ij =P(x t ∈q i |x t-1 ∈q j )

[0069] After normalizing the elements in each row, the resulting transfer matrix W is the Markov matrix, but this way of constructing the matrix ignores the conditional relationship between the time series distribution and the time step dependency. Removing the time dependency will result in a large amount of signal information loss. In order to overcome this limitation, each calculated transition probability is rearranged according to the order of the time series, thereby expanding the original Markov matrix. Through this process, a matrix of size [N,N] can be obtained, which is called the Markov transition field matrix M. This matrix records the complete probability information of each possible state transition in the time series in more detail, and its expression is shown in the following formula:

[0070]

[0071] Among them, the quantile interval q i and q j are the data at time steps i and j respectively; M ij for q i to q j The migration probability.

[0072] The Markov transition field matrix obtained through this reconstruction method makes up for the shortcomings of the original transfer matrix through time arrangement and has time correlation. The two-dimensional image constructed by Markov transition field matrix encoding can describe the dynamic changes of time and frequency of one-dimensional signals, making the constructed image dataset have generalization performance.

[0073] The one-dimensional acoustic signal and vibration signal are respectively encoded by Markov image. Figure 3 shows a schematic diagram of MTF image encoding of vibration signals, Figure 4 A schematic diagram of acoustic signal MTF image encoding is shown. Figure 3 , Figure 4 The following are the coded images of acoustic signals and vibration signals under four working conditions (normal working condition, roller stuck, bearing abnormality and support frame abnormality).

[0074] Since the images of the vibration signal under the four different working conditions are different, but after the acoustic signal image is encoded, the images under the normal working condition and the roller stuck condition are relatively similar, and the images of the two working conditions of bearing abnormality and support frame abnormality are relatively similar, so in order to better realize the fault image recognition and classification of the roller acoustic signal, the encoded image still needs to be enhanced.

[0075] Specifically, the image enhancement process may adopt the Laplace algorithm and the average pixel enhancement algorithm.

[0076] Among them, the Laplace algorithm is used for image enhancement processing, including: performing second-order spatial differentiation on the pixel values ​​of the encoded image according to the Laplace algorithm, or performing a convolution operation on the encoded image. The two-dimensional image enhancement method can use the Laplace algorithm to enhance the image edge by performing second-order spatial differentiation on the pixel values ​​of the image, or enhance the high-frequency details in the image through image convolution operations, thereby improving the image clarity and edge information.

[0077] Among them, the image enhancement processing is performed using a mean pixel enhancement algorithm, including: calculating the time domain characteristic parameters of the acoustic signal to obtain the time domain mean of the acoustic signal; using the time domain mean as a multiplication factor to expand the intensity of each channel of each pixel point to obtain the pixel value of each pixel point; normalizing the pixel value of each pixel point and then redistributing each channel value between (0, 255) to obtain a new pixel value and reconstructing the image according to the new pixel value.

[0078] Figure 5 FIG. 2 shows a flow chart of the average pixel enhancement algorithm. Figure 5 As shown in the figure, for the MTF image, the channel value of each pixel (R channel, G channel and B channel) is first calculated, and the time domain characteristic parameters of the acoustic signal are calculated to obtain the time domain mean of the signal, and the mean is used as a multiplication factor to simultaneously expand the intensity of each channel, and then the pixel value is normalized and the channel value is redistributed between (0, 255), and a new MTF image is reconstructed after obtaining a new pixel value. If the amplified channel value is greater than 255, it is normalized, and if the amplified channel value is less than 255, it is directly output. The MTF image is reconstructed based on the output pixel value.

[0079] go through Figure 5 The processing flow can only enhance the size of the image pixel value without changing the contrast between the pixels based on the signal time domain characteristic parameters, and can still retain the time-frequency domain characteristics of the acoustic signal after image encoding to a certain extent.

[0080] S208, inputting the image after image enhancement processing into a pre-trained Biformer-RegNet roller fault diagnosis model to obtain a fault diagnosis result. The Biformer-RegNet roller fault diagnosis model is trained by a training set and a test set divided by the image after image enhancement processing.

[0081] The characteristic of the Biformer-RegNet roller fault diagnosis model is that the Biformer two-layer routing attention mechanism is introduced in the Block module of the RegNet network.

[0082] In this embodiment, RegNet and the transformer network based on double-layer routing attention are first analyzed, and then the Biformer attention mechanism is introduced into the Block module of the RegNet network. The model is trained and optimized through existing data sets, and the results are compared with other lightweight networks. Finally, a roller fault diagnosis system is built based on the optimized model.

[0083] Specifically, the Biformer-RegNet roller fault diagnosis model includes: an input layer CBR, a backbone layer Body and an output layer RegHead.

[0084] Among them, the input layer CBR refers to the combination of convolution (Conv), batch normalization (Batch Normalization, BN) and ReLU activation function (ReLU). Convolution is used to extract the features of input data, batch normalization is used to accelerate the training process and improve the generalization ability of the model, and the ReLU activation function is used to increase the nonlinearity of the model and improve the expressiveness of the model.

[0085] The backbone layer Body includes multiple Restage modules connected in sequence, and the Restage module includes at least one BotteleNeck_x submodule, and a Biformer submodule is added to the BotteleNeck_x submodule. The above-mentioned BotteleNeck_x submodule includes: a 3×3 deep convolution submodule, a two-layer routing attention submodule, a normalization layer and a multi-layer perceptron. This embodiment introduces the Biformer attention mechanism in the Block module of the RegNet network, specifically, the Biformer attention mechanism is added to the BotteleNeck_x submodule in the Restage. The Biformer submodule located after the group convolution can focus on the feature distribution of the acoustic signal image and improve the feature extraction ability of the model.

[0086] The output layer RegHead includes a 1×1 convolution, a GAP layer, a Dropout layer, and a fully connected layer connected in sequence; the fully connected layer is used to finally output the working status classification label.

[0087] The fault diagnosis method for two-dimensional imaging of belt conveyor roller acoustic signals provided by an embodiment of the present invention performs denoising preprocessing on acoustic vibration signals based on a wavelet threshold denoising model with a dual-parameter three-segment threshold function, thereby improving the generalization ability of the denoising model; an acoustic signal imaging method based on MTF image coding, and on this basis, by combining the Laplace algorithm with mean pixel enhancement, the time domain characteristics of the acoustic signal are integrated into the two-dimensional imaging process of the signal, thereby improving the effectiveness of fault texture feature extraction; a RegNet network model and a Biformer attention mechanism are combined to build a lightweight two-dimensional convolutional neural network model, thereby achieving low-cost, high-efficiency and high-precision classification of roller fault types.

[0088] The pre-processed signal is analyzed in the time domain and frequency domain using traditional feature extraction methods, and some time domain parameter indicators such as peak value, root mean square value, etc. can be used to preliminarily judge the time domain characteristics of the signal under different working conditions. The traditional feature extraction method can make a simple judgment and classification of the above four working conditions, and analyze the time domain and frequency domain results. It is known that the traditional time domain and frequency domain extraction features have limitations and are not suitable for the intelligent inspection process of the future belt conveyor inspection robot. Therefore, the embodiment of the present invention provides a one-dimensional signal imaging method based on MTF image coding, which integrates the time domain feature parameters of the acoustic signal into the two-dimensional signal imaging process, and uses the CNN model for recognition and classification tests.

[0089] For the graphical feature recognition of the acoustic signal of the roller working status, this embodiment introduces the Biformer attention mechanism on the basis of the RegNet network model and builds an improved RegNet network recognition model RegNet-Biformer. Figure 6 The Biformer-RegNet model structure in an embodiment of the present invention is shown.

[0090] In the input module CBR_x (x represents the number of CBR modules connected in series), the input image set is processed by three sub-modules: Conv2d two-dimensional convolution layer, BN batch normalization layer, and ReLU activation function layer to obtain the feature image set; in the Body layer composed of four Restage modules, the feature image set passes through four Restage stages in sequence, and the Biformer attention mechanism is added to the BotteleNeck_x sub-module in the Restage. The Biformer sub-module located after the grouped convolution can focus on the feature distribution of the acoustic signal image and improve the feature extraction ability of the model; in the RegHead layer, the feature image set processed by the Body layer passes through 1×1 convolution, global average pooling, Dropout layer and fully connected layer in sequence, and finally outputs the model classification label, that is, the roller working state type.

[0091] Figure 7The RegNet model space structure diagram in the embodiment of the present invention is shown. The RegNet network needs to use different network model structures when processing different classification tasks. The main way is to use adjustment parameters to change the model distribution, thereby achieving the purpose of adjusting the model structure. Figure 7 The parameters involved include: the number of Blocks in each RegNet Stage d i , the number of channels of the output matrix w i And the width g of each Group in the Block.

[0092] In the RegNet model structure, as shown in Table 1, the specific input, output size and structure type of each layer of the model are given, where GAP is global average pooling, Dropout is the random inactivation ratio of neuron functions during the training process of the model, and FC is the fully connected layer of the model. The main body of the spatial network is mainly composed of the input layer Steam, the backbone layer Body and the output layer RegHead.

[0093]

[0094] Table 1

[0095] The Steam layer is used to preprocess input image sets of different sizes. The size of the training set is adjusted to a 3×224×224-dimensional input set through center cropping, four-sided padding, and triple cropping. The input image undergoes a common two-dimensional convolution with a step size of 2 and a convolution kernel size of 3×3, and outputs a 32×112×112-dimensional feature map. The backbone layer Body increases the model capacity by accumulating four Regstages, which can extract the two-dimensional data features of the signal under four working conditions and enhance the model's representation ability.

[0096] Figure 8 The schematic diagram of the structure of Block in the RegNet model in the embodiment of the present invention is shown. Each Regstage is composed of multiple Block layers stacked together, where the main branches of the Block layer are 1×1 and 3×3 grouped convolutions; when the step size is equal to 1, no processing is performed on the shortcut branch, and when the step size is equal to 2, a 1×1 convolution is added to the shortcut branch for downsampling, such as Figure 8 As shown (the left picture shows a structure with a step size of 1, and the right picture shows a structure with a step size of 2). The output layer RegHead is composed of three structures: 1×1 convolution, GAP, Dropout, and FC. Its function is to classify and diagnose the extracted feature information of the roller working status.

[0097] Fig. 9A schematic diagram of the classification process of the RegNet model in an embodiment of the present invention is shown. The image input set passes through four Restage stages in sequence and finally outputs a classification label.

[0098] In the RegNet model structure, the width and depth of the Block module are given by the fitted linear function. Its expression is:

[0099] u j =w 0 +w a j, 0≤j<d

[0100] Among them, the initial width w of the model 0 >0, the slope of the fitted channel number change line w a >0, the depth of the network is d, and the u value calculated by the above formula 5.1 is the number of channels in each layer.

[0101] Considering that the number of channels in each Block module in the training model should be integers of equal size, a hyperparameter w is introduced. m and the intermediate variable s j , and calculate the intermediate variables of the j-th network layer to satisfy the conditions of the following formula.

[0102]

[0103] In order to j Expressed as an integer, for s in the above formula j Round off and express it as Substituting this into the formula, we can see that the number of channels calculated is:

[0104]

[0105] For the number of channels u j Further quantization calculations are performed so that for each Block module, their widths are consistent, and the depth of the layer (the number of Blocks) is is equal to the number of i, as follows:

[0106]

[0107] It can be seen that the parameterization of the RegNet model space is to use grid search and automatically optimize d, w 0 、w a and w m , the parameter value range is: d<64, w 0 <256, w a <256 and 0≤w m ≤3. The classification process of the input set to the RegNet network model is as follows Fig. 9 shown.

[0108] Fig.10 The structure diagram of the Biformer network model in an embodiment of the present invention is shown. The core of the Biformer network model lies in the two-stage attention calculation: the first stage is to divide the input image into multiple coarse-grained regions of a certain size, perform self-attention in the region, and then calculate the correlation between adjacent regions and obtain their relationship matrix, and then perform sparse processing on these matrices and retain the K elements with the largest value; the second stage is to further perform self-attention calculation on the fine-grained region based on the coarse-grained sparse matrix.

[0109] Its network structure is as follows Fig.10 As shown in the figure, the input image dataset goes through four feature extraction stages from left to right, and obtains feature layers of different widths and heights after multiple linear mappings and matrix calculations. It mainly consists of four modules: 3×3 deep convolution module, bi-level routing attention module (BRA), normalization layer and multi-layer perceptron, and its core module is the BRA module.

[0110] Fig.11 The structure diagram of the Biformer attention mechanism in an embodiment of the present invention is shown. BRA is essentially a dynamic sparse attention mechanism, which realizes the computational sparseness of the attention module through bi-level routing, and realizes the resource allocation of the computer and the perception of the image content. Since it was applied in the Biformer network model when it was first proposed, it is also called the Biformer attention mechanism. Its core idea is to filter irrelevant key-value pairs in the coarse-grained area of ​​the image input set, and to achieve screening by constructing and pruning the region-level directed graph and fine-grained token to token attention.

[0111] like Fig.11 The BRA module shown is the structural diagram of the Biformer attention mechanism. In the figure, S is the number of non-overlapping regions into which the feature map is divided, Q, K, and V are the linear mappings of each region after the deformation operation, k represents the number of routing regions, A is the affinity adjacency matrix between each region, and Q represents the value obtained after the attention operation of K and V. The calculation method is shown in Formula 5.11.

[0112] O=Attention(Q,K g ,V g )+LCE(V)

[0113] Among them, K g 、Vg represents the key-value tensor collected from K and V, LCE(V) is the local context enhancement term, and the function LCE(·) is parameterized using depthwise convolution with a convolution kernel size of 5.

[0114] In this embodiment, the Biformer-RegNet model can be compared with other lightweight network diagnosis and recognition accuracy through experiments. Specifically, the data sets with the same number of samples are put into the Shufflenet V2 network model, RegNetx_200MF network model, Efficientnet V2 network model, and Mobilenet V3 network model for training respectively, and the model diagnosis and recognition accuracy is obtained after 50 trainings.

[0115] It can be seen from the experimental results that the network model constructed in this embodiment has a higher classification accuracy than other lightweight convolutional networks. As can be seen from Table 2, although the Biformer-RegNet network model has a higher parameter volume than the ShufflenetV2 model, RegNetx_200MF model and Mobilenet V3 model, the accuracy of fault identification and classification is improved by 1.99%, 1.48% and 2.66% respectively; not only that, when compared with the Efficientnet V2 network model with a larger parameter volume and computational complexity, the fault diagnosis accuracy of the Biformer-RegNet model is also relatively improved by 1.15%.

[0116]

[0117] Table 2

[0118] The embodiment of the present invention further provides a fault diagnosis system for two-dimensional imaging of the acoustic signal of a belt conveyor roller, which is used to execute the above-mentioned fault diagnosis method for two-dimensional imaging of the acoustic signal of a belt conveyor roller.

[0119] Fig.12 The functional structure diagram of the roller fault diagnosis system provided by the embodiment of the present invention is shown. The roller fault diagnosis system includes two functional modules: a system login module and a fault diagnosis module. The system login module includes the following functions: user name verification, password verification, and administrator verification. The fault diagnosis module includes the following functions: image import, roller fault identification, and result display.

[0120] Optionally, the system includes: a login system module and a fault diagnosis module;

[0121] The login system module is used to set the usage permissions. When the user name or password is entered incorrectly, the error information can be fed back to the login interface in a timely manner. After the user with usage permissions successfully logs in, he / she can enter the fault diagnosis main interface;

[0122] The fault diagnosis module is equipped with a trained Biformer-RegNet network model. When an image is imported into the fault diagnosis module, it can recognize and classify the image and display the recognition result at the specified position of the interface.

[0123] The fault diagnosis system for two-dimensional imaging of belt conveyor roller sound signals provided in the above embodiment can implement each process in the embodiment of the fault diagnosis method for two-dimensional imaging of belt conveyor roller sound signals, and will not be described again here to avoid repetition.

[0124] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, each process of the above-mentioned belt conveyor roller sound signal two-dimensional imaging fault diagnosis method embodiment is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0125] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the control device through a computer, and the program can be stored in a computer-readable storage medium. When the program is executed, it may include the processes of the above-mentioned method embodiments, wherein the storage medium may be a memory, a disk, an optical disk, etc.

[0126] In this article, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0127] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0128] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault diagnosis method for belt conveyor roller acoustic signal two-dimensional imaging, characterized in that: The method comprises: Obtain the sound and vibration signals of the belt conveyor rollers when they are working; Performing denoising preprocessing on the acoustic vibration signal based on a wavelet threshold denoising model with a two-parameter three-stage threshold function; Performing Markov image coding on the acoustic vibration signal after noise reduction preprocessing to obtain an encoded image, and performing image enhancement processing on the encoded image; the image enhancement processing adopts Laplace algorithm and average pixel enhancement algorithm; The image after image enhancement processing is input into a pre-trained Biformer-RegNet roller fault diagnosis model to obtain a fault diagnosis result; the Biformer-RegNet roller fault diagnosis model introduces a Biformer attention mechanism in the Block module of the RegNet network, and the Biformer-RegNet roller fault diagnosis model is trained by a training set and a test set divided by the image after image enhancement processing.

2. The method according to claim 1, characterized in that The performing image enhancement processing on the encoded image comprises: A second-order spatial differentiation is performed on the pixel values ​​of the encoded image according to the Laplace algorithm, or a convolution operation is performed on the encoded image.

3. The method according to claim 1 or 2, characterized in that: The performing image enhancement processing on the encoded image comprises: Calculating the time domain characteristic parameters of the acoustic signal to obtain the time domain mean of the acoustic signal; The intensity of each channel of each pixel point is expanded by using the time domain mean as a multiplication factor to obtain a pixel value of each pixel point; After normalizing the pixel values ​​of each pixel point, each channel value is redistributed between (0, 255) to obtain a new pixel value and reconstruct the image according to the new pixel value.

4. The method according to claim 1, characterized in that: The Biformer-RegNet roller fault diagnosis model includes: an input layer CBR, a backbone layer Body and an output layer RegHead; The backbone layer Body includes a plurality of Restage modules connected in sequence, and the Restage module includes at least one BotteleNeck_x submodule, and a Biformer submodule is added to the BotteleNeck_x submodule.

5. The method according to claim 4, characterized in that The BotteleNeck_x submodule includes: a 3×3 deep convolution submodule, a two-layer routing attention submodule, a normalization layer and a multi-layer perceptron.

6. The method according to claim 4, characterized in that The output layer RegHead includes a 1×1 convolution, a global average pooling layer, a Dropout layer and a fully connected layer connected in sequence; the fully connected layer is used to finally output the working status classification label.

7. The method according to claim 1, characterized in that The wavelet threshold denoising model based on a two-parameter three-segment threshold function performs denoising preprocessing on the acoustic vibration signal, including: The acoustic vibration signal is decomposed by multi-resolution by wavelet decomposition, and the signal is decomposed by reducing order by using Mallat pyramid algorithm; A threshold is selected according to the noise level and signal characteristics, and the wavelet coefficients are zeroed or scaled down according to the threshold.

8. A fault diagnosis system for belt conveyor roller acoustic signals in two-dimensional images, characterized in that: A fault diagnosis method for executing two-dimensional imaging of belt conveyor roller sound signals as described in any one of claims 1-7.

9. The belt conveyor roller sound signal two-dimensional imaging fault diagnosis system according to claim 8, characterized in that: include: Login system module and fault diagnosis module; The login system module is used to set the usage authority. When the user name or password is entered incorrectly, the error information can be fed back to the login interface in time. After the user with usage authority successfully logs in, he / she can enter the fault diagnosis main interface; The fault diagnosis module is configured with a trained Biformer-RegNet network model. When an image is imported into the fault diagnosis module, it can recognize and classify the image and display the recognition result at a specified position on the interface.

Images