Ship Variable-Speed Bearing Fault Diagnosis Method Based on Multi-Feature Fusion and Improved ShuffleNetV2

Abstract

The present invention requests protection for a ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2, which includes the following steps: First, perform CEEMD noise reduction processing on the original bearing vibration signal to reduce signal redundant information; based on the noise-reduced signal, perform angular domain resampling to obtain angular domain signals; use Hilbert transform to process the angular domain signals to obtain envelope spectrum signals, and fuse the noise-reduced signal, angular domain signals and envelope spectrum signals into an R-G-B color image to reduce signal redundant information and enhance fault features. Then, add an SE attention module to ShuffleNetV2 and replace the ReLU activation function with the H-Swish activation function to improve the accuracy in object detection and avoid the phenomenon of neuron "necrosis". Finally, input the processed signal into the improved ShuffleNetV2 to achieve fault diagnosis and recognition.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent fault diagnosis and is applicable to ship variable speed bearing fault diagnosis scenarios. Specifically, it relates to a ship variable speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2. Background Art

[0002] With the advent of Industry 4.0, mechanical equipment is gradually tending towards high performance, high load and complexity. Among the diverse mechanical equipment, rolling bearings have become one of the important parts of rotating mechanical equipment, and are widely used in ship propeller shaft systems, large wind turbines, water conservancy projects, and aerospace equipment. Due to the complex working environment and variable operating conditions of rolling bearings, they are prone to failure. According to statistics, rolling bearing failures account for more than 30% of the total mechanical equipment failures. If a rolling bearing fails, it usually causes the entire mechanical equipment to fail, which is very likely to cause irreparable losses. Therefore, the state detection and fault diagnosis research of rolling bearings is of great significance to ensure the safe operation of mechanical equipment, eliminate safety hazards, improve production efficiency, and ensure the safety of life and property.

[0003] In recent years, with the continuous improvement of science and technology, the explosive growth of data, the collection of large quantities of bearing data has become increasingly convenient, the cost has gradually decreased, and it can be effectively labeled. Deep learning methods have made significant progress in bearing fault diagnosis and have received great attention in the industry. Since at least 70% of the fault sources of rotating machinery can be found through vibration feature analysis, rolling bearing vibration signals are usually used in combination with deep learning for fault diagnosis. Compared with traditional machine learning methods, deep learning has a powerful feature extraction capability. It can mine the hidden nonlinear relationship in the original signal without using complex signal processing technology, so as to input the original vibration signal to the output fault diagnosis structure and realize the "end-to-end" working mode. Using deep learning methods to process variable, complex and non-stationary vibration signals under big data has good timeliness and versatility.

[0004] At present, most of the technical research on the fault diagnosis of rolling bearings is based on the assumption of constant rotational speed. However, in actual engineering, rolling bearings do not always operate under the condition of constant rotational speed. Typical variable-speed conditions include when mechanical equipment starts and stops. At this time, the rotational speed of the bearing will increase or decrease with time. In addition, many mechanical equipment often operate under continuous variable-speed conditions. For example, the rotating propeller shaft system on a ship often changes its rotational speed according to the needs of navigation and maneuvering. The rolling bearings in the ship's propeller shaft system will operate under variable-speed conditions. Different from the constant rotational speed condition, the impact caused by faults under variable-speed conditions no longer has periodic characteristics but changes with the change of rotational speed. In addition, the vibration signal itself under variable conditions is more complex, and the change of rotational speed will make the non-stationarity of the signal more intense. At the same time, using a single feature for fault diagnosis cannot fully and comprehensively reflect the vibration signal when rolling bearings have faults. To address the problem of signal instability under variable rotational speed, the vibration signal under variable-speed conditions can be resampled in terms of angle, converting the non-stationary signal in the time domain into a stationary signal in the angle domain, thereby eliminating the influence of signal fluctuations under variable rotational speed. To address the problem that a single feature cannot fully reflect the fault signal, first, CEEMD is used to denoise the time-domain signal. Then, the time-domain signal is converted into an angle-domain signal using resampling technology. At the same time, the envelope spectrum of the angle-domain signal is obtained using the Hilbert transform. Finally, the above three signals are fused into a red-green-blue color signal to enhance the sample features and magnify the differences under different states.

[0005] Compared with other traditional deep convolutional neural networks, ShuffleNetV2 achieves a good balance between speed and accuracy under the same complexity. ShuffleNetV2 directly inputs the feature map into two branches. Among them, in order to reduce the dimensionality of the length and width of the feature map, both branches use a 3×3 depth convolution with a stride of 2, thus reducing the computational amount of the network. Then, after the outputs of the two branches, a Concat operation is performed, and the number of channels will double after addition compared to the original input, increasing the width of the network. Thus, the number of channels is increased without significantly increasing the computational amount, enhancing the network's ability to extract features. Finally, a channel shuffle operation is performed to achieve information exchange between different channels. However, the original ShuffleNetV2 also has certain problems. Its activation function is the ReLU activation function. Although it has the characteristics of simple and efficient calculation and can effectively alleviate the overfitting problem, when the input function value is 0, it will cause the negative gradient to be set to zero, and the neuron may no longer be activated, resulting in the phenomenon of neuron "necrosis". To address this problem, the H-Swish activation function can be used to replace the ReLU activation function. This function can effectively solve the phenomenon of neuron "necrosis" in ReLU, and at the same time can reduce the network depth and maintain performance. In practice, it shows excellent performance and the inference speed remains almost unchanged. At the same time, since the SE attention module has a very simple structure and is easily deployable, it has good characteristics in terms of model and computational complexity. Therefore, introducing the SE attention module into the network can effectively improve the performance of the convolutional neural network.

[0006] CN113916535A, a variable-speed bearing fault diagnosis method, system, device and medium based on time-frequency analysis and CNN. The method includes: First, obtain multiple bearing vibration signals under variable-speed working conditions; Second, divide the bearing vibration signals into multiple segmented signals and perform Hilbert transform and time-frequency analysis to obtain the segmented time-frequency diagrams of each bearing vibration signal; Then, input the segmented time-frequency diagrams into a convolutional neural network to obtain the fault categories of each bearing vibration signal; Furthermore, perform downsampling processing on each bearing vibration signal with a determined fault category, and obtain the fault feature curve through Hilbert transform and short-time Fourier transform in sequence; Finally, calculate the fault feature coefficients of different fault categories and combine the fault feature frequencies on the fault feature curve to obtain the bearing speed information.

[0007] CN113916535A, a variable-speed bearing fault diagnosis method, system, device and medium based on time-frequency analysis and CNN. Indeed, combining time-frequency single features and CNN can achieve good results. Time-domain analysis can effectively analyze the frequency and energy distribution of signals in a global sense and is applicable to most steady-state signals. When analyzing non-steady-state signals, frequency ambiguity often occurs, and it cannot truly and effectively reflect the frequency characteristics and time-varying laws, especially for problems such as the order noise of engines, motors or rotating machinery and the fault diagnosis of bearings, gears, rotors, etc. At the same time, single features can only obtain partial information from vibration signals, and their robustness and generalization ability are poor. Moreover, the sizes of traditional CNN models are very large and not suitable for engineering applications.

[0008] This patent aims at the non-stationary working conditions of rotating machinery. By angular resampling, non-stationary signals are converted into angular quasi-stationary signals. Aiming at the problem that single features cannot provide more information and expand the differences between various faults, this patent uses three feature signals and fuses the feature signals to enhance the features of the samples. First, complete ensemble empirical mode decomposition with adaptive noise (CEEMD) is used to denoise the time-domain signal. Secondly, the resampling technology is used to convert the denoised time-domain signal into an angular-domain signal. The envelope spectrum of the angular-domain signal is obtained through Hilbert transform, and the three signals are fused into the form of an RGB image. Multi-features can make full use of the complementary information of different features and improve the accuracy of fault diagnosis. In order to be better applied in the industrial field in the future, this patent adopts the lightweight neural network ShuffleNetV2. At the same time, this patent improves the lightweight neural network ShuffleNetV2. After the branch concatenation operation of ShuffleNetV2, a squeeze-and-excitation (SE) block is added to improve the recognition accuracy of the neural network, and the ReLU activation function is replaced with the HardSwish activation function to avoid necrosis. Summary of the Invention

[0009] The present invention aims to solve the above problems of the prior art. A ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2 is proposed. The technical solution of the present invention is as follows:

[0010] A ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2, which includes the following steps:

[0011] Step 1: First, collect the original vibration signal through an accelerometer installed near the rotating bearing; use the complementary ensemble empirical mode decomposition (CEEMD) denoising method for the collected original vibration time-domain signal to denoise the vibration signal;

[0012] Step 2: Then, perform angular domain resampling on the denoised time-domain signal, and design an energy compensation coefficient to compensate for the energy loss caused by angular domain resampling;

[0013] Step 3: Next, after obtaining the angular domain signal, use Hilbert transform and squared envelope spectrum SES analysis for envelope analysis to obtain the squared envelope spectrum signal;

[0014] Step 4: Finally, arrange the obtained denoised time-domain signal, angular domain signal, and envelope spectrum signal in matrix form respectively, and fuse the time-domain vibration signal features, angular domain vibration signal features, and envelope spectrum signal features into a color image form with R-G-B three channels;

[0015] Step 5: Build the original ShuffleNet V2 network and improve and optimize the original network; add an SE module after splicing the features of each branch, and use the SE module to automatically obtain the weights of each feature channel; use the HardSwish activation function to replace the ReLU activation function in the original network;

[0016] Step 6: Divide the generated signal images into training data and test data. Then, input the training set into the improved neural network for training. After training using the convolutional neural network, use the test set for verification to achieve the fault classification of the variable-speed bearing of the ship.

[0017] Furthermore, the CEEMD denoising method is used for the original vibration time-domain signal, and the specific steps of CEEMD denoising include:

[0018] (1) Multiply white noise by a noise ratio coefficient and add it to the original vibration time-domain signal. Then, use EMD to decompose the obtained new signal to obtain a series of IMF components. Then, continue to add N different white noises and use EMD for corresponding N decompositions, and calculate the overall average of the IMF components obtained after N decompositions:

[0019]

[0020] where, \(x(t)\) represents the original vibration time-domain signal, \(\omega\) i represents white noise, \(\varepsilon\) represents the noise ratio coefficient, IMF 1 represents the mean value of the \(N\)th first-order IMF component obtained by decomposing and locking \(N\) different white noises \(N\) times, and \(E\) i represents the \(i\)th order component obtained after each EMD decomposition;

[0021] (2) Calculate the residual obtained by removing the first-order IMF component from the original vibration time-domain signal:

[0022] r 1= x(t) - IMF 1

[0023] where r 1 represents the residual after removing the first - order IMF component;

[0024] (3) Add the obtained residual to new white noise and continue with EMD decomposition. Perform the iterative decomposition in step (1) until a component that meets the conditions of the second - order IMF component is obtained. Then, calculate the average of the N second - order IMF components to obtain the second - order IMF component of the original vibration time - domain signal:

[0025]

[0026] where IMF 2 represents the mean of the Nth second - order IMF component obtained by decomposing with N different white noises added N times;

[0027] (4) Repeat the above steps until the residual of the kth order is calculated;

[0028] r k = r k-1 - IMF k

[0029] (5) Calculate the average of all N kth - order IMF components obtained to get the (k + 1)th - order IMF component of the original vibration time - domain signal:

[0030]

[0031] (6) The termination condition of the above - mentioned loop iteration is to terminate when the number of extreme points in the obtained residual is less than or equal to 2. The final residual is:

[0032]

[0033] where K represents the order of the IMF components obtained after decomposition;

[0034] (7) The final source signal can be expressed after CEEMD decomposition as:

[0035]

[0036] Furthermore, the angular - domain resampling of the noise - reduced signal is carried out. The specific steps of angular - domain resampling include:

[0037] (1) Calculate and determine the relationship between the rotational angular displacement and time. Considering the rotational speed change within a short time as constant, the rotational angle of the rotational speed is calculated using a quadratic - curve equation:

[0038] θ(t)= b 0 + b 1t + b 2 t 2

[0039] Among them, b 0 , b 1 and b 2 are three undetermined coefficients, which can be calculated through three consecutive time points (t 0 , t 1 , t 2 ) and the rotation angle Δθ, and the specific expression is:

[0040]

[0041] Among them, Δφ represents the equal-angle interval between two rotational speed pulses; finally, b 0 , b 1 and b 2 The calculation formula is expressed as:

[0042]

[0043] (2) According to the relationship between the rotational angular displacement and time, calculate the corresponding moments of the equal-angle interval resampling points; after obtaining b 0 , b 1 and b 2 , substitute them into the rotation angle calculation formula in (1) to obtain the time node t corresponding to any rotation angle θ in the range of (0, 2Δφ) in the time-domain signal:

[0044]

[0045] (3) According to the calculated time points of the angular domain resampling, the amplitude of the corresponding denoised time-domain vibration signal is obtained through the interpolation algorithm, realizing the conversion from the time-domain vibration signal to the angular domain vibration signal; design an energy compensation coefficient k to compensate for the energy loss of the angular domain resampling:

[0046]

[0047] Among them, x(t i ) is the amplitude of the denoised time-domain signal, Δt is the time increment, that is, the sampling interval, N is the wavelength of the signal, x'(t j ) is the amplitude of the angular domain vibration signal, Δt' j is the corresponding time increment between two resampling points, and the final amplitude of the angular domain signal should be multiplied by the coefficient k.

[0048] Furthermore, in step 3, secondly, after obtaining the angular domain signal, use Hilbert transform and SES analysis to perform envelope analysis to obtain the squared envelope spectrum signal. The specific steps include:

[0049] (1) Calculate the analytic signal from the discrete signal using the Hilbert transform:

[0050]

[0051] where \(x(n)\) represents the discrete signal sequence, \(j\) represents the imaginary unit, and \(H[]\) represents the Hilbert transform of the signal within the brackets;

[0052] (2) The fault signals of the bearing and gear are approximately regarded as CS2 signals; envelope analysis is used to obtain their harmonic characteristics, and the SES calculation is expressed as:

[0053]

[0054] where \(\alpha\) represents the cyclic frequency, \(FS\) represents the sampling frequency, \(\widetilde{x}\) represents the analytic signal calculated from the discrete signal using the Hilbert transform, and DFT represents the discrete Fourier transform.

[0055] Furthermore, the specific steps of step 4 for fusing the time-domain vibration signal characteristics, angular-domain vibration signal characteristics, and envelope spectrum signal characteristics into a color image form with R-G-B three channels are as follows:

[0056] First, convert the three signals into single-channel grayscale images respectively. To obtain an image of size \(M\times M\), a segmented signal of length \(M\) is randomly obtained from the original signal: 2 of

[0057]

[0058] where \(L(i)\), \(i = 1,\ldots,M\) 2 , represents the value of the signal segment, and \(P(j,k)\), \(j = 1,\ldots,M\), \(k = 1,\ldots,M\) represents the pixel intensity of the image; the \(round(\cdot)\) function is the rounding function, and the pixel values of the grayscale image are normalized from 0 to 255; then, the three grayscale images are merged to create a new image; use the \(merge(mode,bands)\) function in the built-in PIL library in Python; the \(merge()\) module can merge a group of single-band images into a new multi-band image; it includes two parameters, \(mode\) for the mode of the output image, and \(bands\) is a tuple or list of images, and the mode of each channel is described by the variable \(mode\); all channels must have the same size; the relationship between the variable \(mode\) and the variable \(bands\) can be expressed as: \(len(ImageMode.getmode(mode).bands)=len(bands)\). The \(len()\) module method represents the lengths of the variable \(mode\) and the variable \(bands\).

[0059] Further, the specific steps for building the ShuffleNet V2 neural network are as follows:

[0060] First, construct the basic unit of ShuffleNet. The basic unit of ShuffleNet is improved based on a residual unit. The basic unit includes: first, a 1x1 convolution, then a 3x3 depthwise convolution DWConv, followed by a 1x1 convolution, and finally a shortcut connection that directly adds the input to the output.

[0061] Then, improve the basic unit. First, replace the dense 1x1 convolution with a 1x1 group convolution, and add a channel shuffle operation after the 1x1 convolution. The specific operation of Channel shuffle is as follows: Assume that the input layer is divided into g groups, the total number of channels is g×n, where n is the number of input channels. Then, split the channel dimension into two dimensions (g, n), transpose these two dimensions to (n, g), and finally reshape them into one dimension. Then, add another branch to the original input. In this branch, add a 3x3 global pooling layer with a stride of 2 to obtain a feature map of the same size as the output, and connect the obtained feature map with the output.

[0062] Further, the specific steps for building the SE module and using the H-Swish activation function to improve the ShuffleNet V2 network are as follows:

[0063] First, construct the SE module. The SE module consists of two parts: Squeeze and Excitation. The network learns the feature weights according to the loss. The specific steps are as follows:

[0064] (1) Convolution operation: Use a conventional convolution operation. The calculation formula is as follows:

[0065]

[0066] where v c represents the c-th convolution kernel, x S represents the s-th input covered by the current convolution kernel, and C' represents the number of convolution kernels.

[0067] (2) In the Squeeze stage, first perform global average pooling to generate global spatial information, and then compress the global spatial information into a channel descriptor. Specifically, the H×W spatial dimension of the entire image is compressed to where the c-th element of z is calculated as follows;

[0068]

[0069] F sq (uc ) represents the global spatial information compression operation, where H and W represent the height and width of the feature image respectively, and u c represents the feature output after the convolution operation.

[0070] (3) To utilize the information aggregated in the Squeeze operation, a simple gating mechanism allows each channel to learn the weights of specific samples through a channel-dependent self-selection gating mechanism; this mechanism learns to use global information, using the sigmoid function:

[0071] s = F ex (z, W) = σ(g(z, W)) = σ(W 2 δ(W 1 z))

[0072] where δ is the ReLU activation function. The use of the ReLU activation function limits the complexity of the model and helps with training; W 1 and W 2 represent linear layers; s is the core of this module, used to represent the weights of each channel;

[0073] (4) Use the channel scaling operation to re-weight and map the feature map:

[0074]

[0075] where u c represents different channels, and s c represents the weights of the channels;

[0076] After constructing the SE module, add the SE module to the ShuffleNetV2 network.

[0077] Furthermore, after constructing the SE module, adding the SE module to the ShuffleNetV2 network, there are three ways to apply the SE block to ShuffleNet V2:

[0078] The first method is to place the SE block in parallel with the Shuffle unit in the network on the identity connection; the second method is to place the SE block directly after the last convolutional layer in the hierarchy; the third method is to place the SE block after the branch feature concatenation; using the third method, only use the SE block after the branch feature concatenation. There are only three concatenation operations in ShuffleNetV2, and only three SE blocks need to be added;

[0079] Use H-Swish to solve the phenomenon of "neuron necrosis" in ReLU, and at the same time, it can also reduce the network depth and maintain performance.

[0080] Further, the specific steps for diagnosing the faults of the variable-speed bearings of the ship are as follows:

[0081] Process the collected bearing vibration signals to obtain an R-G-B color image that fuses the noise-reduced time-domain signal, angular-domain signal, and envelope spectrum signal, and then input the color image into the trained ShuffleNetV2 neural network for detection to achieve fault diagnosis and recognition.

[0082] The advantages and beneficial effects of the present invention are as follows:

[0083] (1) During the operation of the ship's propeller shaft system, the rolling bearings do not always operate under constant-speed conditions. Different from the constant-speed conditions, the impacts caused by faults under variable-speed conditions no longer have periodic characteristics but change with the change of speed. In addition, the vibration signals under variable conditions are more complex, and the change of speed will make the non-stationarity of the signals stronger. At the same time, using a single feature for fault diagnosis cannot completely and comprehensively reflect the vibration signals when the rolling bearings have faults. However, in the current bearing fault diagnosis, most of them only use a single time-frequency analysis to process the bearing signals, resulting in poor robustness and generalizability of the fault diagnosis. Therefore, a new signal processing method is proposed. Aiming at the problem of unstable signals under variable speeds, the vibration signals under variable-speed conditions can be angle-resampled to convert the non-stationary signals in the time domain into stationary signals in the angular domain, thereby eliminating the influence of signal fluctuations under variable speeds. Aiming at the problem that a single feature cannot fully reflect the fault signals, first use CEEMD to denoise the time-domain signals, then convert the time-domain signals into angular-domain signals using resampling technology, and at the same time obtain the envelope spectrum of the angular-domain signals using Hilbert transform. Finally, fuse the above three signals into a red-green-blue color signal to enhance the sample features and amplify the differences under different states.

[0084] (2) At present, although many deep convolutional neural networks can achieve high recognition rates, they have the defects of large computational complexity and complex models, and are not suitable for deployment on edge devices and mobile terminals. In recent years, with the development of convolutional neural networks, lightweight neural networks have emerged. However, lightweight neural networks are usually used in natural scenes and have not been widely applied in the industrial field. Therefore, aiming at the problems in the current ship propeller shaft bearing fault diagnosis method based on deep learning, such as the relatively complex model, poor adaptability when deployed on edge devices and mobile terminals with limited computing resources, and poor real-time and accurate identification of bearing faults, an improved ShuffleNetV2 neural network is proposed. Based on the ShuffleNetV2 unit, an SE attention module is introduced to improve the accuracy in object detection. At the same time, replacing the ReLU activation function with the H-Swish activation function can not only reduce the number of ShuffleNetV2 units used in each Stage module of the network structure, but also avoid the "necrosis" phenomenon of neurons when the input function value is 0 when using the ReLU activation function. Description of the Drawings

[0085] Figure 1 is the flowchart of the ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2 provided by the preferred embodiment of the present invention;

[0086] Figure 2 is the schematic diagram of the signal-image conversion method;

[0087] Figure 3 is the schematic diagram of the fault classification step structure;

[0088] Figure 4 is the schematic diagram of the improved ShuffleNetV2 network model structure. Detailed Embodiment

[0089] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0090] The technical solution of the present invention to solve the above technical problems is:

[0091] Aiming at the problem of difficult fault diagnosis of the rolling bearing in the ship propeller shaft under variable-speed working conditions, a ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2 is proposed. The technical solution of the present invention is as follows:

[0092] A ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2, the process framework of this method is as Figure 1As shown in the figure, it includes the following steps:

[0093] Under different operating conditions, vibration signals are collected by an accelerometer installed near the rotating bearing; Signal processing steps: First, the CEEMD noise reduction method is used for the collected original vibration time-domain signal to reduce the noise of the vibration signal and reduce the influence of redundant signals on subsequent processing; Then, the time-domain signal after noise reduction is subjected to angular domain resampling processing. To avoid the influence of the vibration signal energy on the diagnostic result in angular domain resampling, an energy compensation coefficient is designed to compensate for the energy loss that may be generated by angular domain resampling; Secondly, after obtaining the angular domain signal, Hilbert transform and SES analysis are used for envelope analysis to obtain the squared envelope spectrum signal; Finally, the obtained noise-reduced time-domain signal, angular domain signal, and envelope spectrum signal are respectively arranged in matrix form, and the time-domain vibration signal characteristics, angular domain vibration signal characteristics, and envelope spectrum signal characteristics are fused into a color image form with R-G-B three channels;

[0094] Neural network optimization steps: Build the original ShuffleNet V2 network and improve and optimize the original network. An SE module is added after the splicing of each branch feature. The SE module can automatically obtain the weights of each feature channel, and then use these weights to enhance useful features and suppress features that are not important for the current task. The ReLU activation function in the original network is replaced with the HardSwish activation function to avoid the phenomenon of neural necrosis when the gradient is less than zero.

[0095] Furthermore, for the original vibration time-domain signal, the CEEMD noise reduction method is used, and the specific steps of CEEMD noise reduction include:

[0096] (1) Multiply white noise by a noise ratio coefficient and add it to the original vibration time-domain signal, then use EMD to decompose the obtained new signal to obtain a series of IMF components. Then continue to add N different white noises and use EMD for corresponding N decompositions, and find the overall average of the IMF components obtained after N decompositions:

[0097]

[0098] Among them, x(t) represents the original vibration time-domain signal, ω i represents white noise, ε represents the noise ratio coefficient, IMF 1 represents the mean value of the Nth first-order IMF component obtained by decomposing and locking N different white noises added N times, and E i represents the ith-order component obtained after each EMD decomposition.

[0099] (2) Calculate the residual amount remaining after removing the first-order IMF component from the original vibration time-domain signal:

[0100] r 1 = x(t) - IMF 1

[0101] where r 1 represents the residual after removing the first-order IMF component.

[0102] (3) Add the obtained residual to new white noise and continue with EMD decomposition. Perform the iterative decomposition in (1) until a component that satisfies the second-order IMF component condition is obtained. Then, calculate the average of the N second-order IMF components to obtain the second-order IMF component of the original vibration time-domain signal:

[0103]

[0104] where IMF 2 represents the mean of the Nth second-order IMF component obtained by decomposing and locking with N different white noises added N times.

[0105] (4) Repeat the above steps until the residual of the kth order is calculated:

[0106] r k = r k-1 - IMF k

[0107] (5) Calculate the average of all N kth-order IMF components obtained to obtain the (k + 1)th-order IMF component of the original vibration time-domain signal:

[0108]

[0109] (6) The termination condition for the above loop iteration is to terminate when the number of extreme points in the obtained residual is less than or equal to 2. The final residual is:

[0110]

[0111] where K represents the order of the IMF component obtained after decomposition.

[0112] (7) The final source signal can be expressed after CEEMD decomposition as:

[0113]

[0114] Furthermore, the angular domain resampling of the noise-reduced signal is performed. The specific steps of the angular domain resampling include:

[0115] (1) Calculate and determine the relationship between the rotational angular displacement and time. Considering the rotational speed change within a short period as constant, the rotational angle of the rotational speed can be calculated using a quadratic curve equation:

[0116] θ(t) = b0 +b 1 t + b 2 t 2

[0117] Among them, b 0 , b 1 and b 2 are three undetermined coefficients, which need to be obtained from three equations. Through three consecutive time points (t 0 , t 1 , t 2 ) and the rotation angle Δθ, the calculation can be obtained, and the specific performance is:

[0118]

[0119] Among them, Δφ represents the equal-angle interval between two rotational speed pulses. Finally, b 0 , b 1 and b 2 The calculation formula is expressed as:

[0120]

[0121] (2) According to the relationship between the rotational angular displacement and time, calculate the corresponding moments of the equal-angle interval resampling points; after obtaining b 0 , b 1 and b 2 , substitute them into the rotation angle calculation formula in (1), and the time node t corresponding to any rotation angle θ within the range of (0, 2Δφ) in the time-domain signal can be obtained:

[0122]

[0123] (3) According to the calculated time points of the angular domain resampling, the amplitude of the corresponding noise-reduced time-domain vibration signal can be obtained through the interpolation algorithm, and then the conversion from the time-domain vibration signal to the relatively stable angular domain vibration signal is realized. In the angular domain resampling, the energy contained in the vibration signal may affect the diagnostic result. Therefore, an energy compensation coefficient k is designed to compensate for the energy loss of the angular domain resampling:

[0124]

[0125] Among them, x(t i ) is the amplitude of the noise-reduced time-domain signal, Δt is the time increment, that is, the sampling interval, N is the wavelength of the signal, x'(t j ) is the amplitude of the angular domain vibration signal, and Δt' j is the corresponding time increment between two resampling points. The final amplitude of the angular domain signal should be multiplied by the coefficient k.

[0126] Further, the specific steps for performing envelope analysis using Hilbert transform and SES analysis to obtain the squared envelope spectrum signal are as follows:

[0127] (1) Calculate the analytic signal from the discrete signal using Hilbert transform:

[0128]

[0129] where x(n) represents the discrete signal sequence, j represents the imaginary unit, and H[] represents performing Hilbert transform on the signal within the brackets;

[0130] (2) The fault signals of bearings and gears can be approximately regarded as CS2 signals; generally, envelope analysis is used to obtain their harmonic characteristics. SES is a typical envelope analysis method and is often used to calculate CS2; the SES calculation is expressed as:

[0131]

[0132] where α represents the cyclic frequency, FS represents the sampling frequency, x~ represents the analytic signal calculated from the discrete signal using Hilbert transform, and DFT represents the discrete Fourier transform.

[0133] Further, the specific steps for fusing the time-domain vibration signal characteristics, angular-domain vibration signal characteristics, and envelope spectrum signal characteristics into a color image form with R-G-B three channels are as follows:

[0134] First, convert the three signals into single-channel grayscale images respectively. To obtain an image of size M×M, a segmented signal of length M is randomly obtained from the original signal: 2 segments:

[0135]

[0136] where L(i), i = 1, …, M 2, representing the value of the signal segment. P(j,k), where j = 1, …, M and k = 1, …, M represent the pixel intensities of the image; the round(·) function is a rounding function, and the pixel values of the grayscale image are normalized from 0 to 255; then, three grayscale pictures are merged to create a new image; here, the merge(mode, bands) function of the built-in PIL library in Python is used; the merge() module can merge a set of single-band images into a new multi-band image; it includes two parameters, mode for the mode of the output image, and bands as a tuple or list of images, and the mode of each channel is described by the variable mode; all channels must have the same size; the relationship between the variable mode and the variable bands can be expressed as: len(ImageMode.getmode(mode).bands) = len(bands).

[0137] Furthermore, the specific steps for building the ShuffleNetV2 neural network include:

[0138] First, construct the basic unit of ShuffleNetV2. The basic unit of ShuffleNetV2 is improved based on a residual unit; the basic unit mainly includes: first, a 1x1 convolution, then a 3x3 depthwise convolution (DWConv, mainly to reduce the computational amount), followed by a 1x1 convolution, and finally a shortcut connection that directly adds the input to the output;

[0139] Then, improve the basic unit. First, replace the dense 1x1 convolution with a 1x1 group convolution, and add a channel shuffle operation after the 1x1 convolution; the specific operation of Channel shuffle is as follows: assume that the input layer is divided into g groups, the total number of channels is g×n, where n is the number of input channels, then split the channel dimension into two dimensions (g, n), transpose these two dimensions to (n, g), and finally reshape them into one dimension; then, add another branch to the original input, and add a 3x3 global pooling layer with a stride of 2 in the branch to obtain a feature map of the same size as the output, and concatenate the obtained feature map with the output.

[0140] Furthermore, build the SE module. The specific steps for further improving the ShuffleNetV2 network are as follows:

[0141] First, construct the SE module, which mainly consists of two parts: Squeeze and Excitation. The model is trained by the network to learn feature weights according to the loss, so that the effective feature maps have larger weights and the ineffective or less effective feature maps have smaller weights to achieve better results. The specific steps are as follows:

[0142] (1) Convolution operation. The convolution here is a conventional convolution operation, and the calculation formula is as follows:

[0143]

[0144] Among them, v c represents the c-th convolution kernel, x S represents the s-th input under the coverage of the current convolution kernel, and C' represents the number of convolution kernels;

[0145] (2) In the Squeeze stage, first perform global average pooling to generate global spatial information, and then compress the global spatial information into the channel descriptor. Specifically, the H×W spatial dimension of the entire image is compressed to The c-th element of z is calculated as follows;

[0146]

[0147] (3) To utilize the information aggregated in the Squeeze operation, a simple gating mechanism allows each channel to learn the weights of specific samples through a channel-dependent self-selection gating mechanism; this mechanism learns to use global information, selectively emphasizes information-rich features, and suppresses less useful features, usually using the sigmoid function:

[0148] s = F ex (z, W) = σ(g(z, W)) = σ(W 2 δ(W 1 z))

[0149] Among them, δ is the ReLU activation function. The use of the ReLU activation function limits the complexity of the model and helps with training; W 1 and W 2 represent linear layers; s is the core of this module, used to represent the weights of each channel;

[0150] (4) Use channel scaling operation to re-weight and map the feature map:

[0151]

[0152] Among them, u c represents different channels, sc Represents the weight of the channel;

[0153] After constructing the SE module, adding the SE module to the ShuffleNetV2 network, there are three ways to apply the SE block to ShuffleNetV2: The first way is to place the SE block in parallel with the Shuffle unit in the network on the identity connection; The second way is to place the SE block directly after the last convolutional layer in the hierarchy; The third way is to place the SE block after the branch feature concatenation; If the first and second methods are used, the SE block needs to be added to all 16 modules of ShuffleNetV2, which will undoubtedly lead to network redundancy and higher usage costs; Therefore, using the third method to use the SE block only after the branch feature concatenation, there are only three concatenation operations in ShuffleNetV2, and only three SE blocks need to be added, which can greatly reduce the cost.

[0154] However, since the ReLU activation function causes the gradient to be 0 when the input gradient is less than 0, this results in negative gradients being set to zero in this ReLU, and this neuron may never be activated by any data again, resulting in "necrosis"; Using H-Swish can effectively solve the phenomenon of neuron "necrosis" in ReLU, and at the same time can also reduce the network depth and maintain performance.

[0155] Furthermore, the specific steps for the proposed method to perform ship variable-speed bearing fault diagnosis are as follows:

[0156] Process the collected bearing vibration signal to obtain an R-G-B color image that fuses the denoised time-domain signal, angular-domain signal, and envelope spectrum signal, and then input the color image into the trained ShuffleNetV2 neural network for detection to achieve fault diagnosis and recognition.

[0157] The systems, devices, modules, or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0158] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements not only includes those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising said element.

[0159] The above embodiments should be understood as being only for illustrative purposes of the present invention and not for limiting the scope of protection of the present invention. After reading the content described in the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A fault diagnosis method for ship variable-speed bearings based on multi-feature fusion and improved ShuffleNetV2, characterized in that, it includes the following steps: Step 1: First, collect the original vibration signal through an accelerometer installed near the rotating bearing; use the complementary ensemble empirical mode decomposition (CEEMD) noise reduction method for the collected original vibration time-domain signal to perform noise reduction processing on the vibration signal; Step 2: Then, perform angular domain resampling on the noise-reduced time-domain signal, and design an energy compensation coefficient to compensate for the energy loss generated by angular domain resampling; Step 3: Secondly, after obtaining the angular domain signal, use Hilbert transform and squared envelope spectrum (SES) analysis for envelope analysis to obtain the squared envelope spectrum signal; Step 4: Finally, arrange the obtained noise-reduced time-domain signal, angular domain signal, and envelope spectrum signal into matrix forms respectively, and fuse the time-domain vibration signal features, angular domain vibration signal features, and envelope spectrum signal features into a color image form with R-G-B three channels; Step 5: Build the original ShuffleNetV2 network and improve and optimize the original network; add an SE module after splicing each branch feature, and use the SE module to automatically obtain the weights of each feature channel; Replace the ReLU activation function in the original network with the HardSwish activation function; Step 6: Divide the generated signal images into training data and test data; then, input the training set into the neural network for training; after training using the convolutional neural network, use the test set for verification to achieve fault classification of ship variable-speed bearings.

2. The fault diagnosis method for ship variable-speed bearings based on multi-feature fusion and improved ShuffleNetV2 according to claim 1, characterized in that, when using the CEEMD noise reduction method for the original vibration time-domain signal, the specific steps of CEEMD noise reduction include: (1) Multiply white noise by a noise ratio coefficient and add it to the original vibration time-domain signal, then use EMD to decompose the obtained new signal to obtain a series of IMF components. Then continue to add N different white noises and use EMD for corresponding N times of decomposition, and find the overall average of the IMF components obtained after N times of decomposition: where x(t) represents the original vibration time-domain signal, ω i represents white noise, ε represents the noise ratio coefficient, and IMF 1 represents the mean of the Nth first-order IMF component obtained by adding N different white noises and unlocking the decomposition N times, and E i represents the ith-order component obtained after each EMD decomposition; (2) Calculate the residual obtained by removing the first-order IMF component from the original vibration time-domain signal: r 1 = x(t) - IMF 1 where r 1 represents the residual after removing the first-order IMF component; (3) Add the obtained residual to new white noise and continue with EMD decomposition, and perform the iterative decomposition in step (1) until the components that meet the second-order IMF component conditions are obtained, and then find the average of the N second-order IMF components to obtain the second-order IMF component of the original vibration time-domain signal: Among them, IMF 2 represents the mean value of the Nth second-order IMF component obtained by unlocking the decomposition of N different white noises added N times; (4) Repeat the above steps until the k-th order residual is calculated; r k = r k-1 - IMF k (5) Find the average of all N k-th order IMF components obtained to obtain the (k + 1)-th order IMF component of the original vibration time-domain signal: (6) The termination condition of the above loop iteration is to terminate when the number of extreme points in the obtained residual is less than or equal to 2, and the final residual is: where K represents the order of the IMF components obtained after decomposition; (7) The final source signal can be expressed after CEEMD decomposition as follows:

3. The ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2 according to claim 2, characterized in that the angular domain resampling of the denoised signal, and the specific steps of angular domain resampling include: (1) Calculate and determine the relationship between the rotational angular displacement and time. Regarding the rotational speed change within a short time as constant, the rotation angle of the rotational speed is calculated using a quadratic curve equation: θ(t) = b 0 + b 1 t + b 2 t 2 Among them, b 0 , b 1 and b 2 are three undetermined coefficients, which can be calculated through three consecutive time points (t 0 , t 1 , t 2 ) and the rotation angle Δθ, and the specific manifestation is as follows: where Δφ represents the equal angular interval between two rotational speed pulses; finally, b 0 , b 1 and b 2 are calculated by the following formulas: (2) Calculate the moments corresponding to the resampling points at equal angular intervals according to the relationship between the rotational angular displacement and time; after obtaining b 0 , b 1 and b 2 , substitute them into the rotational angle calculation formula in (1) to obtain the time node t corresponding to any rotational angle θ in the time domain signal within the range of (0, 2Δφ): (3) According to the calculated time points of angular domain resampling, the amplitude of the corresponding denoised time-domain vibration signal is obtained through an interpolation algorithm, realizing the conversion from the time-domain vibration signal to the angular domain vibration signal; design an energy compensation coefficient k to compensate for the energy loss during angular domain resampling: where x(t i ) is the amplitude of the noise-reduced time-domain signal, Δt is the time increment, i.e., the sampling interval, N is the wavelength of the signal, x′(t j ) is the amplitude of the angular-domain vibration signal, Δt′ j is the corresponding time increment of two resampled points, and the final amplitude of the angular-domain signal should be multiplied by the coefficient k.

4. The ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2 according to claim 3, characterized in that in step 3, secondly, after obtaining the angular domain signal, Hilbert transform and SES analysis are used for envelope analysis to obtain the squared envelope spectrum signal, and the specific steps include: (1) Calculate the analytic signal from the discrete signal using Hilbert transform: where x(n) represents the discrete signal sequence, j represents the imaginary unit, and H[] represents performing Hilbert transform on the signal within the brackets; (2) The fault signals of the bearing and gear are approximately regarded as CS2 signals; envelope analysis is used to obtain their harmonic characteristics, and the SES calculation is expressed as: where α represents the cyclic frequency, FS represents the sampling frequency, x~ represents the analytic signal calculated from the discrete signal using Hilbert transform, and DFT represents the discrete Fourier transform.

5. The ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShufffeNetV2 according to claim 4, characterized in that the specific steps of step 4 for fusing the time-domain vibration signal features, angular domain vibration signal features, and envelope spectrum signal features into a color image form with R-G-B three channels include: First, convert the three signals into single-channel grayscale images respectively. To obtain an image of size M×M, a segmented signal of length M is randomly obtained from the original signal: 2 ​ Among them, L(i), i = 1, ..., M 2 , represents the value of the signal segment, P(j, k), j = 1, ..., M, k = 1, ..., M represents the pixel intensity of the image; the round(·) function is a rounding function, and the pixel values of the grayscale image are normalized from 0 to 255; then, three grayscale pictures are merged to create a new image; use the merge(mode, bands) function of the built-in PIL library in python; the merge() module can merge a group of single-band images into a new multi-band image; it includes two parameters, mode is used for the mode of the output image, and bands is a tuple or list of images, and the mode of each channel is described by the variable mode; all channels must have the same size; the relationship between the variable mode and the variable bands can be expressed as: len(ImageMode.getmode(mode).bands) = len(bands), where len represents the lengths of the variable mode and the variable bands.

6. The ship variable-speed bearing fault diagnosis method based on multi-feature fusion and improved ShuffleNetV2 according to claim 5, characterized in that the specific steps of building the ShuffleNet V2 neural network include: First, construct the basic unit of ShuffleNet. The basic unit of ShuffleNet is improved based on a residual unit; the basic unit includes: first, a 1x1 convolution, then a 3x3 depthwise convolution DWConv, followed by a 1x1 convolution, and finally a shortcut connection that directly adds the input to the output; Then, the basic unit is improved. First, the dense 1x1 convolution is replaced with 1x1 group convolution, and a channel shuffle operation is added after the 1x1 convolution. The specific operation of Channel shuffle is as follows: Assume that the input layer is divided into g groups, the total number of channels is g×n, where n is the number of input channels. Then, the channel dimension is split into two dimensions (g, n), these two dimensions are transposed to (n, g), and finally reshaped into one dimension. Then, another branch is added to the original input, and a 3x3 global pooling layer with a stride of 2 is added to this branch to obtain a feature map of the same size as the output, and the obtained feature map is concatenated with the output.

7. The method for diagnosing faults of a ship's variable-speed bearing based on multi-feature fusion and improved ShuffleNetV2 according to claim 6, wherein, the specific steps for building the SE module and using the H-Swish activation function to improve the ShuffleNet V2 network are as follows: First, build the SE module. The SE module includes two parts: Squeeze and Excitation. The network learns the feature weights according to the loss. The specific steps are as follows: (1) Convolution operation; Use the conventional convolution operation, and the calculation formula is as follows: Among them, v c represents the c-th convolutional kernel, and x S represents the s-th input under the coverage of the current convolutional kernel, and C′ represents the number of convolutional kernels; (2) In the Squeeze stage, first, global average pooling is performed to generate global spatial information, and then the global spatial information is compressed into the channel descriptor. Specifically, the H×W spatial dimension of the entire image is compressed to where the c-th element of z is calculated as follows; F sq (u c ) represents the global spatial information compression operation, where H and W respectively represent the height and width of the feature image, and u c represents the feature output after the convolution operation; (3) To utilize the information aggregated in the Squeeze operation, a simple gating mechanism allows each channel to learn the weight of a specific sample through a channel-dependent self-selection gating mechanism. This mechanism learns to use global information and uses the sigmoid function: s = F ex (z, W) = σ(g(z, W)) = σ(W 2 δ(W 1 z)) Among them, δ is the ReLU activation function. The use of the ReLU activation function limits the complexity of the model and helps with training; W 1 and W 2 represent linear layers; s is the core of this module and is used to represent the weights of each channel; (4) Re-weight the mapping of the feature map using the channel scaling operation: Among them, u c represents different channels, and s c represents the weight of the channel; After building the SE module, add the SE module to the ShuffleNet V2 network.

8. The method for diagnosing faults of a ship's variable-speed bearing based on multi-feature fusion and improved ShuffleNetV2 according to claim 7, wherein, after building the SE module, adding the SE module to the ShuffleNetV2 network, there are three methods to apply the SE block to ShuffleNetV2: The first method is to place the SE block in parallel with the Shuffle unit in the network connected by identity. The second method is to directly place the SE block after the last convolutional layer of the layer. The third method is to place the SE block after the branch feature concatenation. Using the third method, only use the SE block after the branch feature concatenation. There are only three concatenation operations in ShuffleNetV2, and only three SE blocks need to be added; Use H-Swish to solve the "necrosis" phenomenon of neurons in ReLU, and at the same time, it can also reduce the network depth and maintain the performance.

9. The method for diagnosing faults of a ship's variable-speed bearing based on multi-feature fusion and improved ShuffleNetV2 according to claim 8, wherein, the specific steps for diagnosing faults of the ship's variable-speed bearing are as follows: Process the collected bearing vibration signals to obtain an R-G-B color image that fuses the noise-reduced time-domain signal, angular-domain signal, and envelope spectrum signal. Then, input the color image into the trained ShuffleNetV2 neural network for detection to achieve fault diagnosis and recognition.

Images