Design method, device and equipment of cnn model for bearing fault diagnosis

Abstract

The application relates to a design method, device and equipment of a CNN model for bearing fault diagnosis. The method firstly designs the length of a CNN model input signal based on a fault characteristic frequency and resolution requirement; secondly, determines the width and size of the input signal based on the rotating frequency of a rotating shaft; then, calculates the width of a convolution kernel based on the exponential decay characteristic of a fault peak; and finally, determines the channel number of the CNN model and the size parameter of each channel based on the number of fault categories. The CNN model design method can directly design the CNN model based on fault characteristics, significantly improves the design efficiency of the CNN model, and improves the interpretability of the network structure and parameters. In addition, when the CNN model designed by the method is used for bearing fault diagnosis, the CNN model has higher precision, smaller uncertainty and faster convergence speed.

Description

Technical Field

[0001] This application relates to the field of deep convolutional neural network (CNN) model design technology, and in particular to a design method, apparatus and device for a CNN model used for bearing fault diagnosis. Background Technology

[0002] As a representative of deep learning algorithms, deep convolutional neural networks (CNNs) have been widely used in fault diagnosis and early warning of rotating equipment.

[0003] Publication CN117171544A discloses a multi-channel fusion convolutional neural network and its application to fault diagnosis of motor vibration. Publication CN117708574A discloses a physically embedded CNN by embedding the envelope order spectrum of rolling bearing acceleration into a convolutional neural network and applying it to fault diagnosis of bearings with variable speeds. Publication CN117932262A discloses a fault diagnosis method for ZPW-2000A track circuits based on CNN-LSTM-Attention. This method uses a CNN layer for local feature extraction, an LSTM encoding layer for abstract feature extraction, and an attention layer to introduce an attention mechanism, ultimately achieving accurate track circuit fault classification through a fault diagnosis layer.

[0004] While the methods described above have achieved good results in bearing fault diagnosis, the network structure and parameters are derived from image recognition / speech recognition fields or obtained through repeated parameter tuning. For example, the length of the input signal is often defined as 1024, 2048, 4096, etc., while the size of the convolutional kernel is usually defined as 3×3, 5×5, 7×7, etc. This leads to two obvious problems: first, manually tuning or optimizing algorithms to find the optimal parameters reduces the network design efficiency; second, network parameters suitable for image recognition / speech recognition may not be suitable for fault diagnosis, or at least not optimal, and lack mechanistic interpretability. Summary of the Invention

[0005] Therefore, it is necessary to provide a design method, apparatus, and device for a CNN model for bearing fault diagnosis to address the aforementioned technical problems.

[0006] A design method for a CNN model for bearing fault diagnosis, the method comprising:

[0007] Based on the geometric and operating parameters of the bearing under test, the theoretical fault characteristic frequency value is calculated.

[0008] The length of the input signal is determined based on the theoretical fault characteristic frequency value and the sampling frequency.

[0009] The width and height of the input signal are determined based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal.

[0010] Based on the exponential decay characteristics of the fault spikes and the periodicity of the fault characteristic frequency distribution corresponding to each bearing fault category, the width and height of the convolution kernel in the corresponding channel are determined.

[0011] Based on the number of bearing fault categories, the number of channels in the CNN model and the size parameters of each channel are determined to obtain a CNN model for bearing fault diagnosis.

[0012] The fault diagnosis sample of the bearing to be tested is input into the corresponding channel of the CNN model. Then, the output results of all channels are fused through the fusion layer. The fused result is then processed in the output layer to obtain the bearing fault diagnosis result.

[0013] In one embodiment, theoretical fault characteristic frequency values ​​are calculated based on the geometric parameters and operating parameters of the bearing under test; the theoretical fault characteristic frequency values ​​include: outer ring fault characteristic frequency, inner ring fault characteristic frequency, ball fault characteristic frequency, and cage fault characteristic frequency.

[0014] In one embodiment, determining the length of the input signal based on the theoretical fault characteristic frequency value and the sampling frequency includes:

[0015] Based on the theoretical fault characteristic frequency value and sampling frequency, calculation criteria for determining the length of the input signal are designed in both the time and frequency domains, respectively, to determine the length of the input signal in both the time and frequency domains; wherein, the calculation criteria for the length of the input signal in the time domain are as follows:

[0016]

[0017] Where L1 is the length of the input signal in the time domain, f BPFO f BPFI f BSF and f FTF These represent the characteristic frequencies of outer ring failure, inner ring failure, ball failure, and cage failure, respectively, with λ being an adjustment factor, λ≥1, and f s The sampling frequency.

[0018] The criteria for calculating the length of the input signal in the frequency domain are as follows:

[0019]

[0020] Where L2 is the length of the input signal in the frequency domain, f FCF ∈{f BPFO f BPFI f BSF,f FTF}, i and j are the orders of the fault characteristic frequencies, where f FCF ·i≠f FCF ·j.

[0021] The length of the input signal is determined based on the length of the input signal in the time domain and the length of the input signal in the frequency domain; the length of the input signal is:

[0022] L s =max(L1,L2)

[0023] Among them, L s The length of the input signal.

[0024] In one embodiment, the width and height of the input signal are determined based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal, including:

[0025] Based on the sampling frequency and the rotation frequency of the shaft, the width of the input signal is determined as follows:

[0026]

[0027] Among them, W s f is the width of the input signal. r f is the rotational frequency of the shaft. s Sampling frequency, This is the floor operator.

[0028] Based on the length and width of the input signal, the height of the input signal is determined as follows:

[0029]

[0030] Among them, H s L is the height of the input signal. s The length of the input signal.

[0031] In one embodiment, the width and height of the convolution kernel in the corresponding channel are determined based on the exponential decay characteristics of the fault spikes corresponding to each bearing fault category and the periodicity of the fault characteristic frequency distribution, including:

[0032] Acceleration signals of the faulty bearing under different fault types were collected.

[0033] The acceleration signal is preprocessed to obtain the acceleration signal of the bearing during a single fault cycle.

[0034] Calculate the outer envelope of the acceleration signal and fit the outer envelope using an exponential function.

[0035] Based on the defined attenuation factors, the corresponding attenuation time and signal observation window are determined.

[0036] The number of sample points collected within the time window from the peak decay of the outer envelope to the preset value is calculated, and the number of sample points collected is used as the width of the convolution kernel; the preset value is ζA0, where A0 is the peak value of the outer envelope and ζ is the amplitude decay factor.

[0037] The height of the convolution kernel is determined based on the ratio of the theoretical fault characteristic frequency value to the rotation frequency of the shaft.

[0038] In one embodiment, the height of the convolution kernel is determined based on the ratio of the theoretical fault characteristic frequency value to the rotation frequency of the shaft, including:

[0039] Calculate the ratio of the fault characteristic frequency to the rotation frequency of the shaft, and find a set of positive integers p and q based on the ratio; use the positive integer q as the height of the convolution kernel; wherein the difference between the quotient of the positive integer p divided by q and the ratio is not greater than a preset threshold.

[0040] In one embodiment, the width of the convolution kernel is:

[0041]

[0042] Among them, W k (ζ) represents the kernel width, t ζ B is the time window during which the outer envelope decays from its peak value to a preset value. ζ f is the attenuation coefficient. s Sampling frequency, This is the floor operator.

[0043] In one embodiment, the number of channels in the CNN model and the size parameters of each channel are determined based on the number of bearing fault categories to obtain a CNN model for bearing fault diagnosis, including:

[0044] The number of channels in the CNN model is designed to correspond to the number of bearing fault categories, with each channel representing one bearing fault category. Each channel consists of two convolutional modules, each including a convolutional layer and a pooling layer.

[0045] The parameters of the convolutional layer corresponding to each bearing fault category are designed, and the fusion layer and output layer after multiple channels are designed to obtain a CNN model for bearing fault diagnosis.

[0046] A design apparatus for a CNN model for bearing fault diagnosis, the apparatus comprising:

[0047] The theoretical fault characteristic frequency value calculation module is used to calculate the theoretical fault characteristic frequency value based on the geometric parameters and operating parameters of the bearing under test.

[0048] The input signal parameter design module is used to determine the length of the input signal based on the theoretical fault characteristic frequency value and the sampling frequency; and to determine the width and height of the input signal based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal.

[0049] The convolution kernel size design module is used to determine the width and height of the convolution kernel in the corresponding channel based on the exponential decay characteristics of the fault spikes and the periodicity of the fault characteristic frequency distribution for each type of bearing fault.

[0050] The network channel design module is used to determine the number of channels in the CNN model and the size parameters of each channel based on the number of bearing fault categories, so as to obtain a CNN model for bearing fault diagnosis.

[0051] The bearing fault diagnosis module is used to input the fault diagnosis samples of the bearing to be tested into the corresponding channels of the CNN model, and then fuse the output results of all channels through the fusion layer. The fusion result is then processed in the output layer to obtain the bearing fault diagnosis result.

[0052] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of any of the methods described above.

[0053] The aforementioned design method, apparatus, and equipment for a CNN model used in bearing fault diagnosis firstly designs the length of the CNN model input signal based on the fault feature frequencies and their resolution requirements; secondly, it determines the width and size of the input signal based on the rotation frequency of the shaft; then, it calculates the width of the convolution kernel based on the exponential decay characteristics of the fault spikes; finally, it determines the number of channels in the CNN model and the size parameters of each channel based on the number of fault categories. The CNN model design method proposed in this invention can directly design a CNN model based on fault features, significantly improving the design efficiency of the CNN model and enhancing the interpretability of the network structure and parameters. Furthermore, the CNN model designed using this method exhibits higher accuracy, lower uncertainty, and faster convergence speed when used for bearing fault diagnosis. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the design method of a CNN model for bearing fault diagnosis in one embodiment;

[0055] Figure 2 Here is a flowchart for designing the convolution kernel width in another embodiment;

[0056] Figure 3 Here is a flowchart for designing the kernel height in another embodiment;

[0057] Figure 4Here is the structure of a CNN model in another embodiment;

[0058] Figure 5 Here is a comparison chart of CNN performance analysis in another embodiment, where (a) is the percentage improvement of CNN performance by different improvement measures, and (b) is a comparison chart of the performance of the CNN model designed by this method and the ordinary CNN model.

[0059] Figure 6 This is a structural block diagram of a design device for a CNN model used for bearing fault diagnosis in one embodiment;

[0060] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] In one embodiment, such as Figure 1 As shown, a design method for a CNN model for bearing fault diagnosis is provided, which includes the following steps:

[0063] Step 100: Calculate the theoretical fault characteristic frequency value based on the geometric parameters and operating parameters of the bearing under test.

[0064] Specifically, in theory, the fault signals of rotating equipment have inherent characteristics in the time and frequency domain. For example, when different components malfunction, different fault characteristic frequencies will appear in the acceleration spectrum.

[0065] By analyzing the time-frequency domain of fault signals, the correspondence between the structure and parameters of the CNN model and the fault characteristic frequencies is established, and this correspondence is used in the design of the CNN structure and parameters.

[0066] Step 102: Determine the length of the input signal based on the theoretical fault characteristic frequency value and the sampling frequency.

[0067] Specifically, the length of the input signal directly determines the amount of information contained in each sample, while the size of the input signal determines the distribution of information. To design a reasonable sample length, this application proposes design criteria in both the time and frequency domains, calculating the sample length limit. Finally, the sample length is determined by the maximum value calculated from the two criteria.

[0068] The length of the input signal for the CNN model is determined by the bearing fault characteristic frequency and the frequency resolution requirements for different orders and types of fault characteristic frequencies.

[0069] Step 104: Determine the width and height of the input signal based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal.

[0070] Specifically, in order to design the sample size, this application proposes a sample width design method based on the rotation axis frequency. Combined with the sample length, the optimal sample input size can be determined.

[0071] The width of the input signal for a CNN model is determined by the rotation frequency of the axis.

[0072] Step 106: Determine the width and height of the convolution kernel in the corresponding channel based on the exponential decay characteristics of the fault spikes and the periodicity of the fault characteristic frequency distribution for each type of bearing fault.

[0073] Specifically, the size of the convolution kernel defines its receptive field, directly affecting the efficiency and effectiveness of feature extraction. Bearing fault signals have inherent characteristics in the time-frequency domain. To improve the efficiency of feature extraction by the convolution kernel, a kernel width design method based on the attenuation characteristics of fault peaks is proposed. This method involves fitting the outer envelope of the faulty bearing's acceleration using an exponential function, and then defining different attenuation factors to obtain the corresponding kernel width. The height of the convolution kernel is determined by analyzing the distribution period of fault feature peaks in the samples.

[0074] In the time domain, fault spikes decay exponentially. Therefore, these fault mechanism characteristics should be considered when designing the network structure and parameters to improve the design efficiency of CNN models and enhance the interpretability of the network structure and parameters in terms of mechanism.

[0075] Step 108: Based on the number of bearing fault categories, determine the number of channels in the CNN model and the size parameters of each channel to obtain a CNN model for bearing fault diagnosis.

[0076] Specifically, the number of channels in a network has a significant impact on its performance. To further improve network performance, a multi-channel network structure design method is proposed. The number of channels is determined by the number of fault classification categories, and the convolution kernel parameters of each channel are calculated from the fault feature frequencies of the fault types they process. The results of the multi-channel network are finally fused through a fusion layer, and the output layer ultimately achieves fault diagnosis.

[0077] The number of channels in a CNN model is determined by the number of fault categories to be diagnosed, with each channel corresponding to one fault category.

[0078] The structure and main size parameters of the CNN model are directly derived from the fault mechanism, establishing a direct relationship between each parameter and the frequency of fault characteristics. The entire CNN network has good physical interpretability.

[0079] Step 110: Input the fault diagnosis sample of the bearing to be tested into the corresponding channel of the CNN model, then fuse the output results of all channels through the fusion layer, and process the fused result in the output layer to obtain the bearing fault diagnosis result.

[0080] Specifically, when the CNN model designed for bearing fault diagnosis is used for bearing fault diagnosis, the network exhibits higher accuracy, lower uncertainty, and faster convergence speed.

[0081] In the aforementioned design method for a CNN model for bearing fault diagnosis, the method first designs the length of the CNN model input signal based on the fault feature frequencies and their resolution requirements; secondly, it determines the width and size of the input signal based on the rotation frequency of the shaft; then, it calculates the width of the convolution kernel based on the exponential decay characteristics of the fault spikes; finally, it determines the number of channels in the CNN model and the size parameters of each channel based on the number of fault categories. The CNN model design method proposed in this invention can directly design a CNN model based on fault features, significantly improving the design efficiency of the CNN model and enhancing the interpretability of the network structure and parameters. Furthermore, the CNN model designed using this method exhibits higher accuracy, lower uncertainty, and faster convergence speed when used for bearing fault diagnosis.

[0082] In one embodiment, the theoretical fault characteristic frequency values ​​in step 100 include: outer ring fault characteristic frequency, inner ring fault characteristic frequency, ball fault characteristic frequency, and cage fault characteristic frequency.

[0083] Specifically, when different components of a bearing fail, corresponding fault characteristic frequencies and their higher harmonics will appear in the acceleration spectrum. For example, if the outer race of the bearing fails, the corresponding fault characteristic frequency is the Ball Passing Frequency of Outer race (BPFO); if the inner race fails, the corresponding fault characteristic frequency is the Ball Passing Frequency of Inner race (BPFI); if the bearing cage fails, the corresponding fault characteristic frequency is the Fundamental Train Frequency (FTF); and if the bearing balls fail, the corresponding fault characteristic frequency is the Ball Spin Frequency (BSF). The calculation formulas for the above fault characteristic frequencies are shown below:

[0084]

[0085] Among them, fr Let d be the rotational frequency of the shaft, d be the diameter of the bearing balls, D be the bearing's mean diameter, α be the bearing's initial contact angle, and n be the number of bearing balls. Therefore, given the geometric and operating parameters of the bearing under test, four theoretical fault characteristic frequency values ​​can be calculated.

[0086] In one embodiment, step 102 includes: designing calculation criteria for determining the length of the input signal in the time domain and frequency domain, respectively, based on the theoretical fault characteristic frequency value and the sampling frequency, to determine the length of the input signal in the time domain and the length of the input signal in the frequency domain; wherein, the calculation criteria for the length of the input signal in the time domain is as follows:

[0087]

[0088] Where L1 is the length of the input signal in the time domain, f BPFO f BPFI f BSF and f FTF These represent the characteristic frequencies of outer ring failure, inner ring failure, ball failure, and cage failure, respectively, with λ being an adjustment factor, λ≥1, and f s The sampling frequency is the acceleration. The specific calculation formulas for the outer ring fault characteristic frequency, inner ring fault characteristic frequency, ball fault characteristic frequency, and cage fault characteristic frequency are given in equations (1)-(4).

[0089] The criteria for calculating the length of the input signal in the frequency domain are as follows:

[0090]

[0091] Where L2 is the length of the input signal in the frequency domain, f FCF ∈{f BPFO f BPFI f BSF ,f FTF}, i and j are the orders of the fault characteristic frequencies, where f FCF ·i≠f FCF ·j.

[0092] The length of the input signal is determined based on the length of the input signal in the time domain and the length of the input signal in the frequency domain; the length of the input signal is:

[0093] L s =max(L1,L2) (7)

[0094] Among them, L s The length of the input signal.

[0095] Specifically, the length of the input signal directly determines the amount of information contained in each sample. If the length is too short, the sample contains too little information, which is not conducive to fault diagnosis. Conversely, if the sample is too long, there will be information redundancy, increasing the time for network training and inference. Therefore, designing an appropriate sample length is of great value for fault diagnosis.

[0096] To design a reasonable sample length, this invention proposes two design criteria: ① In the time domain, the sample length should be greater than or equal to the number of sample points collected within a single fault cycle under any fault type. ② In the frequency domain, after spectral transformation of the samples, the frequency resolution should be high enough to distinguish different fault characteristic frequencies under a given order.

[0097] Criterion ① is shown in formula (5); Criterion ② is shown in formula (6).

[0098] Based on the above two criteria, the length of the output signal can be determined to be the maximum value of L1 and L2. That is, the input length of the sample should satisfy the above two design criteria at the same time, as shown in formula (7).

[0099] In one embodiment, step 104 includes: determining the width of the input signal based on the sampling frequency and the rotation frequency of the shaft.

[0100]

[0101] Among them, W s f is the width of the input signal. r f is the rotational frequency of the shaft. s Sampling frequency, This is the floor operator.

[0102] Based on the length and width of the input signal, the height of the input signal is determined as follows:

[0103]

[0104] Among them, H s L is the height of the input signal. s The length of the input signal.

[0105] Specifically, as can be seen from formulas (1) to (4), different fault characteristic frequencies all share a common component, namely the rotational frequency f of the shaft. r This means that the signal acquired within one rotation cycle of the shaft is the minimum cycle common to different fault types. Therefore, the width of the input signal is defined as the number of sample points acquired within one rotation of the shaft, and its calculation formula is shown in Equation (8).

[0106] After determining the width W of the input signal sThen, combined with the length L of the input signal s This allows for further determination of the height H of the input signal. s The calculation formula is shown in equation (9).

[0107] Therefore, the size of the input signal can be determined as (W) s H s ).

[0108] In one embodiment, step 106 includes: acquiring acceleration signals of the faulty bearing under different fault types; preprocessing the acceleration signals to obtain the acceleration signals of the bearing within a single fault cycle; calculating the outer envelope of the acceleration signals and fitting the outer envelope using an exponential function; determining the corresponding decay time and signal observation window according to different defined decay factors; calculating the number of sample points acquired within the time window when the outer envelope decays from its peak to a preset value, and using the number of acquired sample points as the width of the convolution kernel; the preset value is ζA0, where A0 is the peak value of the envelope and ζ is the amplitude decay factor; and determining the height of the convolution kernel according to the ratio of the theoretical fault characteristic frequency value to the rotation frequency of the shaft.

[0109] In one embodiment, the width of the convolution kernel is:

[0110]

[0111] Among them, W k (ζ) represents the kernel width, t ζ B is the time window during which the outer envelope decays from its peak value to a preset value. ζ f is the attenuation coefficient. s Sampling frequency, This is the floor operator.

[0112] Specifically, the convolutional kernels of CNN models essentially extract features from signals at different scales, and the size of the convolutional kernel defines the receptive field range for feature extraction. Therefore, designing a reasonable convolutional kernel size is crucial for improving the efficiency and effectiveness of feature extraction.

[0113] like Figure 2 As shown, the convolution kernel design method proposed in this application consists of 6 sub-steps. The content of each sub-step will be described below:

[0114] (1) Collect the acceleration of the faulty bearing

[0115] First, the acceleration signals of the bearing under different fault types are collected by an acceleration sensor. The acceleration sensor can be placed in the bearing housing or mounted on the outer ring of the bearing by drilling holes.

[0116] (2) Obtain the signal of a single fault cycle

[0117] To facilitate the analysis of the acceleration response characteristics of a faulty bearing, the continuously acquired acceleration data is segmented to obtain the bearing's acceleration within a single fault cycle. Specifically, the maximum value of the acceleration within a certain range (greater than one fault cycle but less than two fault cycles) is defined as the peak value of the outer envelope, and this peak value is used as the starting point of each fault cycle. Then, based on the fault characteristic frequency and the acceleration sampling frequency, the number of fault cycle sampling points N under different fault types can be calculated. FCF As shown in formula (11).

[0118]

[0119] Starting from the fault peak, the subsequent N FCF -1 points represent the sample points collected within one fault cycle, i.e., the signals collected in a single fault cycle.

[0120] (3) Calculate the outer envelope of acceleration

[0121] Since raw acceleration often contains a lot of noise and its signal amplitude fluctuates greatly, it is not convenient to directly model it. Therefore, this invention calculates the outer envelope of the acceleration, considering only the positive portion of the acceleration. Subsequent steps will also be based on the signal's outer envelope.

[0122] (4) Fit the outer envelope using an exponential function.

[0123] Since the acceleration response of the faulty bearing exhibits an exponential decay trend, after obtaining the outer envelope of the acceleration, this invention uses the model of formula (12) to model the outer envelope of the acceleration.

[0124]

[0125] Where A0 is the peak value of the acceleration envelope, B ζ ζ is the attenuation coefficient, ζ is the amplitude attenuation factor (0 < ζ < 1), t is time, and A(t) is the envelope amplitude at time t.

[0126] (5) Define different attenuation factors

[0127] The fault spike decays from its peak value A0, and the time required to decay to different proportions varies, resulting in different numbers of sample points collected within different time windows. The larger ζ is, the smaller the peak decay ratio and the narrower the corresponding signal window; conversely, the smaller ζ is, the larger the peak decay ratio and the wider the corresponding signal window. Therefore, different signal observation windows of different sizes can be obtained by defining different attenuation factors ζ, as shown in formula (13).

[0128]

[0129] (6) Calculate the kernel width under different decay factors.

[0130] Different attenuation factors correspond to different attenuation times and signal observation windows, representing the response characteristics of acceleration observed at different scales, which corresponds precisely to the receptive field of the convolution kernel. Therefore, this invention proposes a convolution kernel width design method based on attenuation factors. The method calculates the number of sample points collected within the time window from the peak A0 to ζA0 of the acceleration envelope, and defines this value as the convolution kernel width. The specific calculation formula is shown in Equation (10).

[0131] In one embodiment, determining the height of the convolution kernel based on the ratio of the theoretical fault characteristic frequency value to the rotation frequency of the shaft includes: calculating the ratio of the fault characteristic frequency to the rotation frequency of the shaft; finding a set of positive integers p and q based on the ratio; using the positive integer q as the height of the convolution kernel; wherein the difference between the quotient of the positive integer p divided by q and the ratio is not greater than a preset threshold.

[0132] Specifically, when a convolution kernel performs a convolution operation with the input signal, it is generally performed from left to right and from top to bottom. For fault diagnosis, the convolution kernel extracts fault features within its receptive field, and for a faulty bearing, the most important feature is the fault spike. To improve the accuracy of fault diagnosis, theoretically, each convolution operation should contain at least one fault spike. Therefore, to design a reasonable convolution kernel height, this invention proposes an analysis method based on the distribution pattern of fault spikes. For example... Figure 3 As shown, the sample points collected in each fault cycle can be calculated by formula (14), and the width of the input signal has been determined in step 3, and the specific value can be obtained by formula (15). In order to ensure that there is at least one fault spike in the receptive field of each convolution operation, theoretically, an integer number of fault cycles should be completely distributed in the integer rows of the input samples. This relationship can be expressed by formula (16), where p and q are both positive integers. Further derivation shows that the ratio of positive integers p and q is exactly equal to the ratio of each fault feature frequency to the rotation frequency of the axis. Therefore, this application proposes a method for designing the convolution kernel height based on this ratio. Specifically, for different fault types, the ratio of fault feature frequency to axis frequency is first calculated, and then two positive integers p and q are found such that As close to k as possible f Finally, the height of the convolution kernel can be designed as q.

[0133]

[0134] p×N FCF =q×N r(16)

[0135]

[0136] Where p and q are both positive integers, and k f It is the ratio of the fault characteristic frequency to the rotational frequency of the shaft.

[0137] In one embodiment, step 108 includes: designing the number of channels of the CNN model to correspond to the number of bearing fault categories, with each channel corresponding to one bearing fault category; each channel consists of two convolutional modules; wherein the convolutional module includes a convolutional layer and a pooling layer; designing the parameters of the convolutional layer corresponding to each bearing fault category; designing a fusion layer and an output layer after multiple channels to obtain a CNN model for bearing fault diagnosis.

[0138] Specifically, besides parameter design, the network structure, especially the number of channels, also affects network performance. Therefore, this invention proposes a multi-channel design method, with the structure as follows: Figure 4 As shown. Specifically, the number of channels *n* in the network is determined by the number of fault categories. For example, if there are 3 different types of faults in the samples to be classified, then *n* = 3. The number of channels for other fault categories is designed similarly. Each channel consists of two convolutional modules; each convolutional module includes a convolutional layer and a pooling layer. The parameters of the convolutional layer are calculated based on the fault type processed by that channel. For example, if channel 1 processes outer ring faults, the size of the convolutional kernel is calculated from the characteristic frequency of outer ring faults; if channel 2 processes inner ring faults, its convolutional kernel size is calculated from the characteristic frequency of inner ring faults. The outputs of multiple channels are combined through a fusion layer, and finally, the output layer provides the result. The output layer can generally be softmax. For the sample size design under multi-channel conditions, the length and size of the optimal input signal for each channel are first calculated, and then the maximum value of all channels is taken as the input size of the entire network.

[0139] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0140] In a verification embodiment, the performance of the CNN model designed by this method is compared with that of a regular CNN model. The CNN performance analysis comparison chart is shown below. Figure 5 As shown, (a) represents the percentage improvement in CNN performance due to different improvement measures, and (b) is a performance comparison between the CNN model designed in this method and the ordinary CNN model. Figure 5 It can be observed that the input signal size design, convolution kernel size design, and multi-channel design proposed in this method can effectively improve the fault diagnosis accuracy of CNN models. The CNN designed based on this method has better performance than conventional CNNs in terms of accuracy, robustness, speed, and stability.

[0141] In one embodiment, such as Figure 6 As shown, a design device for a CNN model for bearing fault diagnosis is provided, including: a theoretical fault feature frequency value calculation module, an input signal parameter design module, a convolution kernel size design module, a network channel design module, and a bearing fault diagnosis module, wherein:

[0142] The theoretical fault characteristic frequency value calculation module is used to calculate the theoretical fault characteristic frequency value based on the geometric parameters and operating parameters of the bearing under test.

[0143] The input signal parameter design module is used to determine the length of the input signal based on the theoretical fault characteristic frequency value and the sampling frequency; and to determine the width and height of the input signal based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal.

[0144] The convolution kernel size design module is used to determine the width and height of the convolution kernel in the corresponding channel based on the exponential decay characteristics of the fault spikes and the periodicity of the fault characteristic frequency distribution for each type of bearing fault.

[0145] The network channel design module is used to determine the number of channels in the CNN model and the size parameters of each channel based on the number of bearing fault categories, so as to obtain a CNN model for bearing fault diagnosis.

[0146] The bearing fault diagnosis module is used to input the fault diagnosis samples of the bearing to be tested into the corresponding channels of the CNN model, and then fuse the output results of all channels through the fusion layer. The fusion result is then processed in the output layer to obtain the bearing fault diagnosis result.

[0147] In one embodiment, the theoretical fault characteristic frequency values ​​in the theoretical fault characteristic frequency value calculation module include: outer ring fault characteristic frequency, inner ring fault characteristic frequency, ball fault characteristic frequency, and cage fault characteristic frequency.

[0148] In one embodiment, the input signal parameter design module is further configured to design calculation criteria for determining the length of the input signal in the time domain and frequency domain, respectively, based on the theoretical fault characteristic frequency value and the sampling frequency, thereby determining the length of the input signal in the time domain and the length of the input signal in the frequency domain; wherein, the length of the input signal in the time domain is as shown in formula (5). The length of the input signal in the frequency domain is as shown in formula (6); the maximum value between the length of the input signal in the time domain and the length of the input signal in the frequency domain is taken as the length of the input signal.

[0149] In one embodiment, step 104 includes: determining the width of the input signal as shown in formula (8) based on the sampling frequency and the rotation frequency of the shaft; and determining the height of the input signal as shown in formula (9) based on the length and width of the input signal.

[0150] In one embodiment, the convolution kernel size design module is further used to acquire acceleration signals of the faulty bearing under different fault types; preprocess the acceleration signals to obtain the acceleration signals of the bearing within a single fault cycle; calculate the outer envelope of the acceleration signals and fit the outer envelope using an exponential function; determine the corresponding decay time and signal observation window according to different defined decay factors; calculate the number of sample points acquired within the time window when the outer envelope decays from the peak to a preset value, and use the number of acquired sample points as the width of the convolution kernel; the preset value is ζA0, where A0 is the peak value of the outer envelope and ζ is the amplitude decay factor; and determine the height of the convolution kernel according to the ratio of the theoretical fault characteristic frequency value to the rotation frequency of the shaft.

[0151] In one embodiment, the convolution kernel size design module is further configured to calculate the ratio of the fault characteristic frequency to the rotation frequency of the shaft, find a set of positive integers p and q based on the ratio, and use the positive integer q as the height of the convolution kernel; wherein the difference between the quotient of the positive integer p divided by q and the ratio is not greater than a preset threshold.

[0152] In one embodiment, the width of the convolution kernel in the kernel size design module is as shown in formula (10).

[0153] In one embodiment, the network channel design module is further configured to design the number of channels of the CNN model to correspond to the number of bearing fault categories, with each channel corresponding to one bearing fault category; each channel consists of two convolutional modules; wherein the convolutional module includes a convolutional layer and a pooling layer; the parameters of the convolutional layer corresponding to each bearing fault category are designed according to each bearing fault category, and a fusion layer and an output layer after multiple channels are designed to obtain a CNN model for bearing fault diagnosis.

[0154] Specific limitations regarding the design apparatus for the CNN model used for bearing fault diagnosis can be found in the above-mentioned limitations on the design method of the CNN model for bearing fault diagnosis, and will not be repeated here. Each module in the aforementioned design apparatus for the CNN model used for bearing fault diagnosis can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0155] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a design method for a CNN model for bearing fault diagnosis. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0156] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0157] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiment.

[0158] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0159] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A design method for a CNN model for bearing fault diagnosis, characterized in that, The method includes: Calculate the theoretical fault characteristic frequency value based on the geometric parameters and operating parameters of the bearing under test; The length of the input signal is determined based on the theoretical fault characteristic frequency value and the sampling frequency. The width and height of the input signal are determined based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal. Based on the exponential decay characteristics of the fault peaks and the periodicity of the fault characteristic frequency distribution corresponding to each bearing fault category, the width and height of the convolution kernel in the corresponding channel are determined. Specifically, this includes: acquiring acceleration signals of the faulty bearing under different fault types; preprocessing the acceleration signals to obtain the acceleration signal of the bearing within a single fault cycle; calculating the outer envelope of the acceleration signal and fitting the outer envelope using an exponential function; determining the corresponding decay time and signal observation window according to different defined decay factors; calculating the number of sample points acquired within the time window when the outer envelope decays from the peak to a preset value, and using the number of acquired sample points as the width of the convolution kernel; the preset value is... ,in, The peak value of the outer envelope. The amplitude attenuation factor is used; the height of the convolution kernel is determined based on the ratio of the theoretical fault characteristic frequency value to the rotation frequency of the shaft; specifically, it includes: calculating the ratio of the fault characteristic frequency to the rotation frequency of the shaft, finding a set of positive integers p and q based on the ratio; using the positive integer q as the height of the convolution kernel; wherein, the difference between the quotient of the positive integer p divided by q and the ratio is not greater than a preset threshold. Based on the number of bearing fault categories, determine the number of channels in the CNN model and the size parameters of each channel to obtain a CNN model for bearing fault diagnosis. The fault diagnosis sample of the bearing to be tested is input into the corresponding channel of the CNN model. Then, the output results of all channels are fused through the fusion layer. The fused result is then processed in the output layer to obtain the bearing fault diagnosis result.

2. The method according to claim 1, characterized in that, Calculate the theoretical fault characteristic frequency value based on the geometric parameters and operating parameters of the bearing under test; The theoretical fault characteristic frequency values ​​include: outer ring fault characteristic frequency, inner ring fault characteristic frequency, ball fault characteristic frequency, and cage fault characteristic frequency.

3. The method according to claim 1, characterized in that, Based on the theoretical fault characteristic frequency value and sampling frequency, the length of the input signal is determined, including: Based on the theoretical fault characteristic frequency value and sampling frequency, calculation criteria for determining the length of the input signal are designed in both the time and frequency domains to determine the length of the input signal in both the time and frequency domains; wherein, the calculation criteria for the length of the input signal in the time domain are as follows: in, The length of the input signal in the time domain. , , and These are the characteristic frequencies of outer ring failure, inner ring failure, ball failure, and cage failure, respectively. As a regulating factor, , The sampling frequency; The criteria for calculating the length of the input signal in the frequency domain are as follows: in, The length of the input signal in the frequency domain. , , , , and Let be the order of the fault characteristic frequencies, where ; The length of the input signal is determined based on the length of the input signal in the time domain and the length of the input signal in the frequency domain; the length of the input signal is: in, The length of the input signal.

4. The method according to claim 1, characterized in that, Based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal, the width and height of the input signal are determined, including: Based on the sampling frequency and the rotation frequency of the shaft, the width of the input signal is determined as follows: in, The width of the input signal. The rotational frequency of the shaft. Sampling frequency, This is the floor function operator; Based on the length and width of the input signal, the height of the input signal is determined as follows: in, The height of the input signal, The length of the input signal.

5. The method according to claim 1, characterized in that, The width of the convolution kernel is: in, The width of the convolution kernel. The time window for the outer envelope to decay from its peak value to a preset value. The attenuation coefficient is... Sampling frequency, This is the floor operator.

6. The method according to claim 1, characterized in that, Based on the number of bearing fault categories, the number of channels in the CNN model and the size parameters of each channel are determined to obtain a CNN model for bearing fault diagnosis, including: The number of channels in the CNN model is designed to correspond to the number of bearing fault categories, with each channel representing one bearing fault category; each channel consists of two convolutional modules; wherein each convolutional module includes a convolutional layer and a pooling layer. The parameters of the convolutional layer corresponding to each bearing fault category are designed, and the fusion layer and output layer after multiple channels are designed to obtain a CNN model for bearing fault diagnosis.

7. A design apparatus for a CNN model for bearing fault diagnosis, characterized in that, The device includes: The theoretical fault characteristic frequency value calculation module is used to calculate the theoretical fault characteristic frequency value based on the geometric parameters and operating parameters of the bearing under test. The input signal parameter design module is used to determine the length of the input signal based on the theoretical fault characteristic frequency value and the sampling frequency; and to determine the width and height of the input signal based on the sampling frequency, the rotation frequency of the shaft, and the length of the input signal. The convolution kernel size design module is used to determine the width and height of the convolution kernel in the corresponding channel based on the exponential decay characteristics of the fault peak and the periodicity of the fault characteristic frequency distribution for each bearing fault category. Specifically, it includes: acquiring acceleration signals of the faulty bearing under different fault types; preprocessing the acceleration signals to obtain the acceleration signal of the bearing within a single fault cycle; calculating the outer envelope of the acceleration signal and fitting the outer envelope using an exponential function; determining the corresponding decay time and signal observation window according to different defined decay factors; calculating the number of sample points acquired within the time window when the outer envelope decays from the peak to a preset value, and using the acquired sample points as the width of the convolution kernel; the preset value is... ,in, The peak value of the outer envelope. The amplitude attenuation factor is used; the height of the convolution kernel is determined based on the ratio of the theoretical fault characteristic frequency value to the rotation frequency of the shaft; specifically, it includes: calculating the ratio of the fault characteristic frequency to the rotation frequency of the shaft, finding a set of positive integers p and q based on the ratio; using the positive integer q as the height of the convolution kernel; wherein, the difference between the quotient of the positive integer p divided by q and the ratio is not greater than a preset threshold. The network channel design module is used to determine the number of channels in the CNN model and the size parameters of each channel based on the number of bearing fault categories, so as to obtain a CNN model for bearing fault diagnosis. The bearing fault diagnosis module is used to input the fault diagnosis sample of the bearing to be tested into the corresponding channel of the CNN model, and then fuse the output results of all channels through the fusion layer. The fusion result is then processed in the output layer to obtain the bearing fault diagnosis result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

Images