Rolling bearing fault identification method based on stochastic resonance and convolutional neural network

Abstract

The invention discloses a rolling bearing fault identification method based on stochastic resonance and a convolutional neural network, and the method comprises the steps: setting artificial fault points on an inner ring, an outer ring and a rolling body of a rolling bearing, collecting bearing vibration signals of the rolling bearing at different rotating speeds, constructing a self-adaptive variable-scale stochastic resonance detection system based on an improved signal-to-noise ratio, and carrying out the fault identification of the rolling bearing. The method comprises the following steps: respectively carrying out horizontal axis unified scale processing on spectrograms, forming an original image set by the processed spectrograms, marking and classifying the spectrograms in the original image set based on artificial fault points, dividing the marked and classified original image set into a training set and a test set, building a convolutional neural network model, and carrying out training and testing. And determining parameters of the convolutional neural network model according to a test result, representing the parameters as an identification model, performing fault identification on the obtained to-be-identified bearing vibration signal based on the identification model, and obtaining a fault type. The method solves the problem of weak fault identification of the rolling bearing under the working conditions of strong noise and rotating speed variation.

Description

Technical Field

[0001] The invention relates to the technical field of image recognition, and in particular to a rolling bearing fault recognition method based on stochastic resonance and convolutional neural network. Background Art

[0002] Rolling bearing is an important component widely used in mechanical equipment. It is mainly used to support rotating shaft parts and needs to withstand complex, cyclical alternating stress for a long time. Therefore, rolling bearing is a vulnerable part. The quality of its running state will play a vital role in whether the whole equipment can operate healthily and reliably. Therefore, rolling bearing fault diagnosis, especially early rolling bearing fault diagnosis, has very important practical significance.

[0003] Surface damage faults (such as pitting, spalling and abrasions) are common faults in the operation of rolling bearings. For this type of fault, the damage points on the working surface of the bearing element repeatedly hit the surfaces of other elements in contact with it during operation, which will produce "passing vibrations". The occurrence cycle is regular, and the frequencies generated at different fault locations are different. Detecting these signals is an effective means of diagnosing bearing faults. Usually, rolling bearing faults are manually judged by experts in various fields based on professional experience. The emergence of shallow intelligent models such as neural networks and support vector machines has replaced the traditional manual diagnosis process and ushered in the era of intelligent fault diagnosis. However, since rolling bearing fault diagnosis is often in the early stage of fault, and the fault signal is often accompanied by large noise, and the working conditions of rolling bearings are relatively complex, the working conditions are often unstable, such as the fluctuation of the speed, which causes the fault frequency of rolling bearings to fluctuate within a certain range, and the accuracy of identifying the type of rolling bearing fault is low. Summary of the invention

[0004] The present invention provides a rolling bearing fault identification method based on stochastic resonance and convolutional neural network to overcome the above technical problems.

[0005] Rolling bearing fault identification method based on stochastic resonance and convolutional neural network, including

[0006] S1. Artificial fault points are set on the inner and outer rings and rolling elements of the rolling bearing, and the vibration signals of the rolling bearing at different speeds are collected based on the acceleration sensor above the rolling bearing.

[0007] S2. constructing an adaptive variable-scale stochastic resonance detection system based on an improved signal-to-noise ratio, wherein the adaptive variable-scale stochastic resonance detection system is used to detect the collected bearing vibration signal to obtain a frequency spectrum corresponding to an artificial fault point.

[0008] S3, respectively perform horizontal axis uniform scale processing on the spectrograms, form the processed spectrograms into an original image set, mark and classify the spectrograms in the original image set based on the artificial fault points, and divide the marked and classified original image set into a training set and a test set,

[0009] S4. Build a convolutional neural network model, train the convolutional neural network model based on the training set, test the trained convolutional neural network model based on the test set and obtain the test results, determine the parameters of the convolutional neural network model according to the test results and express it as a recognition model,

[0010] S5. Obtain a bearing vibration signal to be identified, perform fault identification on the bearing vibration signal to be identified based on the identification model, and obtain the fault type.

[0011] Preferably, said S2 includes,

[0012] S21, building a particle swarm optimization model, including setting the number of particles to M, the maximum number of iterations to N, the search range of particles and the maximum search speed, the objective function and the detection range and detection position of the objective function, and initializing the particle swarm optimization model,

[0013] S22, obtaining a bearing vibration signal, setting a classification condition based on amplitude, frequency or noise, classifying the bearing vibration signal based on the classification condition, determining a variable scale stochastic resonance solution method, and decomposing the classified bearing vibration signal in the bearing vibration signal according to the variable scale stochastic resonance solution method,

[0014] S23, select a stochastic resonance system, input the M particles in the initialized particle swarm optimization model and the bearing vibration signal after variable scale decomposition into the stochastic resonance system and obtain system parameters, and obtain the optimal system parameters after N iterations.

[0015] S24. Based on the stochastic resonance system with optimal system parameters, after the bearing vibration signal is input into the stochastic resonance system, a frequency spectrum corresponding to the artificial fault point is obtained.

[0016] Preferably, the variable-scale stochastic resonance solution method is the Runge-Kutta method.

[0017] Preferably, the horizontal axis unification scale processing of the spectrum diagram includes making the total length of the coordinate axis of the horizontal axis of the spectrum diagram a fixed multiple of the rotational speed, that is, the horizontal axis range is 0-nr, n is a constant, r is the rotational speed, and the three fault characteristic frequencies are divided by nr and represented as the horizontal axis, that is, the position of the fault point in the spectrum diagram does not change with the change of the rotational speed.

[0018] Preferably, the stochastic resonance system is a bistable stochastic resonance system.

[0019] The present invention provides a rolling bearing fault identification method based on stochastic resonance and convolutional neural network. The collected vibration signal is detected by an adaptive variable-scale random detection resonance system, and then the frequency spectrum output by the adaptive variable-scale random resonance detection system is processed by horizontal axis uniform scale to construct a training set, and the convolutional neural network model is used for training. The method can identify the fault in the early stage of the rolling bearing fault, and can also complete the identification work under strong noise. The adaptive variable-scale random resonance system based on the improved signal-to-noise ratio can detect the fault signal within a range, so that the interference of unstable working conditions can be avoided. After the frequency spectrum output by the random resonance is processed by horizontal axis uniform scale, the problem of reduced recognition accuracy caused by speed change can be avoided, and the problem of weak fault identification of rolling bearings under strong noise and speed change working conditions can be solved. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0021] Figure 1 is a flow chart of a rolling bearing fault identification method of the present invention;

[0022] Figure 2 This is a spectrum diagram of random resonance detection of inner and outer ring and rolling element faults in an embodiment of the present invention;

[0023] Figure 3 (a) is a stochastic resonance spectrum diagram of a rolling element fault generated at a rotation speed of 100 r / min without uniform horizontal axis scaling, wherein the arrow points to the detected fault characteristic frequency;

[0024] Figure 3 (b) is a stochastic resonance spectrum diagram of a rolling element fault generated at a rotation speed of 200 r / min without horizontal axis uniform scaling, in which the arrow points to the detected fault characteristic frequency;

[0025] Figure 4 (a) is a stochastic resonance spectrum diagram of a rolling element fault generated at a rotation speed of 100 r / min and processed with a uniform horizontal axis scale in an embodiment of the present invention;

[0026] Figure 4 (b) is a stochastic resonance spectrum diagram of a rolling element fault generated at a rotation speed of 200 r / min and processed with a uniform horizontal axis scale in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] Figure 1 It is a flow chart of the method of the present invention, such as Figure 1 As shown, the method of this embodiment includes:

[0029] S1. Artificial fault points are set on the inner and outer rings and rolling elements of the rolling bearing, and the vibration signals of the rolling bearing at different speeds are collected based on the acceleration sensor above the rolling bearing.

[0030] S2. constructing an adaptive variable-scale stochastic resonance detection system based on an improved signal-to-noise ratio, wherein the adaptive variable-scale stochastic resonance detection system is used to detect the collected bearing vibration signal to obtain a frequency spectrum corresponding to an artificial fault point.

[0031] S3, respectively perform horizontal axis uniform scale processing on the spectrograms, form the processed spectrograms into an original image set, mark and classify the spectrograms in the original image set based on the artificial fault points, and divide the marked and classified original image set into a training set and a test set,

[0032] S4. Build a convolutional neural network model, train the convolutional neural network model based on the training set, test the trained convolutional neural network model based on the test set and obtain the test results, determine the parameters of the convolutional neural network model according to the test results and express it as a recognition model,

[0033] S5. Obtain a bearing vibration signal to be identified, perform fault identification on the bearing vibration signal to be identified based on the identification model, and obtain the fault type.

[0034] The present invention provides a rolling bearing fault identification method based on stochastic resonance and convolutional neural network. The collected vibration signal is detected by an adaptive variable-scale stochastic resonance system, and then the frequency spectrum output by the stochastic resonance system is processed by horizontal axis unified scale to construct a training set, and the convolutional neural network model is used for training. The method can identify the fault in the early stage of the rolling bearing fault, and can also complete the identification work under strong noise. The adaptive variable-scale stochastic resonance system based on the improved signal-to-noise ratio can detect the fault signal within a range, so that the interference of unstable working conditions can be avoided. After the frequency spectrum output by the stochastic resonance is processed by horizontal axis unified scale, the problem of reduced recognition accuracy caused by speed change can be avoided, and the problem of weak fault identification of rolling bearings under strong noise and speed change working conditions can be solved.

[0035] Specifically, this embodiment provides the details of the rolling bearing fault identification method, including:

[0036] S1. Artificial fault points are set on the inner and outer rings and rolling elements of the rolling bearing. The vibration signals of the rolling bearing at different speeds are collected based on the acceleration sensor above the rolling bearing. Specifically, an artificial fault with a width of 0.5 mm is cut out on the inner and outer rings and rolling elements of the rolling bearing by an electric spark wire cutting machine to simulate the weak faults in the actual operation of the rolling bearing. The vibration signals of different bearing faults are collected by the acceleration sensor arranged above the bearing seat of the rolling bearing. The speed of the bearing is adjusted from 40r / min to 200r / min by a speed regulator with a step length of 20r / min. The bearing runs for 10 minutes under each speed condition to collect the vibration signals of the bearings with different faults.

[0037] S2. constructing an adaptive variable-scale stochastic resonance detection system based on an improved signal-to-noise ratio, wherein the adaptive variable-scale stochastic resonance detection system is used to detect the collected bearing vibration signal to obtain a frequency spectrum corresponding to an artificial fault point.

[0038] The S2 includes:

[0039] S21, building a particle swarm optimization model, including setting the number of particles to M, the maximum number of iterations to N, the search range of particles and the maximum search speed, the objective function and the detection range and detection position of the objective function, and initializing the particle swarm optimization model,

[0040] S22. Obtain a bearing vibration signal, set classification conditions based on amplitude, frequency or noise, classify the bearing vibration signal based on the classification conditions, determine a variable-scale random resonance solution method, and decompose the classified bearing vibration signal in the bearing vibration signal according to the variable-scale random resonance solution method.

[0041] Specifically, based on amplitude, frequency or noise, the bearing vibration signal is divided into small-scale signal and large-scale signal. Small-scale signal refers to small amplitude, small frequency and small noise signal, and large-scale signal refers to large amplitude, large frequency and large noise signal. The amplitude classification threshold, frequency classification threshold and noise classification threshold are set. The signal with amplitude less than the amplitude classification threshold is represented as a small-scale signal, the signal with amplitude greater than or equal to the amplitude classification threshold is represented as a large-scale signal, the signal with frequency less than the frequency classification threshold is represented as a small-scale signal, the signal with frequency greater than or equal to the frequency classification threshold is represented as a large-scale signal, the signal with noise less than the noise classification threshold is represented as a small-scale signal, and the signal with noise greater than or equal to the noise classification threshold is represented as a large-scale signal. Since the classical stochastic resonance system is constrained by the adiabatic approximation theory and is only applicable to small parameter signal conditions, but the fault signals in actual engineering applications are usually large parameter signals, it is necessary to transform the large parameter signal into a small-scale signal through the variable scale stochastic resonance solution method. The variable-scale stochastic resonance solution method is the Runge-Kutta method. Since the transformation scale R is too large, the iteration step is too large, and the iterative solution does not converge. In this embodiment, the transformation scale R is dynamically adjusted according to the Runge-Kutta method, so that the iterative solution tends to converge.

[0042] S23, selecting a stochastic resonance system, wherein the stochastic resonance system is a bistable stochastic resonance system, inputting each initialized particle and the bearing vibration signal after the scale-variable decomposition into the stochastic resonance system and obtaining system parameters, and obtaining the optimal system parameters after N iterations,

[0043] S24. Based on the stochastic resonance system with optimal system parameters, after the bearing vibration signal is input into the stochastic resonance system, a frequency spectrum corresponding to the artificial fault point is obtained.

[0044] Specifically, the maximum search speed of the particles is 20% of the feasible domain of the independent variable, the number of particles, i.e., the population number, is 50, and the independent variables are the parameters a and b of the stochastic resonance system. Determine the objective function of the particle swarm optimization algorithm, the detection range of the objective function, and the detection position; the objective function of the particle swarm optimization algorithm is the improved signal-to-noise ratio (ISNR), and the expression of ISNR is:

[0045]

[0046] Where: ft is the theoretical fault frequency; S(ft) represents the total energy in the interval [ft-lΔf,ft+lΔf] centered at ft; l represents the detection range; Δf represents the ratio value; and N(ft) represents the total noise energy excluding S(ft).

[0047] Furthermore, the detection range is determined according to the detection position, and the detection position selection formula is:

[0048] kt=fr(K / fs) (2)

[0049] Where: kt is the detection position, K is the number of sampling points, and fs is the sampling frequency; kt is an integer. When kt has a decimal, it is rounded off.

[0050] Randomly initialize the particle's position x i (0) and speed v i (0), each position represents a set of system parameters a and b of stochastic resonance, and the speed represents the update step of system parameters a and b. The individual extreme value pbest of each particle i Set to the optimal position calculated in the current iteration, and set the position with the largest individual extreme value in the entire particle swarm to gbest i ;

[0051] The position of each particle and the preprocessed vibration signal are input into the stochastic resonance system, and the improved signal-to-noise ratio is selected as the fitness function, and then the fitness value of each particle is calculated. If the fitness value of the particle exceeds the fitness value of the current individual extreme value, the position of the particle is taken as the new individual extreme value pbest i If the maximum value of the individual extreme value in the entire particle swarm exceeds the current global optimal value, then the individual extreme value is taken as the new global optimal value gbest i ; Recalculate the speed and position of all particles according to the individual extreme values ​​and the global optimal value obtained; determine whether the current number of iterations reaches the maximum number of iterations N. If it meets the end conditions, directly output the optimal system parameters a and b, and input the optimal system parameters a and b and the bearing vibration signal obtained in the previous step into the bistable stochastic resonance system to detect weak periodic signals.

[0052] The stochastic resonance model mainly includes three elements: a nonlinear system, a weak periodic signal and noise. The nonlinear system used in the stochastic resonance system of this embodiment is a bistable system.

[0053] S3. Perform unified scale processing on the horizontal axis of the spectrum diagram respectively. The unified scale processing on the horizontal axis of the spectrum diagram includes making the total length of the coordinate axis of the horizontal axis of the spectrum diagram a fixed multiple of the rotational speed, that is, the horizontal axis range is (0-nr), n is a constant, r is the rotational speed, and the three fault characteristic frequencies are divided by nr and represented as the horizontal axis, that is, the position of the fault point in the spectrum diagram does not change with the change of the rotational speed.

[0054] The processed spectrograms are combined into an original image set, the spectrograms in the original image set are marked and classified based on the artificial fault points, and the marked and classified original image set is divided into a training set and a test set.

[0055] Specifically, Figure 2This is the frequency spectrum diagram of the random resonance detection of the inner and outer rings and rolling element faults in the present invention. Taking the 22209 bearing as an example, according to the theoretical fault frequency calculation formula, the theoretical fault frequency calculation formula is shown in formulas (3), (4), and (5):

[0056] Inner race fault characteristic frequency:

[0057]

[0058] Outer race fault characteristic frequency:

[0059]

[0060] Rolling element single failure frequency:

[0061]

[0062] Where Z is the number of rolling elements, f r is the rotation frequency, d is the rolling element diameter, D is the bearing pitch diameter, and α represents the bearing contact angle. The characteristic frequencies of the three faults can be calculated as: f I =0.2538r,f o =0.3462r,f c =0.0529r; fix the fault characteristic frequency in the spectrum diagram to avoid interference from speed changes.

[0063] Furthermore, let the total length of the horizontal axis of the stochastic resonance output spectrum be a fixed multiple of the rotation speed, that is, the horizontal axis range is (0-nr), so the positions of the three fault characteristic frequencies in the spectrum can be expressed as f I / nr=0.2538n,f o / nr=0.3462n,f c / nr=0.0529n, n is a constant; that is, the position of the spectrum line of the fault characteristic frequency in the spectrum diagram does not change with the change of the rotation speed. Figure 3 (a) is a stochastic resonance spectrum diagram of rolling element fault generated in the present invention at a speed of 100 r / min without horizontal axis uniform scale processing, wherein the arrow points to the detected fault characteristic frequency; Figure 3 (b) is a stochastic resonance spectrum diagram of rolling element fault generated in the present invention at a speed of 200 r / min without horizontal axis uniform scale processing, wherein the arrow points to the detected fault characteristic frequency; Figure 4 (a) is a stochastic resonance spectrum diagram of rolling element fault generated in the present invention at a rotation speed of 100 r / min and processed with a uniform horizontal axis scale; Figure 4 (b) is a stochastic resonance spectrum diagram of rolling element failure generated in the present invention at a rotation speed of 200 r / min and processed with a uniform horizontal axis scale.

[0064] The original image set is classified according to the fault type and the corresponding labels are set to construct a training set, and 30% of the images are randomly selected from the original image set to construct a test set; the images of the training set and the test set are processed with a unified scale on the horizontal axis, that is, the horizontal axis range is (0-nr), and r is the rotation speed.

[0065] S4, constructing a convolutional neural network model, training the convolutional neural network model based on a training set, testing the trained convolutional neural network model based on a test set and obtaining test results, determining parameters of the convolutional neural network model according to the test results and expressing them as a recognition model,

[0066] Specifically, the convolutional neural network model consists of convolution layer 1, activation layer 1, pooling layer 1, convolution layer 2, activation layer 2, convolution layer 3, activation layer 3, pooling layer 2, deep series layer 1, deep series layer 2, pooling layer 3, deep series layer 3, deep series layer 4, deep series layer 5, deep series layer 6, pooling layer 4, deep series layer 7, deep series layer 8, pooling layer 5, fully connected layer, and classification output layer.

[0067] Each convolutional layer: extracts features from the concentrated images to obtain feature maps;

[0068] Each activation layer: performs nonlinear mapping on the results of the corresponding convolutional layer;

[0069] Each pooling layer: selects the features extracted by the corresponding convolutional layer;

[0070] Each layer of depth concatenation layer: concatenates the results of the corresponding pooling layer;

[0071] Fully connected layer: converts all feature matrices of the pooling layer into one-dimensional feature vectors;

[0072] If the test result of the test set differs from the training result of the training set by more than 1% of the threshold, the learning rate, sampling number, and training round number are changed and retrained.

[0073] S5. Obtain a bearing vibration signal to be identified, perform fault identification on the bearing vibration signal to be identified based on the identification model, and obtain the fault type.

[0074] Overall beneficial effects:

[0075] The present invention provides a rolling bearing fault identification method based on stochastic resonance and convolutional neural network. The collected vibration signal is detected by an adaptive variable-scale stochastic resonance system, and then the frequency spectrum output by the stochastic resonance system is processed by horizontal axis unified scale to construct a training set, and the convolutional neural network model is used for training. The method can identify the fault in the early stage of the rolling bearing fault, and can also complete the identification work under strong noise. The adaptive variable-scale stochastic resonance system based on the improved signal-to-noise ratio can detect the fault signal within a range, so that the interference of unstable working conditions can be avoided. After the frequency spectrum output by the stochastic resonance is processed by horizontal axis unified scale, the problem of reduced recognition accuracy caused by speed change can be avoided, and the problem of weak fault identification of rolling bearings under strong noise and speed change working conditions can be solved.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rolling bearing fault identification method based on stochastic resonance and convolutional neural network, characterized in that: include S1. Artificial fault points are set on the inner and outer rings and rolling elements of the rolling bearing, and the vibration signals of the rolling bearing at different speeds are collected by the acceleration sensor above the rolling bearing. S2. constructing an adaptive variable-scale stochastic resonance detection system based on an improved signal-to-noise ratio, wherein the adaptive variable-scale stochastic resonance detection system is used to detect the collected bearing vibration signal to obtain a frequency spectrum corresponding to an artificial fault point. S3, respectively perform horizontal axis uniform scale processing on the spectrograms, form the processed spectrograms into an original image set, mark and classify the spectrograms in the original image set based on the artificial fault points, and divide the marked and classified original image set into a training set and a test set, S4. Build a convolutional neural network model, train the convolutional neural network model based on the training set, test the trained convolutional neural network model based on the test set and obtain the test results, determine the parameters of the convolutional neural network model according to the test results and express it as a recognition model, S5. Obtain a bearing vibration signal to be identified, perform fault identification on the bearing vibration signal to be identified based on the identification model, and obtain the fault type.

2. The rolling bearing fault identification method based on stochastic resonance and convolutional neural network according to claim 1 is characterized in that: The S2 comprises, S21, building a particle swarm optimization model, including setting the number of particles to M, the maximum number of iterations to N, the search range of particles and the maximum search speed, the objective function and the detection range and detection position of the objective function, and initializing the particle swarm optimization model, S22, obtaining a bearing vibration signal, setting a classification condition based on amplitude, frequency or noise, classifying the bearing vibration signal based on the classification condition, determining a variable scale stochastic resonance solution method, and decomposing the classified bearing vibration signal in the bearing vibration signal according to the variable scale stochastic resonance solution method, S23, select a stochastic resonance system, input the M particles in the initialized particle swarm optimization model and the bearing vibration signal after variable scale decomposition into the stochastic resonance system and obtain system parameters, and obtain the optimal system parameters after N iterations. S24. Based on the stochastic resonance system with optimal system parameters, after the bearing vibration signal is input into the stochastic resonance system with optimal system parameters, a frequency spectrum corresponding to the artificial fault point is obtained.

3. The rolling bearing fault identification method based on stochastic resonance and convolutional neural network according to claim 2 is characterized in that: The variable-scale stochastic resonance solution method is the Runge-Kutta method.

4. The rolling bearing fault identification method based on stochastic resonance and convolutional neural network according to claim 1 is characterized in that: The horizontal axis unification scale processing of the spectrum diagram includes making the total length of the coordinate axis of the horizontal axis of the spectrum diagram a fixed multiple of the rotational speed, that is, the horizontal axis range is 0-nr, n is a constant, r is the rotational speed, and the three fault characteristic frequencies are divided by nr and represented as the horizontal axis, that is, the position of the fault point in the spectrum diagram does not change with the change of the rotational speed.

5. The rolling bearing fault identification method based on stochastic resonance and convolutional neural network according to claim 2 is characterized in that: The stochastic resonance system is a bistable stochastic resonance system.

Images