A method for extracting instantaneous speed of rotating machinery from vibration signals

Abstract

The present invention discloses a method for extracting the instantaneous speed of a rotating machine from a vibration signal. A vibration signal sample set and a corresponding speed signal sample set of a target rotating machine are collected. Each sample set contains N signal samples of length T. The two sample sets are divided into a training set and a test set after preprocessing. The two sets are used to train a built deep learning model S-RNN. After the model training is completed, the model is used for online application to automatically extract the instantaneous speed from the vibration signal.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment status monitoring, and more specifically, relates to a method for extracting the instantaneous rotation speed of a rotating machine from a vibration signal based on deep learning. Background Art

[0002] Rotating machinery is widely used in production and life, such as electric motors, steam turbines, gas turbines, generators, gearboxes, centrifugal pumps, machine tools, diesel engines and gasoline engines. In engineering practice, rotating machinery usually operates under variable speed conditions. For example, elevators need to start and stop frequently to allow passengers to get on and off. Accordingly, the motor driving the elevator needs to increase or decrease the speed frequently. Another example is a wind turbine, whose speed changes randomly with the wind speed.

[0003] For machinery working under variable speed conditions, accurately obtaining speed information is crucial to understanding the operating status of mechanical equipment. At present, there are two main methods to obtain speed. The first is to install a speed sensor for direct measurement, and the second is to extract it from data measured by other sensors. The first method is simple and direct, but in some cases, it is impossible to install a speed sensor due to limitations in the equipment working space (such as a small-sized gearbox), and even if the sensor can be installed, it takes manpower, financial resources and material resources to purchase and install the speed sensor and data collector. The second method is an indirect measurement method, which extracts the speed from the data collected by existing sensors, avoiding the shortcomings of the first method mentioned above. Considering that vibration signals are easy to obtain and are widely collected in rotating machinery condition monitoring systems, the second method is usually used to extract the speed from vibration signals.

[0004] At present, the speed extraction from vibration signals is mainly based on signal processing methods. Its basic principle is to extract the most prominent spectral line in the vibration signal time-frequency spectrum, which is the instantaneous speed or the frequency multiple of the speed. However, this method is greatly affected by environmental noise, and the speed extraction accuracy is low under conditions where the speed fluctuates violently, and the speed extraction accuracy is seriously dependent on the resolution of the time-frequency spectrum. In order to improve the resolution of the time-frequency spectrum, for each newly acquired vibration signal, it is necessary to design reasonable parameters to calculate the time-frequency spectrum and adjust the frequency search direction, etc., which is labor-intensive. Therefore, it is not suitable for real-time acquisition of instantaneous speed in online status monitoring systems. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for extracting the instantaneous speed of a rotating machinery from a vibration signal based on deep learning, which can automatically extract the instantaneous speed directly from the vibration signal.

[0006] To achieve the above-mentioned object of the invention, the present invention provides a method for extracting the instantaneous speed of a rotating machine from a vibration signal, characterized in that it comprises the following steps:

[0007] (1) Collect the vibration signal sample set X∈R of the target rotating machinery N×T And the corresponding speed signal sample set S∈R N ×T , where R represents the real number domain, N is the number of signal samples in the sample set, and T is the sampling time of a single signal sample;

[0008] (2) The vibration signal sample set is normalized, and the speed signal sample set is normalized. Then, the processed vibration signal and speed signal are randomly divided into two non-repetitive subsets: a training set and a test set Among them, N train +N test =N;

[0009] (3) Build a deep learning model S-RNN based on recurrent neural network, and then use a single vibration signal sample X i As input, the corresponding extracted speed signal time series is the output, i=1,2,…,N;

[0010] (4) Training S-RNN: Using training set X train Train S-RNN so that the error between the speed output by S-RNN and the corresponding input speed is Error train Less than the given error threshold Δ Error ;

[0011] (5) Test S-RNN: Use the test set X test Test the trained S-RNN. If the error Error is obtained in the test set test test Less than the given error threshold Δ Error , then go to the next step, otherwise, return to (3);

[0012] (6) Online application: The vibration signal of the target rotating machinery is collected in real time, and after standardized processing, it is input into the S-RNN after the test is completed, so as to extract the corresponding instantaneous speed in real time.

[0013] The object of the invention of the present invention is achieved in this way:

[0014] The present invention discloses a method for extracting the instantaneous speed of a rotating machine from a vibration signal. The method comprises collecting a vibration signal sample set and a corresponding speed signal sample set of a target rotating machine, wherein each sample set comprises N signal samples of length T. The two sample sets are divided into a training set and a test set after preprocessing, and the two sets are used to train a built deep learning model S-RNN. After the model training is completed, the model is used for online application to automatically extract the instantaneous speed from the vibration signal.

[0015] At the same time, the method for extracting the instantaneous speed of a rotating machine from a vibration signal of the present invention also has the following beneficial effects:

[0016] (1) The present invention directly extracts the instantaneous speed from the vibration signal based on the neural network model, without the need to calculate the complex time-frequency spectrum;

[0017] (2) The present invention is simple, efficient, and requires little work, and is suitable for online status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of a method for extracting instantaneous rotation speed of a rotating machine from a vibration signal of the present invention;

[0019] Figure 2 This is a schematic diagram of the deep learning model S-RNN;

[0020] Figure 3 It is the test bench shown in the embodiment;

[0021] Figure 4 is the data set working condition shown in the embodiment;

[0022] Figure 5 is a sample example of an embodiment acceleration signal and an acceleration signal;

[0023] Figure 6 It is the model training accuracy and test accuracy;

[0024] Figure 7 is the instantaneous rotation speed extracted from the vibration signal by the S-RNN tested in the embodiment. DETAILED DESCRIPTION

[0025] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0026] Example

[0027] Figure 1 The present invention is a flow chart of a method for extracting the instantaneous rotation speed of a rotating machine from a vibration signal.

[0028] In this embodiment, if Figure 1 As shown, the present invention provides a method for extracting the instantaneous speed of a rotating machine from a vibration signal, comprising the following steps:

[0029] S1. Collect the vibration signal sample set X∈R of the target rotating machinery N×T And the corresponding speed signal sample set S∈R N×T , where R represents the real number domain, N is the number of signal samples in the sample set, and T is the sampling time of a single signal sample;

[0030] In this embodiment, the target rotating machinery includes but is not limited to rotating machinery such as motors, generators, gear boxes, pumps and fans; the vibration signal includes but is not limited to acceleration signals, speed signals, displacement signals, torque signals and sound pressure signals, etc.; and the speed signal can be the speed of any rotating shaft of the target machinery.

[0031] S2, normalize the vibration signal sample set, normalize the speed signal sample set, and then randomly divide the processed vibration signal and speed signal into two non-repetitive subsets: a training set and a test set Among them, N train +N test =N;

[0032] In this embodiment, the vibration signal sample set standardization process mainly includes returning the mean of the vibration signal to zero and the standard deviation to unity. The specific formula is:

[0033]

[0034] in, and σ train The training set X train The mean and standard deviation of .

[0035] The speed signal normalization process is mainly to reduce or expand the speed range to a certain range. The specific ratio value can be set according to actual needs, such as [0, 1] or [-1, 1] or (0, 1].

[0036] S3. Build a deep learning model S-RNN based on recurrent neural network (RNN), and then use a single vibration signal sample X i As input, the corresponding extracted speed signal time series is the output, i=1,2,…,N;

[0037] In this embodiment, the deep learning model S-RNN includes at least one recurrent network layer and several fully connected layers (Dense Layer), and the type of the recurrent network layer can be a recurrent neural network (RNN) or a gated recurrent neural network (Gated Recurrent Unit, GRU) or a long short-term memory network (Long Short-term Memory, LSTM), or one of the three network modifications, wherein the information of the recurrent network layer can be transmitted bidirectionally (Bi-directional) or unidirectionally (Unidirectional) on the time axis. Figure 2 As shown, the deep learning model S-RNN includes two recurrent network layers (Layer 1 and Layer 2) and two fully connected layers (Layer 4 and Layer 5). The recurrent neural network type used in the recurrent network layer is LSTM, and information is transmitted bidirectionally on the time axis, that is, BiLSTM, in which LSTM F-RNN uses sequential information transmission, while LSTM B-RNN uses reverse order information transmission. There is an averaging layer (Layer 3) between the second recurrent network layer (Layer 2) and the first fully connected layer (Layer 4), which is used to fuse the information in two directions of BiLSTM. In this embodiment, the fusion method used is to find the average of the two.

[0038] S4. Training S-RNN: Using training set X train Train S-RNN so that the error between the speed output by S-RNN and the corresponding input speed is Error train Less than the given error threshold Δ Error The specific training process will not be described here, and the general neural network training process is consistent.

[0039] In this embodiment, the calculation error can be measured by relative error or absolute error. The calculation formula of absolute error is:

[0040]

[0041] The calculation formula of relative error is:

[0042]

[0043] Among them, s t is the true value of the speed input to S-RNN at time t, is the speed prediction value corresponding to the vibration signal value input to S-RNN at time t;

[0044] In this embodiment, the purpose of S-RNN training is to optimize the parameters of the model. Before optimization, the following loss is minimized, so that the error between the extracted speed and the input speed can be minimized.

[0045]

[0046] in, The speed is predicted by S-RNN, S train is the true value of the rotation speed input to S-RNN; Θ is the model parameter to be optimized, which includes but is not limited to trainable parameters such as weight and bias, and hyperparameters such as the number of network layers, the number of neurons and the type of activation function, etc. Specifically, the back propagation algorithm, steepest descent method, momentum, Adam, etc. can be used to update the parameters;

[0047] S5. Test S-RNN: Use the test set X test Test the trained S-RNN. If the error Error is obtained in the test set test test Less than the given error threshold Δ Error , then go to the next step, otherwise, return to (3); in this embodiment, the purpose of the S-RNN test is to verify the speed extraction accuracy of the trained model for new data. If the accuracy meets the requirements, the model is qualified, otherwise the model hyperparameters need to be reselected.

[0048] S6. Online application: The vibration signal of the target rotating machinery is collected in real time, and after standardized processing, it is input into the S-RNN after the test is completed, so as to extract the corresponding instantaneous speed in real time.

[0049] In order to better illustrate the technical effect of the present invention, a specific embodiment is used to test and verify the present invention. The data used in the embodiment is a bearing state detection data set under variable speed conditions published by the Department of Mechanical Engineering of the University of Ottawa, Canada (http: / / dx.doi.org / 10.17632 / v43hmbwxpm.1). The data set is collected from the MFS-PK5M mechanical fault simulation test bench produced by SpectraQuest, and its structure is as follows: Figure 3 This data set includes vibration acceleration data and speed data of three bearing health states under four speed conditions, such as Figure 4 The three health states are: healthy, inner ring fault, and outer ring fault. The four speed conditions are: speed increase, speed decrease, speed increase first and then decrease, and speed decrease first and then increase.

[0050] In this example, we only use healthy data to train and test S-RNN. Figure 4An acceleration signal and acceleration signal samples in a healthy state. Figure 2 The 6-layer deep learning network S-RNN shown in the figure is used for speed extraction. The model parameters used are as follows. The number of neurons in each layer: 64-256-0-128-64-1, among which the average layer (Layer 3) has no neurons and the number is 0. The activation function of the output layer is a linear function, and the activation functions of other layers are ReLU functions. The model is trained using the Adam method. The absolute error is used as the accuracy measure of the model. The training accuracy and test accuracy of the model are as follows Figure 6 shown.

[0051] Assuming that the above test accuracy meets the requirements, the trained model can be put into use. Use the fault data to simulate the newly collected data. Input its vibration signal into the trained S-RNN and output the extracted instantaneous speed signal. Figure 7 The speed extraction results of data OA-1 are shown. It can be seen that the speed extraction method proposed in the present invention can better extract the instantaneous speed signal from the vibration data. For OA-1, the absolute error of speed extraction is 19.38rpm and the relative error is 1.35%.

[0052] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A method for extracting the instantaneous speed of a rotating machine from a vibration signal, characterized in that: The following steps are involved: (1) Collect the vibration signal sample set X∈R of the target rotating machinery N×T And the corresponding speed signal sample set S∈R N×T , where R represents the real number domain, N is the number of signal samples in the sample set, and T is the sampling time of a single signal sample; (2) The vibration signal sample set is normalized, and the speed signal sample set is normalized. Then, the processed vibration signal and speed signal are randomly divided into two non-repetitive subsets: a training set and a test set Among them, N train +N test =N; (3) Build a deep learning model S-RNN based on recurrent neural network, and then use a single vibration signal sample X i As input, the corresponding extracted speed signal time series is the output, i=1,2,L,N; (4) Training S-RNN: Using training set X train Train S-RNN so that the error between the speed output by S-RNN and the corresponding input speed is Error train Less than the given error threshold Δ Error ; (5) Test S-RNN: Use the test set X test Test the trained S-RNN. If the error Error is obtained in the test set test test Less than the given error threshold Δ Error , then go to the next step, otherwise, return to (3); (6) Online application: The vibration signal of the target rotating machinery is collected in real time, and after standardized processing, it is input into the S-RNN after the test is completed, so as to extract the corresponding instantaneous speed in real time.

2. A method for extracting instantaneous speed of a rotating machine from a vibration signal according to claim 1, characterized in that: The target rotating machines include, but are not limited to, motors, generators, gearboxes, pumps and fans.

3. The method for extracting the instantaneous speed of a rotating machine from a vibration signal according to claim 1, characterized in that: The rotation speed may be the rotation speed of any rotating shaft of the target rotating machine.

4. The method for extracting the instantaneous speed of a rotating machine from a vibration signal according to claim 1, characterized in that: The vibration signal includes but is not limited to an acceleration signal, a velocity signal, a displacement signal, a torque signal and a sound pressure signal.

5. The method for extracting the instantaneous speed of a rotating machine from a vibration signal according to claim 1, characterized in that: The method for standardizing the vibration signal sample set is: The mean of the vibration signal is reset to zero and the standard deviation is normalized to one. The specific formula is: in, and σ train The training set X train The mean and standard deviation of .

6. The method for extracting the instantaneous speed of a rotating machine from a vibration signal according to claim 1, characterized in that: The normalization processing method of the rotation speed signal sample set is: reducing or expanding the range of the rotation speed to a certain range.

7. The method for extracting the instantaneous speed of a rotating machine from a vibration signal according to claim 1, characterized in that: The S-RNN includes at least one recurrent network layer and several fully connected layers Dense Layer, wherein the information of the recurrent network layer is transmitted bidirectionally on the time axis, and the type of the recurrent network layer is a recurrent neural network RNN, a gated recurrent neural network GRU, a long short-term memory network LSTM, or one of the three network modifications.

Images