Vibration fault diagnosis algorithm

Abstract

The invention discloses a vibration fault diagnosis algorithm, which relates to the technical field of fault diagnosis, and comprises the following steps: S1, data acquisition: acquiring laboratory fault data and public fault data, and acquiring field equipment data by using a vibration acceleration sensor; s2, data preprocessing: carrying out mean value removal processing on the acquired data and the collected data, then obtaining spectrum data fftresult by using fft, and carrying out frequency normalization processing and amplitude normalization processing on the spectrum data; according to the method, the frequency domain data is used for analysis and calculation, the frequency and amplitude of the frequency domain data are subjected to data normalization processing, so that the data under different working conditions can be fused together for training and reasoning, the generalization ability of a neural network model is improved, meanwhile, a regression mode instead of a common classification mode is adopted, and the accuracy of the neural network model is improved. And the reasoning speed is greatly improved.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis, and in particular to a vibration fault diagnosis algorithm. Background Art

[0002] Rotating equipment and reciprocating equipment are widely used in industry. Their key components are prone to various problems in long-term operation due to various situations, leading to failures or even catastrophic accidents. The existing neural network-based fault diagnosis algorithms have the following shortcomings:

[0003] 1. The current algorithm (WDCNN) targets time domain data, which is susceptible to collector noise. The differences in working conditions are difficult to standardize and unify. The generalization ability of the neural network model based on the benchmark is weak. It is necessary to train the model separately for different working conditions, which is not suitable for industrial application.

[0004] 2. The algorithms in the prior art usually use the 1vs 1 or 1vs reset classification method to classify fault categories when performing fault category identification. This has high computational complexity and is difficult to give the probability of identifying a category, and is unable to provide more analysis data for subsequent analysis systems.

[0005] Based on the above problems, in order to improve the generalization ability of the neural network model and significantly increase the reasoning speed, a vibration fault diagnosis algorithm is proposed. Summary of the invention

[0006] The purpose of the present invention is to solve the problems in the prior art and to propose a vibration fault diagnosis algorithm.

[0007] A vibration fault diagnosis algorithm comprises the following steps:

[0008] S1. Data acquisition: Obtain laboratory fault data and public fault data, and use vibration acceleration sensors to collect field equipment data;

[0009] S2, data preprocessing: remove the mean value of the acquired data and the collected data, then use fft to obtain the spectrum data fft_result, and perform frequency normalization and amplitude normalization on the spectrum data;

[0010] S3, neural network training: data analysis is performed manually, and the corresponding data points and fault types are labeled for the fault data to form label data, and then the neural network model is trained. The output value is compared with the label data through loss function calculation, and the loss value is calculated for reverse calculation in the neural network to finally obtain the neural network model;

[0011] S4. Neural network reasoning: Use the neural network model to reason about the field equipment data to infer whether a fault has occurred, the fault type and the fault probability data.

[0012] Preferably, in step S1, the sampling frequency of the collected field device data is not less than 32 kHz, and the duration of each sampling is not less than 1 second.

[0013] Preferably, in step S2, the horizontal axis of the spectrum data is the frequency data freq, and the frequency normalization process is to actively discard the frequency data of the negative half axis, only use the frequency data of the positive half axis, and use the proportion based on the power frequency as the horizontal axis data of the normalized data:

[0014] freq_ratio=fft_result_freq / operating_freq

[0015] After the above processing, the horizontal axis of the spectrum data becomes freq_ratio;

[0016] The horizontal axis is sampled using the Largest Triangle ThreeBucket algorithm, and the sampling rules are as follows:

[0017] (1) For the horizontal coordinate [0.0, 1.0), the number of sampling points is S0 = [64, 256];

[0018] (2) For the horizontal coordinate of [1.0, +∞], the number of sampling points is S1=min(4096, count(freq_ratio))-S0.

[0019] Preferably, in step S2, the vertical axis of the spectrum data is the amplitude data fft_amplitude corresponding to the frequency, and the amplitude normalization process adopts energy-based normalization process. Since the vertical axis data of the spectrum data is a complex number, we first obtain the absolute value of the data as the value for subsequent processing:

[0020] amplitude i =|fft_amplitude i |;

[0021] Take the square of all frequency amplitudes to form the energy value of each frequency component:

[0022]

[0023] All energy values ​​are added together to get the total energy value:

[0024]

[0025] Then find the proportion of each frequency component to the total energy value:

[0026] normalized_spectrum i =amplitude i / energy.

[0027] Preferably, in step S3, the loss function adopts weighted cosine similarity:

[0028] (1) Calculate the standard cosine similarity:

[0029]

[0030] (2) Assign weights to annotation points:

[0031]

[0032] The weight of the power frequency point is set to the maximum, and the weights of other multiples and divisions are reduced in turn;

[0033] (3) Calculate the total loss:

[0034]

[0035] Preferably, in step S2, the data preprocessing of the field equipment also includes data cleaning, noise removal and filtering.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] The present invention uses frequency domain data for analysis and calculation, and adopts data normalization processing for both the frequency and amplitude of the frequency domain data, so that data of different working conditions can be integrated for training and reasoning, thereby improving the generalization ability of the neural network model. At the same time, regression is adopted instead of the commonly used classification method, which greatly improves the reasoning speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the process structure of the present invention.

[0039] Figure 2 This is a diagram of the neural network architecture of the present invention. DETAILED DESCRIPTION

[0040] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0041] Reference Figure 1-2 As shown, a vibration fault diagnosis algorithm includes the following steps:

[0042] S1. Data acquisition: Obtain laboratory fault data and public fault data, and use vibration acceleration sensors to collect field equipment data;

[0043] S2, data preprocessing: remove the mean value of the acquired data and the collected data, then use fft to obtain the spectrum data fft_result, and perform frequency normalization and amplitude normalization on the spectrum data. The purpose of data preprocessing is to obtain unified data at different speeds and different working conditions, so that different data can be compared when the neural network is trained and inferred;

[0044] S3, neural network training: data analysis is performed manually, and the corresponding data points and fault types are labeled for the fault data to form label data, and then the neural network model is trained. The output value is compared with the label data through loss function calculation, and the loss value is calculated for reverse calculation in the neural network to finally obtain the neural network model;

[0045] S4. Neural network reasoning: Use the neural network model to reason about the field equipment data to infer whether a fault has occurred, the fault type and the fault probability data.

[0046] In step S1, the sampling frequency of the collected field equipment data is not less than 32kHz, and the duration of each sampling is not less than 1 second.

[0047] In step S2, the horizontal axis of the spectrum data is the frequency data freq, and the frequency normalization process is to actively discard the frequency data of the negative half axis and only use the frequency data of the positive half axis (the spectrum data at this time is not comparable for different speeds. In order to be able to compare the spectrum data at different speeds, the frequency itself needs to be normalized), and the proportion based on the power frequency is used as the horizontal axis data of the normalized data:

[0048] freq_ratio=fft_result_freq / operating_freq

[0049] After the above processing, the horizontal axis of the spectrum data becomes freq_ratio;

[0050] The horizontal axis is sampled using the Largest Triangle ThreeBucket algorithm, and the sampling rules are as follows:

[0051] (1) For the horizontal coordinate [0.0, 1.0), the number of sampling points is S0 = [64, 256];

[0052] (2) For the horizontal coordinate of [1.0, +∞], the number of sampling points is S1=min(4096, count(freq_ratio))-S0.

[0053] In step S2, the vertical axis of the spectrum data is the amplitude data fft_amplitude corresponding to the frequency, and the amplitude normalization process adopts the energy-based normalization process. Since the vertical axis data of the spectrum data is a complex number, we first obtain the absolute value of the data as the value for subsequent processing:

[0054] amplitude i =|fft_amplitude i |;

[0055] Take the square of all frequency amplitudes to form the energy value of each frequency component:

[0056]

[0057] All energy values ​​are added together to get the total energy value:

[0058]

[0059] Then find the proportion of each frequency component to the total energy value:

[0060] normalized_spectrum i =amplitude i / energy.

[0061] In step S3, the loss function uses weighted cosine similarity:

[0062] (1) Calculate the standard cosine similarity:

[0063]

[0064] (2) Assign weights to annotation points:

[0065]

[0066] The weight of the power frequency point is set to the maximum, and the weights of other multiples and divisions are reduced in turn;

[0067] (3) Calculate the total loss:

[0068]

[0069] In step S2, the data preprocessing of the field equipment also includes data cleaning, noise removal and filtering.

[0070] like Figure 2As shown, the neural network consists of convolution layer 1, convolution layer 2, deconvolution layer 3, convolution layer 4, concatenation layer 5, convolution layer 6, convolution layer 7, fully connected layer 8 and fully connected layer 9. Convolution layer 6 and convolution layer 7 are followed by a maximum pooling layer. Each of the above convolution layers (including deconvolution layers) is followed by a batch normal layer and a LeakyRelu (0.1) layer to correct and activate the data. Both fully connected layer 8 and fully connected layer 9 have a LeakyRelu (0.1) activation layer.

[0071] Example

[0072] Step 1: Sensor Installation and Connection

[0073] Install acceleration vibration sensors at key parts of the equipment (such as bearings, gearboxes, etc.). The sensors should be as close as possible to the monitored equipment to obtain data with sufficient signal-to-noise ratio. Connect the sensors to the data acquisition module, ensure that the output signal of the sensor matches the input interface of the data acquisition module, and connect them through appropriate cables.

[0074] Step 2: Data Collection

[0075] Start the data acquisition module and begin to receive data from the sensor in real time to ensure the continuity and accuracy of data acquisition. Transfer the collected data to the data storage module and use appropriate communication protocols and data transmission lines to stably transfer the data to the data storage module to ensure the integrity and traceability of the data.

[0076] Step 3: Data processing and analysis

[0077] 1. Data preprocessing

[0078] The collected raw data are preprocessed, including data cleaning, denoising and filtering, to ensure the quality and reliability of the data.

[0079] 2. Data processing

[0080] In the training phase, data analysis is first performed manually, and corresponding data points and fault types are annotated for laboratory fault data and public fault data. Then, the neural network model is trained to obtain the neural network model.

[0081] In the inference stage, after preprocessing, the field equipment data is inferred through a neural network to obtain data such as whether a fault has occurred, the type of fault, and the probability of fault.

[0082] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A vibration fault diagnosis algorithm, characterized in that: The following steps are involved: S1. Data acquisition: Obtain laboratory fault data and public fault data, and use vibration acceleration sensors to collect field equipment data; S2, data preprocessing: remove the mean value of the acquired data and the collected data, then use fft to obtain the spectrum data fft_result, and perform frequency normalization and amplitude normalization on the spectrum data; S3, neural network training: data analysis is performed manually, and the corresponding data points and fault types are labeled for the fault data to form label data, and then the neural network model is trained. The output value is compared with the label data through loss function calculation, and the loss value is calculated for reverse calculation in the neural network to finally obtain the neural network model; S4. Neural network reasoning: Use the neural network model to reason about the field equipment data to infer whether a fault has occurred, the fault type and the fault probability data.

2. A vibration fault diagnosis algorithm according to claim 1, characterized in that: In the step S1, the sampling frequency of the collected field device data is not less than 32 kHz, and the duration of each sampling is not less than 1 second.

3. A vibration fault diagnosis algorithm according to claim 1, characterized in that: In step S2, the horizontal axis of the spectrum data is the frequency data freq, and the frequency normalization process is to actively discard the frequency data of the negative half axis, only use the frequency data of the positive half axis, and use the proportion based on the power frequency as the horizontal axis data of the normalized data: freq_ratio=fft_result_freq / operating_freq After the above processing, the horizontal axis of the spectrum data becomes freq_ratio; The horizontal axis is sampled using the Largest Triangle Three Buckets algorithm, and the sampling rules are as follows: (1) For the horizontal coordinate [0.0, 1.0), the number of sampling points is S0 = [64, 256]; (2) For the horizontal coordinate of [1.0, +∞], the number of sampling points is S1=min(4096, count(Freq_ratio))-S0.

4. A vibration fault diagnosis algorithm according to claim 1, characterized in that: In step S2, the vertical axis of the spectrum data is the amplitude data fft_amplitude corresponding to the frequency, and the amplitude normalization process adopts the energy-based normalization process. Since the vertical axis data of the spectrum data is a complex number, we first obtain the absolute value of the data as the value for subsequent processing: amplitude i =|fft_amplitude i |; Take the square of all frequency amplitudes to form the energy value of each frequency component: All energy values ​​are added together to get the total energy value: Then find the proportion of each frequency component to the total energy value: normalized_spectrum i =amplitude i / energy。 5. A vibration fault diagnosis algorithm according to claim 1, characterized in that: In step S3, the loss function uses weighted cosine similarity: (1) Calculate the standard cosine similarity: (2) Assign weights to annotation points: The weight of the power frequency point is set to the maximum, and the weights of other multiples and divisions are reduced in turn; (3) Calculate the total loss:

6. A vibration fault diagnosis algorithm according to claim 1, characterized in that: In step S2, the data preprocessing of the field equipment also includes data cleaning, noise removal and filtering.

Images