A highly versatile and multi-scenario intelligent rolling bearing fault diagnosis system

Abstract

A highly versatile and multi-scenario intelligent rolling bearing fault diagnosis system, including a wireless vibration sensor, an industrial edge computing gateway, and a device health management cloud platform; the vibration signals of the rolling bearing are collected by the wireless vibration sensor and wirelessly transmitted to the industrial edge computing gateway; the data format of the vibration signals is converted by the industrial edge computing gateway and transmitted to the device health management cloud platform; through the intelligent analysis of the HSIM module of the device health management cloud platform, the final classification prediction value of the vibration signals is given; the HSIM module includes: a one-dimensional convolutional neural network automatically extracts signal features, a fully connected classifier and an SVM classifier are combined in parallel with a probability threshold decision to classify samples, the classification result is determined through a decision function, and a simple diagnostic report or a fault category suggestion is generated according to the classification result for secondary manual diagnosis; the present invention has the advantages of strong versatility and high recall rate.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing fault diagnosis, and in particular to a rolling bearing fault diagnosis system with high versatility and multi-scenario intelligence (High Scalable Intelligent Method, HSIM). Background Art

[0002] Rolling bearings are the most widely used mechanical components in electromechanical equipment, and are also one of the most vulnerable components. Data show that 70% of the faults in rotating machinery are caused by rolling bearings, and among all kinds of gearbox faults, bearing failures are second only to gears and account for 19%, and 80% of motor failures are motor bearing failures; and the failure of rolling bearings will inevitably lead to abnormal operation of mechanical devices, and even cause catastrophic consequences, so it is very important to study the common failures of rolling bearings.

[0003] At present, most of the various rolling bearing testing and diagnosis instruments and systems developed and produced at home and abroad are made based on the principle of vibration analysis. Vibration analysis is to monitor the bearing vibration signal through a vibration sensor installed in the appropriate position of the bearing seat or box, and analyze and process the signal to judge the bearing working condition and fault. More than 80% of the literature on bearing monitoring and diagnosis discusses vibration analysis. From the perspective of applicability, practicality and effectiveness, there is currently no better rolling bearing monitoring and diagnosis method than vibration analysis.

[0004] Traditional vibration analysis methods require manual extraction of vibration signal features, and rely on professional spectrum analysis instruments and professional vibration analysis engineers to make judgments based on experience. This is costly and requires vibration analysts to have a high level of professional knowledge and experience. Currently, rolling bearing fault analysis and prediction in the industry require the cooperation of professional vibration analysis engineers and high-precision spectrum analysis instruments to accurately diagnose faults on-site equipment. The instruments are costly, have a low level of intelligence, and require vibration analysts to have a high level of professional knowledge and experience. Since the number of rotating equipment in an enterprise is often large, a large team of vibration analysts is required to complete the work of rolling bearing fault diagnosis and prediction. Since the service life of bearings is relatively long, faulty signals often only account for a small part of all signals in the signals to be detected. As a result, vibration analysts need to spend a lot of time on the signals of normal equipment, resulting in a large waste of manpower and time costs.

[0005] In recent years, machine learning and deep learning knowledge have been gradually used in rolling bearing fault diagnosis. Methods such as one-dimensional convolutional network (1DCNN) and wide convolutional network (WDCNN) are widely used. However, due to the characteristics of industrial data, such as small data volume and data imbalance, the machine learning model is seriously overfitted and the prediction accuracy is poor. Due to the requirements of the neural network model on data distribution, the fault prediction of the mechanical system generally has one machine and one model. Even the sampling rate of the vibration signal of the same equipment under the same working condition is different, and the judgment model is also different. Therefore, each equipment needs to train a model separately, which is time-consuming and labor-intensive. In industrial practice, it is impractical to customize a diagnostic device for each equipment to integrate the corresponding model, so the technology is difficult to implement. In addition, the machine learning classification model focuses on judging the fault category of the rolling bearing. In the industrial practice of rolling bearing fault diagnosis, the engineers are not very concerned about the category of rolling bearing faults (as long as there is a fault, the equipment needs to be repaired). Therefore, there is no need to classify the fault very accurately, but to accurately detect whether the equipment has a fault, that is, the recall rate of the equipment fault sample must reach 100% to avoid industrial accidents. Summary of the invention

[0006] In order to solve the problems existing in the diagnosis of industrial rolling bearings at this stage and overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a highly versatile, multi-scenario intelligent (HSIM) rolling bearing fault diagnosis system with the advantages of strong versatility and high recall rate.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A highly versatile, multi-scenario intelligent rolling bearing fault diagnosis system includes a wireless vibration sensor, an industrial edge computing gateway and an equipment health management cloud platform; the vibration signal of the rolling bearing is collected by the wireless vibration sensor, and the vibration signal is wirelessly transmitted to the industrial edge computing gateway; the data format of the vibration signal is converted by the industrial edge computing gateway and transmitted to the equipment health management cloud platform; the final classification prediction value of the vibration signal is given through the intelligent analysis of the HSIM module of the equipment health management cloud platform.

[0009] The HSIM module comprises:

[0010] 1) The received vibration signal X is subjected to preprocessing operations such as standardization and data enhancement to obtain a sample set x with a sample number of N;

[0011] 2) Input the sample set x into the classification module pre-integrated in the HSIM module. The processing steps of the classification module are:

[0012] 2.1) First, the signal features of the sample set x are automatically extracted through a one-dimensional convolutional neural network;

[0013] 2.2) The output of the one-dimensional convolutional network is passed through a fully connected classifier and an SVM classifier, and threshold decisions are made on the output of the softmax layer. The decision method of the threshold decision layer is to increase the probability threshold of the normal signal to improve the recall rate of the fault sample, and classify the samples in parallel to obtain 2N classification prediction values. The threshold decision method is as follows:

[0014]

[0015] x is the sample set, is the output value of the t-th sample in the y-th neuron of the softmax layer, y∈Y,Y={0,1,2,3}; h t (x) is the output category prediction value of the t-th sample; 0 is a normal sample, non-0 is a faulty sample, where 1 is a rolling element fault, 2 is an inner race fault, and 3 is an outer race fault; S threshold is the probability threshold. When the predicted category is normal, the probability value of the softmax predicted normal category must be greater than the threshold;

[0016] 2.3) Determine the classification result through the decision function;

[0017] The decision function is:

[0018]

[0019] x is the sample set; Y is the sample classification set; y is the classification prediction value of a single sample; N is the number of samples in the sample set; h n (x) is the output category prediction value of the n / 2th sample, when n%2=0, it is the prediction value of the fully connected classifier, and when n%2=1, it is the prediction value of the SVM classifier; H(X) is the final classification prediction value of the sample set.

[0020] The HSIM module issues a simple diagnosis report for vibration signals classified as normal, and gives a suggested fault category for vibration signals classified as faults for reference in secondary manual diagnosis.

[0021] The beneficial effects of the present invention are:

[0022] 1) The present invention effectively combines traditional manual detection technology and machine learning intelligent detection technology, taking advantage of each other and complementing their shortcomings;

[0023] 2) The present invention focuses on assisting vibration analysts in completing the screening of normal signals and faulty signals, screening out faulty signals with a 100% recall rate, while excluding as many normal signals as possible, saving the time and energy of vibration analysts, allowing personnel to only focus on faulty signals for further analysis and verification, greatly improving the work efficiency of personnel and reducing costs and increasing efficiency for enterprises.

[0024] 3) The present invention automatically extracts equipment vibration signals in real time through wireless vibration sensors, and gives real-time diagnostic suggestions through intelligent analysis of wireless signal transmission and equipment health management cloud platform.

[0025] 4) The present invention automatically extracts features from vibration signals through a one-dimensional convolutional neural network, reduces the screening and exclusion of normal signals, and provides a simple diagnostic report; it also provides a recommended fault category and manually performs a secondary judgment on the fault signal; it effectively reduces the work intensity and difficulty of vibration analysts by more than 80%, saving costs.

[0026] 5) The present invention improves the existing machine learning model, and the threshold decision and multi-classifier parallel voting decision improve the recall rate of fault samples under different working conditions, ensures that the fault samples can be effectively diagnosed, solves the one-machine-one-model problem, and improves the versatility of the model; BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the structure of the present invention.

[0028] Figure 2 This is the flow chart of the HSIM module.

[0029] Figure 3 Comparison of fault sample recall rates between the traditional method and the present invention (dataset B).

[0030] Figure 4 Comparison of the prediction accuracy results of the traditional method and the present invention (dataset B).

[0031] Figure 5 is the recall rate of normal samples and fault samples of the present invention (dataset C).

[0032] Figure 6 Comparison of the prediction accuracy results of the traditional method and the present invention (data set C). DETAILED DESCRIPTION

[0033] The present invention is further described in detail below in conjunction with embodiments and drawings.

[0034] like Figure 1As shown, a highly versatile, multi-scenario intelligent rolling bearing fault diagnosis system includes a wireless vibration sensor, an industrial edge computing gateway and an equipment health management cloud platform; the vibration signal of the rolling bearing is collected by the wireless vibration sensor, and the vibration signal is wirelessly transmitted to the industrial edge computing gateway; the data format of the vibration signal is converted by the industrial edge computing gateway and transmitted to the equipment health management cloud platform; the final classification prediction value of the vibration signal is given through the intelligent analysis of the HSIM module of the equipment health management cloud platform.

[0035] like Figure 2 As shown, the HSIM module includes:

[0036] 1) The received vibration signal X is subjected to preprocessing operations such as standardization and data enhancement to obtain a sample set x with a sample number of N;

[0037] 2) Input the sample set x into the classification module pre-integrated in the HSIM module. The processing steps of the classification module are:

[0038] 2.1) First, the signal features of the sample set x are automatically extracted through a one-dimensional convolutional neural network;

[0039] 2.2) The output of the one-dimensional convolutional network is passed through a fully connected classifier and an SVM classifier, and threshold decisions are made on the output of the softmax layer. The decision method of the threshold decision layer is to increase the probability threshold of the normal signal to improve the recall rate of the fault sample, and classify the samples in parallel to obtain 2N classification prediction values. The threshold decision method is as follows:

[0040]

[0041] x is the sample set, is the output value of the t-th sample in the y-th neuron of the softmax layer, y∈Y,Y={0,1,2,3}; h t (x) is the output category prediction value of the t-th sample; 0 is a normal sample, non-0 is a faulty sample, where 1 is a rolling element fault, 2 is an inner race fault, and 3 is an outer race fault; S threshold is the probability threshold. When the predicted category is normal, the probability value of the softmax predicted normal category must be greater than the threshold;

[0042] 2.3) Determine the classification result through the decision function;

[0043] The decision function is:

[0044]

[0045] x is the sample set; Y is the sample classification set; y is the classification prediction value of a single sample; N is the number of samples in the sample set; h n(x) is the output class prediction value of the n / 2-th sample (the prediction value of the fully connected classifier when n%2 = 0, and the prediction value of the SVM classifier when m%2 = 1); H(X) is the final classification prediction value of the sample set.

[0046] The HSIM module issues a simple diagnostic report for the vibration signals classified as normal, as shown in Table 1; and gives the recommended fault categories for the vibration signals classified as faulty, as shown in Table 2, for reference in secondary manual diagnosis.

[0047] Table 1 Example of Diagnostic Report for Normal Signals

[0048] Device Name Belt conveyor Device Status normal sensor Motor end_NB_20000565 state Running temperature 33 Power normal Collection time 2021-7-13 12:00:32 Real-time data Check Historical data Check

[0049] Table 2 Example of Diagnostic Results for Fault Signals

[0050]

[0051]

[0052] The present invention is verified through the following embodiments:

[0053] a) Dataset Introduction: To verify the generality and fault sample recall rate of the present invention under the same working conditions and different working conditions, this embodiment involves three datasets A, B, and C:

[0054] Dataset A comes from the Bearing Data Center of Case Western Reserve University in the United States. The vibration samples in this dataset are collected from the rolling bearings of the motor drive shaft, and the sampling frequency is 12 kHz. The fault categories include normal and rolling element faults (0.007 / 0.014 / 0.021 inch), inner race faults (diameter 0.007 / 0.014 / 0.021 inch), and outer race faults (6 o'clock direction, diameter 0.007 / 0.014 / 0.021 inch) in four normal states. The bearing data in each state is collected under different loads (0HP, 1HP, 2HP, 3HP). Dataset A is used for the training of the HSIM module.

[0055] Dataset B also comes from the Bearing Data Center of Case Western Reserve University in the United States. The vibration samples in this dataset are collected from the rolling bearings of the motor drive shaft, and the sampling frequency is 12 kHz. The fault categories include normal and rolling element faults (0.028 inch), inner race faults (0.028 inch), and outer race faults (3 / 12 o'clock direction, 0.007 / 0.021 inch).

[0056] Dataset C comes from the publicly available data of the bearing data competition. Dataset C is a real bearing vibration signal dataset, and there are 10 working states, as shown in Table 3.

[0057] Table 3, Fault Categories of Bearings

[0058]

[0059] During prediction evaluation, categories 3, 6, and 9 are mapped to label 1; categories 2, 5, and 8 are mapped to label 2; categories 1, 4, and 7 are mapped to label 3; and category 0 is mapped to label 0.

[0060] b) Classification model training and testing:

[0061] Since the bearing vibration characteristics are obvious and the dimension is low. To prevent overfitting, this embodiment uses a small number of convolution kernels and a small number of convolution layers. In order to fully extract data features and ensure the receptive field of the last layer of neurons, the convolution kernel size and the number of convolution layers cannot be too small. The present invention uses data set A to train multiple classification models for HSIM module evaluation and universality testing. The recall rate in the test set reaches 100%, and the prediction accuracy is over 98%;

[0062] c) HSIM module evaluation and universality test:

[0063] The parameters of the classification model trained in b) are fixed. In this embodiment, the trained classification model is used to predict the datasets B and C to evaluate the universality of the HSIM module.

[0064] Test results of data set B:

[0065] This embodiment aims to improve the prediction and classification accuracy while ensuring the recall rate of fault samples. Figure 3 It can be seen that the recall rate of fault samples has been significantly improved. The recall rate of fault samples in the classification models of 2 to 5 convolutional networks has reached 100%. The 6-layer convolutional network has overfitting and poor fault tolerance due to the high-dimensional features extracted. Therefore, the recall rate of fault samples after improvement is still slightly lower than 100%, but it is still significantly higher than the recall rate of traditional classification models.

[0066] The 28 vibration signals were preprocessed to form 28 sample sets, each with 100 samples. Figure 4 It can be seen that compared with traditional prediction methods, the present invention has significantly improved the prediction accuracy and can provide more effective suggestions for secondary manual diagnosis.

[0067] Test results of data set C:

[0068] Figure 5 and Figure 6 The recall and precision results of dataset C are given by using the classification model trained with dataset A. The sampling rate and bearing speed of dataset C are unknown, and the data show that the number of cycle points is less than 400. Figure 5It can be seen that the present invention has perfect versatility and can separate normal samples from faulty samples with 100% accuracy—the recall rates of normal samples and faulty samples are both 100%. Figure 6 It can be seen that the accuracy of the prediction results of the present invention is also improved to varying degrees.

[0069] The above-described embodiments of the present invention do not constitute a limitation on the protection scope of the present invention. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A highly versatile, multi-scenario intelligent rolling bearing fault diagnosis system, Features: It includes wireless vibration sensors, industrial edge computing gateways and equipment health management cloud platforms. The vibration signals of rolling bearings are collected by wireless vibration sensors and wirelessly transmitted to industrial edge computing gateways. The data format of vibration signals is converted by industrial edge computing gateways and transmitted to equipment health management cloud platforms. Through intelligent analysis of the HSIM module of the equipment health management cloud platform, the final classification prediction value of the vibration signal is given; The HSIM module comprises: 1) The received vibration signal X is subjected to a preprocessing operation of standardization and data enhancement to obtain a sample set x with a sample number of N; 2) Input the sample set x into the classification module pre-integrated in the HSIM module. The processing steps of the classification module are: 2.1) First, the signal features of the sample set x are automatically extracted through a one-dimensional convolutional neural network; 2.2) The output of the one-dimensional convolutional network is passed through a fully connected classifier and an SVM classifier, and threshold decisions are made on the output of the softmax layer. The decision method of the threshold decision layer is to increase the probability threshold of the normal signal to improve the recall rate of the fault sample, and classify the samples in parallel to obtain 2N classification prediction values. The threshold decision method is as follows: x is the sample set, is the output value of the t-th sample in the y-th neuron of the softmax layer, y∈Y,Y={0,1,2,3}; h t (x) is the output category prediction value of the t-th sample; 0 is a normal sample, non-0 is a faulty sample, where 1 is a rolling element fault, 2 is an inner race fault, and 3 is an outer race fault; S threshold is the probability threshold. When the predicted category is normal, the probability value of the softmax predicted normal category must be greater than the threshold; 2.3) Determine the classification result through the decision function; The decision function is: x is the sample set; Y is the sample classification set; y is the classification prediction value of a single sample; N is the number of samples in the sample set; h n (x) is the output category prediction value of the n / 2th sample, when n%2=0, it is the prediction value of the fully connected classifier, and when n%2=1, it is the prediction value of the SVM classifier; H(X) is the final classification prediction value of the sample set; The HSIM module issues a simple diagnosis report for vibration signals classified as normal, and gives a suggested fault category for vibration signals classified as faults for reference in secondary manual diagnosis.

Images