A bearing fault intelligent detection system and method based on inspection and vibration signals

Abstract

The present invention discloses an intelligent bearing fault detection system and method based on inspection and vibration signals, which is characterized by comprising the following steps: the staff of the background control center issues an inspection command to the inspection robot, the inspection robot starts to inspect the workshop equipment, collects images, temperature and sound signals, and transmits them to the background control center; then the data is cleaned, normalized, and incremented; then a deep neural network model is established, and the model is pre-trained, and then the field sampling bearing training set is retrained by the transfer learning method, and the trained optimal model is deployed as a fault detection model in the model deployment machine, and the image signal, sound signal, and temperature signal obtained by field collection are used to perform multi-source data fault detection, and a fault alarm is generated. The advantages are: it can detect and warn the bearing fault of the field equipment in time, and effectively maintain the normal operation of the mechanical equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment fault detection, and in particular relates to an intelligent bearing fault detection system and method based on inspection and vibration signals. Background Art

[0002] The grinding and magnetizing workshop is the main building of a magnetite beneficiation plant. Its main equipment includes ball mills, vertical mills, cyclones, slurry pumps, vertical ring strong magnets, and permanent magnet drum separators. These devices are complex in structure, and any failure can not only halt production but, in severe cases, damage the equipment and even threaten employee safety. Therefore, to ensure safe operation, equipment failures are typically diagnosed through manual inspections. Conventional methods include visual inspection, touch, listening, tapping, and smelling. With the advancement of information technology, some intelligent detection equipment has been incorporated into the inspection process. However, these devices suffer from the following significant shortcomings: 1) Limited sensing range: Due to limitations in the inspection environment and technical requirements, current intelligent detection equipment primarily consists of non-contact infrared thermal imaging temperature measurement and contact vibration detection, resulting in low detection efficiency and difficulties in ensuring employee safety. 2) Lack of intelligent analysis: Currently, after acquiring test data, effective intelligent judgment methods are lacking, relying primarily on manual judgment. Inspection results also rely heavily on the experience of inspectors, placing a heavy workload on inspectors and hindering further improvement in inspection efficiency and accuracy.

[0003] As the requirements for safe production and operation of equipment become increasingly higher, it is very necessary to research, develop and use a set of intelligent trackless robot systems and detection methods that can not only automatically collect multiple data on the equipment status, but also conduct comprehensive and integrated analysis of the obtained data. When extracting data features based on traditional machine learning algorithms, it is often difficult to manually extract signal features due to the different working conditions of different bearing equipment, and even the different vibration conditions of different sampling points of the same equipment, and the model generalization ability is poor. The fault detection model based on deep learning can automatically extract signal features, but requires a large amount of fault sampling data for model training. There are very few fault samples for bearing equipment that has been running for a long time in the field. The present invention pre-trains the model with the disclosed bearing vibration data set, which can solve the problem of bearing fault diagnosis under different working conditions in the field environment with a small sample size. At the same time, the present invention proposes a new abnormal data cleaning algorithm that can reduce noise interference or system false alarms, and improve the recognition accuracy of the bearing operating status through data preprocessing and data enhancement algorithms. Summary of the Invention

[0004] In response to the problem that existing bearing equipment lacks intelligent detection technology for online detection, the purpose of the present invention is to provide an intelligent bearing fault detection system and method based on inspection and vibration signals. The inspection is completed autonomously by using an intelligent inspection robot. By analyzing the characteristics of the vibration timing signal during the operation of the rolling bearing and combining multi-source signal acquisition and analysis technology, the status characteristics of the equipment under various operating conditions are excavated. The present invention can effectively detect the operating status of the bearing and provide timely warnings.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] The present invention provides an intelligent bearing fault detection system based on inspection and vibration signals, characterized in that the intelligent inspection system includes an inspection robot, a wireless receiver and a background control center. The inspection robot is composed of a power supply unit, an information collection unit, a control unit, a wireless network unit and a positioning and navigation unit; the background control center is composed of a data server, an algorithm server, a model deployment machine, a client computer and a network router; the wireless receiver is connected to the wireless network unit of the inspection robot and the network router of the background control center respectively;

[0007] The intelligent inspection system includes a plurality of laser vibration sensors arranged around the bearing equipment, which are used to collect vibration time series signal data of the equipment bearings;

[0008] The inspection robot is arranged in a workshop building. In the inspection robot, the power supply unit is electrically connected to the information collection unit, the control unit, the positioning and navigation unit, and the wireless network unit, respectively, for supplying power to the inspection robot during the inspection process;

[0009] The control unit is electrically connected to the information collection unit, the positioning and navigation unit, and the wireless network unit through data transmission cables. The control unit is used to control the automatic inspection of the inspection robot, the positioning and navigation unit is used to determine the position of the automatic inspection of the inspection robot, and the wireless network unit is connected to the wireless receiver through WIFI transmission technology.

[0010] The information collection unit uses a visible light camera, an infrared camera, and a microphone to collect on-site images, temperature, and sound signals, and transmits them to the background algorithm server;

[0011] The data server, algorithm server, model deployment machine and client computer of the background control center are all electrically connected to the network router through data transmission cables, the network router is electrically connected to the wireless receiver, and the background control center realizes information interaction with the inspection robot through the wireless receiver.

[0012] Preferably, the background control center is arranged in a control room.

[0013] The present invention provides an intelligent bearing fault detection method based on inspection and vibration signals, which adopts the above-mentioned intelligent inspection system and is characterized by comprising the following steps:

[0014] A. The staff of the backstage control center issues an inspection command to the inspection robot, which then begins to inspect the bearing equipment in the workshop, collects bearing images, temperature, and sound signals, and transmits them to the backstage control center;

[0015] B. The laser vibration sensor collects the bearing vibration timing signal and transmits it to the background control center;

[0016] C. Data cleaning

[0017] The data server cleans the collected bearing vibration time series signal data;

[0018] D. Data Preprocessing

[0019] The data server normalizes the bearing vibration time series signal data after data cleaning to obtain bearing vibration time series signal sample data; the data server normalizes the collected bearing sound signal to obtain a bearing sound signal sample data set;

[0020] E. Data Incremental Processing

[0021] A sliding rectangular window function is established in the data server and sliding data increment processing is performed on the normalized bearing vibration time series signal sample data to obtain a bearing vibration signal sample data set;

[0022] F. Build a deep neural network model for bearing fault detection in the algorithm server. Use the Case Western Reserve University rolling bearing fault dataset for model pre-training. Then, use transfer learning to retrain the bearing vibration signal sample dataset and the sound signal sample dataset to obtain the trained model.

[0023] G. Deploy the trained model as a fault detection model in the model deployment machine, use the bearing image signals, sound signals and temperature signals collected on site to perform multi-source data-assisted fault detection, and display the detection results on the client computer to generate a fault alarm.

[0024] Preferably, in step C, the formula used for data cleaning of the collected bearing vibration time series signal data is as follows:

[0025] Let X={X1,X2,...,X n} is the vibration signal collected from n sensors in T time series, X i ={x i1 ,x i2 ,...,xiT} T is the data collected by the i-th sensor. Let ΔX = {Δx1, Δx2..., Δx n-1}, where Δx i = {Δx i1 , Δx i2 ..., Δx iT}; and Δx T = |x i,t - x i+1,t |; if x i,t and x i+1,t are within the normal range, there exist a parameter L and a threshold σ such that L - σ ≤ Δ it ≤ L + σ, and the parameter L is the average value of Δx i ;

[0026] If |x i,t - x i+1,t | is not within the range of [L - σ, L + σ], it is necessary to determine which of x i,t or x i+1,t is the abnormal data. This method uses a series of adjacent-time data, x i,Δt , and calculates the median of x i,Δt If then x i,t is abnormal; if abnormal data is detected, then is used to replace the abnormal data.

[0027] Preferably, in step D, the formula used for normalizing the collected vibration time-series signal data and sound signal data is as follows:

[0028]

[0029] [[ID=S9]]

[0030] where, x p is the signal sample after normalization processing, x is the signal sample, μ is the signal sample mean, and σ is the signal sample standard deviation. Preferably, in step E, the process of establishing a sliding rectangular window function in the data server to perform sliding data increment processing on the bearing vibration time-series signal sample data is as follows: First, establish a sliding rectangular window function. The sliding rectangular window slides from the zero moment at a certain step length to intercept a vibration time-series signal sample sequence with the width of the sliding rectangular window. The length of the vibration time-series signal is L, the width of the sliding rectangular window is l, and the sliding step length of the sliding rectangular window is s, (s < l). The width of the sliding rectangular window is l = 2 a l < L, a = 0, 1, 2, 3…, then the sample size after sliding data increment processing is:

[0031] Preferably, in step F, the deep neural network model comprises a one-dimensional convolutional network branch and a multi-layer perceptron network branch. The one-dimensional convolutional network can extract local features of fault information, while the multi-layer perceptron is used to extract global features of the fault. The outputs of the two network branches are then fused through a connection layer and then passed through a fully connected network with the network output dimension being the number of fault categories n. Finally, the output is passed through a Softmax layer to obtain the probability value of each fault category, thereby determining the corresponding fault category.

[0032] The one-dimensional convolutional network branch includes multiple one-dimensional convolutional layers, maximum pooling layers, and fully connected layers, and the activation function of each layer is the ReLU function; to prevent gradient explosion or diffusion and speed up training, a BatchNormalization layer is introduced after each layer, the convolution kernel size of each convolution layer is 2 and the stride is 1, and the pooling size and stride of the maximum pooling layer are both 4; the multi-layer perceptron network branch includes a fully connected layer and a Batch Normalization layer, and the activation function of each layer is also the ReLU function;

[0033] The outputs of the two network branches are fused through a connection layer. The fusion process of the two network branches can splice the outputs of the two branches together to form a concatenate layer, or directly add the outputs of the two network branches to form an add layer. When choosing to directly add the outputs of the two network branches to form an add layer, it is necessary to ensure that the output dimensions of the two network branches are the same.

[0034] Preferably, in step F, during training, the cross entropy loss function is selected as the loss function of the deep neural network model, and the RMSProp is selected as the optimizer.

[0035] Preferably, in step G, the Case Western Reserve University rolling bearing fault dataset includes a deep groove ball bearing model SKF6205. The bearing operating states include normal, inner race fault, rolling element fault, and outer race fault. Each fault type includes three fault diameters: 0.1778 mm, 0.3556 mm, and 0.5334 mm, thus forming 10 operating states. The sampling frequency of the fault data is 12 kHz.

[0036] Compared with the prior art, the advantages of the present invention are:

[0037] 1) The present invention adopts an intelligent bearing fault detection system and method based on inspection and vibration signals, which enhances the detection flexibility and safety and is more sensitive to weak fault information that occurs early in the equipment bearing.

[0038] 2) The present invention comprehensively utilizes advanced technologies such as artificial intelligence and multi-source data analysis to realize the perception of equipment bearing vibration time series information, data analysis, and intelligent judgment of fault defects.

[0039] 3) The present invention can realize intelligent diagnosis and analysis of equipment failures, provide timely warnings without power outages, and promptly notify operation and maintenance personnel to quickly handle the problem, thereby improving equipment safety performance.

[0040] In summary, the present invention utilizes an intelligent inspection robot to autonomously complete equipment inspections, allowing operators to perform remote operations away from the equipment, which is safe and reliable. The detection method of the present invention can improve the predictability of equipment bearing failures through data analysis, provide effective data support for condition-based maintenance or condition assessment, and thus reduce the occurrence of accidents. Before equipment failure occurs, it can quickly issue an early warning and provide preliminary reference information for accident handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is an architectural diagram of the bearing fault intelligent detection system based on inspection and vibration signals of the present invention;

[0042] Figure 2 is a flow chart of the bearing fault intelligent detection method of the present invention;

[0043] Figure 3 It is a deep learning network structure diagram of the present invention;

[0044] Figure 4 It is a deep neural network structure of a specific embodiment. DETAILED DESCRIPTION

[0045] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0046] Example

[0047] like Figure 1 As shown, the present invention is an intelligent bearing fault detection system based on inspection and vibration signals, characterized in that the intelligent inspection system includes an inspection robot, a wireless receiver and a background control center. The inspection robot is composed of a power supply unit, an information collection unit, a control unit, a wireless network unit and a positioning and navigation unit; the background control center is composed of a data server, an algorithm server, a model deployment machine, a client computer and a network router; the wireless receiver is connected to the wireless network unit of the inspection robot and the network router of the background control center respectively;

[0048] The intelligent inspection system includes a plurality of laser vibration sensors arranged around the bearing equipment, which are used to collect the equipment bearing vibration time series signal data;

[0049] The inspection robot is arranged in a workshop building. In the inspection robot, the power supply unit is electrically connected to the information collection unit, the control unit, the positioning and navigation unit, and the wireless network unit, respectively, for supplying power to the inspection robot during the inspection process;

[0050] The control unit is electrically connected to the information collection unit, the positioning and navigation unit, and the wireless network unit through data transmission cables. The control unit is used to control the automatic inspection of the inspection robot, the positioning and navigation unit is used to determine the position of the automatic inspection of the inspection robot, and the wireless network unit is connected to the wireless receiver through WIFI transmission technology.

[0051] The information collection unit uses a visible light camera, an infrared camera, and a microphone to collect on-site images, temperature, and sound signals, and transmits them to the background control center;

[0052] The background control center is set up in the control room. In the background control center, the data server, algorithm server, model deployment machine and client computer are electrically connected to the network router through data transmission cables. The network router is electrically connected to the wireless receiver. The background control center realizes information interaction with the inspection robot through the wireless receiver.

[0053] The specific test preparation work is as follows: Select a location appropriate to the bearing device to be tested as the bearing vibration time series signal collection point. When the equipment is working normally, collect the bearing vibration signal data of sufficient equipment and transmit it to the data server for storage. When the equipment fails, collect the bearing vibration signal data of sufficient equipment and transmit it to the data server for storage. The inspection robot patrols the inspection point and collects equipment images, sounds, and temperature signals;

[0054] In this embodiment, rolling bearing vibration time series signals at 10 positions are collected, and data of approximately 244,000 time points are collected for each vibration time series signal.

[0055] like Figures 2 to 4 As shown, an intelligent detection method for bearing faults of the present invention adopts the above-mentioned intelligent inspection system and is characterized by comprising the following steps:

[0056] A. The staff of the backstage control center issues an inspection command to the inspection robot, which then begins to inspect the bearing equipment in the workshop, collects bearing images, temperature, and sound signals, and transmits them to the backstage control center;

[0057] B. The laser vibration sensor collects the bearing vibration timing signal and transmits it to the background control center;

[0058] C. Data cleaning

[0059] The data server cleans the collected bearing vibration time series signal data;

[0060] The formula used for data cleaning of the collected bearing vibration time series signal data is as follows:

[0061] Let X={X1,X2,...,X n} is the bearing vibration signal collected from n sensors in the time series T, n is 10, X i ={x i1 ,x i2 ,...,x iT} T The data collected by the i-th sensor is sorted by the sensor number according to the position of the sensor. The sensor numbers at adjacent positions are also adjacent. Therefore, in the vibration signal matrix, the column data is time-related and the row data is space-related.

[0062] Let ΔX={Δx1,Δx2...,Δx n-1}, where Δx i ={Δx i1 ,Δx i2 ...,Δx iT} T , and Δx i,t =|x i,t -x i+1,t |. If x i,t and x i+1,t In the normal range, there exists a parameter L and a threshold σ, such that L-σ≤Δ it ≤L+σ. Therefore, if Δx it If it is not within the range of [L-σ,L+σ], then Δx i,t or Δx i+1,t It may be abnormal data.

[0063] The setting of parameter L is critical, usually L is set to Δx i However, when there are abnormal data, Δx i The average value of x may be much higher than i and x i+1 , normal Δx i It may not be within the range of [L-σ,L+σ]. To avoid this problem, use Δx i The median is used instead of the mean because abnormal data may cause a large change in the mean without affecting the median. In this embodiment, L is 0.003 mm and σ is set to 0.02 mm.

[0064] If |x i,t -x i+1,t| is not within the range of [L-σ,L+σ], then it is necessary to judge x i,t or x i+1,t Which one is abnormal data. In the adjacent time period t' of t, |x i,t -x i,t’ If x is greater than a small threshold δ, i,t It may be abnormal data. However, if x i,t ' is also abnormal, then it may be impossible to determine the abnormality. To solve this problem, this method uses a series of adjacent time data, x i,Δt , and calculate x i,Δt the median of if Then x i,t If abnormal data is detected, use In this embodiment, is 0.25mm and δ is 0.05mm.

[0065] D. Data Preprocessing

[0066] The data server normalizes the bearing vibration time series signal data after data cleaning to obtain bearing vibration signal sample data; the data server normalizes the collected bearing sound signal to obtain a bearing sound signal sample data set;

[0067] The formula used for normalizing the collected bearing vibration time series signal data and sound signal data is as follows:

[0068]

[0069] Among them, x p is the normalized signal sample, x is the signal sample, μ is the mean value of the signal sample, and σ is the standard deviation of the signal sample. In this embodiment, μ is 0.25 mm and σ is 0.01 mm.

[0070] E. Data Incremental Processing

[0071] A sliding rectangular window function is established in the data server and sliding data increment processing is performed on the normalized bearing vibration signal sample data to obtain a bearing vibration signal sample data set;

[0072] The process of establishing a sliding rectangular window function in a data server and performing sliding data increment processing on a vibration signal is as follows: first, a sliding rectangular window function is established, wherein the sliding rectangular window slides from time zero according to a certain step length to intercept a vibration time series signal sample sequence of the sliding rectangular window width, wherein the length of the vibration time series signal is L, the width of the sliding rectangular window is l, the sliding step length of the sliding rectangular window is s, (s<l), and the width of the sliding rectangular window is l=2 aIf \(l < L\), where \(a = 0, 1, 2, 3,\cdots\), then the sample size after sliding data increment processing is as follows:

[0073] In this embodiment, the width of the sliding rectangular window is 1024 and the sliding step is 800. According to the sample size calculation formula, there are approximately 300 fault samples for each fault type, and the feature dimension of each sample is 1024.

[0074] F. Establish a deep neural network model for bearing fault detection in the algorithm server. Use the rolling bearing fault dataset of Case Western Reserve University for model pre-training, and then re-train the vibration signal sample dataset and sound signal sample dataset through transfer learning to obtain a training model.

[0075] The deep neural network mentioned above has a network structure including a one-dimensional convolutional network branch and a multi-layer perceptron network branch. The one-dimensional convolutional network can extract local features of fault information, and the multi-layer perceptron is used to extract global features of faults. Subsequently, the outputs of the two network branches are fused through a connection layer, and then passed through a fully connected network. The network output dimension is the number of fault categories \(n\). Finally, through the Softmax layer, the probability value of each fault category is obtained, and then the corresponding fault category is determined.

[0076] The deep neural network model for bearing fault detection constructed in this embodiment is as Figure 3 shown. The vibration time series signal sample data after data increment processing is respectively input into the one-dimensional convolutional network branch and the multi-layer perceptron network branch.

[0077] In the one-dimensional convolutional network branch, five one-dimensional convolutional layers and five max pooling layers are alternately connected, and then four fully connected layers are connected. The ReLU function is selected as the activation function for each layer. To prevent gradient explosion or dispersion and accelerate the training speed, a Batch Normalization layer is introduced after each layer. The convolutional kernel size of each convolutional layer is 2 and the stride is 1. The pooling size and stride of the max pooling layer are both 4. The number of convolutional kernels of the five one-dimensional convolutional layers increases layer by layer, which are 32, 64, 128, 256, and 512 in sequence. After passing through the convolutional layer, the data feature dimension changes from 1 to 512. The number of neurons in the 4 fully connected layers is 256, 128, 64, and 32 in sequence. The output dimension of the one-dimensional convolution branch is 32.

[0078] In the multi-layer perceptron network branch, there are a total of five fully connected layers, and a Batch Normalization layer is introduced after each layer. The activation function is the ReLU function. The number of neurons in each layer is 512, 256, 128, 64, and 32 in sequence. The output dimension of the multi-layer perceptron network branch is 32.

[0079] The outputs of the two network branches are concatenated through the Concatenate layer, and finally the probability values ​​of 10 types of faults are obtained through two fully connected layers and the Softmax layer. The fault category with the largest probability value is selected as the predicted label of the corresponding sample. In addition, the Dropout layer is added at the appropriate position to prevent the model from overtraining and overfitting problems. The loss rate is set to 0.5. In this embodiment, the training batch size is set to 128, the loss function adopts the cross entropy loss function, the optimizer adopts the Adam optimizer, and the learning rate is set to 10 -4 .

[0080] The cross-entropy loss function is selected as the loss function of the deep neural network model, and the RMSProp optimizer is selected; Dropout can be added at appropriate positions in the entire network to prevent the occurrence of overfitting problems.

[0081] The Case Western Reserve University rolling bearing fault dataset described in this paper uses the SKF6205 deep groove ball bearing model. The bearing operating states include normal, inner race fault, rolling element fault, and outer race fault. Each fault type includes three fault diameters: 0.1778 mm, 0.3556 mm, and 0.5334 mm, resulting in 10 operating states. The fault data is sampled at a frequency of 12 kHz.

[0082] G. Deploy the trained model as a fault detection model in the model deployment machine to perform equipment fault detection based on sound and vibration signals. At the same time, use the image signals, sound signals, and temperature signals collected on-site to perform multi-source data-assisted fault detection, and display the detection results on the client computer to generate a fault alarm.

[0083] The image signals, sound signals and temperature signals collected by the inspection robot are used as auxiliary early warnings for the temperature and status of the equipment bearings through manual identification, threshold judgment and other methods.

[0084] In short, the present invention is an intelligent bearing fault detection system and method based on inspection and vibration signals, and its general steps are as follows: first, the inspection robot starts to inspect the equipment in the workshop, collects images, temperature and sound signals, and transmits them to the background control center; the laser vibration sensor collects the vibration time series signal of the equipment bearing and transmits it to the data server; the data server cleans the collected vibration time series signal and normalizes the cleaned vibration time series signal data; then, a sliding rectangular window function is established in the data server to perform sliding data increment processing on the normalized vibration time series signal; the data server also normalizes the collected sound signal; the algorithm server establishes a deep neural network model, uses the Case Western Reserve University rolling bearing fault data set for model pre-training, and then retrains the vibration signal and sound signal training set of the on-site sampled bearing through the transfer learning method. The trained deep neural network model is deployed in the model deployment machine, and the image signal and temperature signal collected on-site are used to perform multi-source data fault detection, and the detection results are displayed on the client computer, and the fault alarm is issued.

Claims

1. An intelligent detection method for bearing faults based on inspection and vibration signals, using an intelligent detection system, the intelligent detection system includes an inspection robot, a wireless receiver and a background control center, the inspection robot is composed of a power supply unit, an information collection unit, a control unit, a wireless network unit and a positioning navigation unit; the background control center is composed of a data server, an algorithm server, a model deployment machine, a client computer and a network router; the wireless receiver is respectively connected to the wireless network unit of the inspection robot and the network router of the background control center; The intelligent detection system includes a laser vibration sensor arranged around the bearing equipment, which is used to collect vibration time series signal data of the equipment bearing; The inspection robot is arranged in a workshop building. In the inspection robot, the power supply unit is electrically connected to the information collection unit, the control unit, the positioning navigation unit and the wireless network unit respectively, and is used for supplying power to the inspection robot during the inspection process; The control unit is electrically connected to the information collection unit, the positioning and navigation unit, and the wireless network unit through data transmission cables. The control unit is used to control the automatic inspection of the inspection robot, the positioning and navigation unit is used to determine the position of the automatic inspection of the inspection robot, and the wireless network unit is connected to the wireless receiver through WIFI transmission technology. The information collection unit uses a visible light camera, an infrared camera, and a microphone to collect on-site images, temperature, and sound signals, and transmit them to the background control center; The data server, algorithm server, model deployment machine and client computer of the background control center are electrically connected to the network router through data transmission cables, the network router is electrically connected to the wireless receiver, and the background control center realizes information interaction with the inspection robot through the wireless receiver; it is characterized in that The following steps are involved: A. The staff of the backstage control center issues inspection instructions to the inspection robot, and the inspection robot begins to inspect the bearing equipment in the workshop, collects the bearing image, temperature and sound signal, and transmits them to the backstage control center; B. The laser vibration sensor collects the bearing vibration timing signal and transmits it to the background control center; C. Data cleaning The data server cleans the collected data of the bearing vibration time series signal; D. Data Preprocessing The data server normalizes the bearing vibration time series signal data after data cleaning to obtain bearing vibration time series signal sample data; The data server normalizes the collected bearing sound signals to obtain a bearing sound signal sample data set; E. Data Increment Processing A sliding rectangular window function is established in the data server and sliding data increment processing is performed on the normalized bearing vibration time series signal sample data to obtain a bearing vibration signal sample data set; F. Establish a deep neural network model for bearing fault detection in the algorithm server, use the Case Western Reserve University rolling bearing fault dataset for model pre-training, and then retrain the bearing vibration signal sample dataset and sound signal sample dataset through transfer learning to obtain a training model; G. Deploy the training model as a fault detection model in the model deployment machine, use the bearing image signal, sound signal, temperature signal and vibration signal collected on site to perform multi-source data fault detection, and display the detection results on the client computer to generate a fault alarm.

2. The bearing fault intelligent detection method based on inspection and vibration signals according to claim 1 is characterized in that: In step C, the formula used for data cleaning of the collected bearing vibration time series signal data is as follows: Let X = {X1, X2, ..., X n } is the vibration signal collected from n sensors in T time series, X i ={x i1 ,x i2 ,...,x iT } T is the data collected by the i-th sensor, let ΔX = {Δx1, Δx2..., Δx n-1 }, where Δx i ={Δx i1 ,Δx i2 ...,Δx iT } T , and Δx i,t =|x i,t -x i+1,t |; if x i,t and x i+1,t In the normal range, there exists a parameter L and a threshold σ1 such that L-σ1≤Δ it ≤L+σ1, parameter L is Δx i The average of If |x i,t -x i+1,t | is not within the range of [L-σ1,L+σ1], then it is necessary to judge x i,t or x i+1,t Which one is the abnormal data? This method uses a series of adjacent time data, x i,Δt , and calculate x i,Δt the median of if δ is the threshold of the median difference, then x i,t is abnormal; if abnormal data is detected, Replace abnormal data.

3. The bearing fault intelligent detection method based on inspection and vibration signals according to claim 1 is characterized in that: In step D, the formula used for normalizing the collected bearing vibration time series signal data and sound signal data is as follows: Among them, x p is the signal sample after normalization, x is the signal sample, μ is the mean of the signal sample, and σ2 is the standard deviation of the signal sample.

4. The bearing fault intelligent detection method based on inspection and vibration signals according to claim 1 is characterized in that: In step E, the process of establishing a sliding rectangular window function in the data server to perform sliding data increment processing on the normalized bearing vibration time series signal sample data is as follows: First, establish a sliding rectangular window function. The sliding rectangular window slides from the zero moment at a certain step length to intercept the vibration time series signal sample sequence with the width of the sliding rectangular window. The length of the vibration time series signal is L, the width of the sliding rectangular window is l, and the sliding step length of the sliding rectangular window is s, (s < l), and the width of the sliding rectangular window is l = 2 a l < L, a = 0, 1, 2, 3…, then the number of samples after sliding data increment processing is:

5. The bearing fault intelligent detection method based on inspection and vibration signals according to claim 1 is characterized in that: In step F, the network structure of the deep neural network model includes a one-dimensional convolutional network branch and a multi-layer perceptron network branch. The one-dimensional convolutional network can extract local features of fault information, and the multi-layer perceptron is used to extract global features of faults. Then, the outputs of the two network branches are fused through a connection layer, and then through a fully connected network. The network output dimension is the number of fault categories n. Finally, through the Softmax layer, the probability value of each fault category is obtained, and then the corresponding fault category is determined; The one-dimensional convolutional network branch includes multiple one-dimensional convolutional layers, maximum pooling layers and fully connected layers, and the activation function of each layer is a ReLU function; In order to prevent gradient explosion or diffusion and speed up training, a BatchNormalization layer is introduced after each layer, the convolution kernel size of each convolution layer is 2 and the step size is 1, and the pooling size and step size of the maximum pooling layer are both 4; the multi-layer perceptron network branch includes a fully connected layer and a Batch Normalization layer, and the activation function of each layer is also a ReLU function; The outputs of the two network branches are fused through a connection layer. The fusion process of the two network branches can splice the outputs of the two branches together to form a Concatenate layer, or directly add the outputs of the two network branches to form an add layer; when choosing to directly add the outputs of the two network branches to form an add layer, it is necessary to ensure that the output dimensions of the two network branches are the same.

6. The bearing fault intelligent detection method based on inspection and vibration signals according to claim 1 is characterized in that: In step F, during training, the loss function of the deep neural network model selects the cross entropy loss function, and the optimizer selects RMSProp.

7. The bearing fault intelligent detection method based on inspection and vibration signals according to claim 1 is characterized in that: In step G, the Case Western Reserve University rolling bearing fault data set has a bearing model of deep groove ball bearing SKF6205, and the operating states of the bearing include normal, inner ring fault, rolling element fault, and outer ring fault. Each type of fault includes three fault diameters of 0.1778 mm, 0.3556 mm, and 0.5334 mm, so 10 operating states can be formed, and the sampling frequency of the fault data is 12 kHz.

Images