Rotary mechanical equipment fault diagnosis method and system

Abstract

The invention discloses a rotary mechanical equipment fault diagnosis method, which comprises the following steps of: 1, acquiring diagnosis data of each test point, and preprocessing and dividing the data to obtain a test set; 2, processing a test set by using a feature extraction model; step 3, obtaining high-dimensional features of the test set; 4, processing the high-dimensional features of the test set by using the classification network model; and step 5, outputting and obtaining a fault type judgment result. The method is accurate and efficient in modeling, high in recognition accuracy and high in overall practicability, as shown in Figure 2, fault data collected by six sensors of an experiment table are classified after model training is completed, the result shows that the model classification effect is excellent, a confusion matrix is drawn for the classification result, it can be known that only a small part of samples of six types of sample data are misjudged, and the result is accurate. And the identification accuracy meets the industrial requirements, which shows that the equipment fault diagnosis method provided by the invention has certain practical value.

Description

Technical Field

[0001] The invention relates to the technical field of equipment diagnosis. Background Art

[0002] In the manufacturing industry, it takes a lot of resources to ensure the smooth operation of equipment. Therefore, real-time monitoring of the operating status of production equipment, predicting the development trend of the equipment health status in advance, and making reasonable maintenance plans can greatly improve production efficiency and economic benefits. At present, enterprises mainly rely on manual inspections to diagnose equipment faults, and judge the health status of equipment by observing the operating status of the equipment or listening to whether there are abnormal sounds during the operation of the equipment. This judgment method is inefficient and has a high false detection rate. It is no longer suitable for the current manufacturing industry.

[0003] Intelligent operation and maintenance is gradually being applied in the manufacturing industry. By installing a large number of sensors on the production line to collect equipment operation data in real time, diagnostic analysts can judge the health status of the equipment based on their own knowledge and experience and issue diagnostic reports. This method of mechanical fault diagnosis relies too much on the prior knowledge of diagnostic analysts and requires professional caregivers. The corresponding talents are scarce and have no significant effect on improving production efficiency. With the development of artificial intelligence technology, algorithm models are gradually being applied in equipment fault diagnosis.

[0004] At present, the most commonly used method is to build an algorithm model through a neural network. For example, the patent application number is 202210449533.2, the publication date is 2022-05-27, and the patent name is "A wind turbine fault diagnosis method based on adaptive residual neural network". The document discloses a wind turbine equipment fault diagnosis technology, including: using continuous wavelet transform to perform time-frequency processing on each sample in the original data set; inputting the two-dimensional time-frequency graph data set into the adaptive resize network to obtain low-dimensional abstract features; inputting the low-dimensional abstract features into the residual network built based on the LeNet-5 network structure to obtain the adaptive resize residual neural network model after joint training; inputting the actual vibration data of the wind turbine bearing into the adaptive resize residual neural network model to obtain the fault classification result. This algorithm model requires a large amount of valid data for model training, but real equipment fault data is relatively rare, and it is difficult to fully train the algorithm model. Summary of the invention

[0005] The technical problem to be solved by the present invention is to realize a rotating mechanical equipment fault diagnosis method and system which does not require large data training.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: a method for diagnosing faults of rotating mechanical equipment, comprising the following steps:

[0007] Step 1: Collect diagnostic data from each measuring point, pre-process and divide the data to obtain a test set;

[0008] Step 2: Use the feature extraction model to process the test set;

[0009] Step 3: Obtain high-dimensional features of the test set;

[0010] Step 4: Use the classification network model to process the high-dimensional features of the test set;

[0011] Step 5: Output the fault type identification result.

[0012] The data in step 1 is preprocessed and divided, including:

[0013] Data preprocessing:

[0014] Adoption Complete missing data;

[0015] In the above formula, x i is the vibration value at time i in the time series signal, x i-1 With x i+1 are the vibration values ​​before and after respectively;

[0016] Data partitioning:

[0017] A whole time series signal is divided into two segments according to the total number of vibration points in the ratio of 8:2. The first segment is the training set and the second segment is the validation set. The number of samples d in each segment can be obtained according to the formula Perform calculations;

[0018] Among them, l is the total length of the signal, m is the step size of the movement during segmentation, and the calculation result is rounded down. The size of m should be equal to the input length of the first branch of the subsequent feature extraction network.

[0019] Also includes: Algorithm model construction steps:

[0020] 1) Collect diagnostic data from each measuring point, pre-process and divide the data to obtain a training set;

[0021] 2) Processing the training set using the training feature extraction network;

[0022] 3) Calculate the loss value after processing;

[0023] 4) Determine whether the training is completed. If so, save the training weights and obtain a feature extraction model based on the data. If not, repeat the previous steps.

[0024] The step of constructing the algorithm model also includes:

[0025] 5) Process the training set using the feature extraction model;

[0026] 6) Obtain high-dimensional features of the training set;

[0027] 7) Use the smote model to expand data;

[0028] 8) Training to obtain a classification network model;

[0029] 9) Determine whether the training is completed. If not, return to the previous step. If so, save the training weights to obtain the classification network model.

[0030] In the step 2), the feature extraction network is composed of three branches, and the three branches extract features from the same vibration signal. The input of each branch is X1, X2, and X3, and feature extraction is performed to obtain x1, x2, and x3.

[0031] The three branches of the feature extraction network calibrate the vibration signal to a length of 2048 and then split it. The single-segment signal after splitting is used as the input X1 of the first branch of the feature extraction network. X1 is then split equally to obtain two vibration signals with a length of 1024 and used as the inputs X2 and X3 of the second and third branches respectively.

[0032] The feature extraction network concatenates x2 and x3, and then fuses the concatenated result with x1.

[0033] Feature fusion formula:

[0034] x 2,3 =x2☉x3

[0035]

[0036] Where ⊙ represents feature concatenation. Expand x2 and x3 separately, and then concatenate them in order. The length of the concatenated signal should be the same as the original signal length. It means to sum the corresponding elements.

[0037] Add the features of corresponding positions to get the complete data features.

[0038] In step 7), the smote algorithm is introduced into the model to achieve sample expansion and sample balance by randomly synthesizing sample type data with a small number of samples and randomly reducing sample type data with a large number of samples. New sample points are inserted between sample points in the minority class samples to expand the data. The synthesized data and the original data are combined into a new data set. The distance between the sample points of the minority class data is measured by the Euclidean distance.

[0039] The Euclidean distance:

[0040]

[0041] Where a and b are sample points. i With b i is the i-th eigenvalue of two sample points in the feature space, and the data dimension is n.

[0042] Select sample point x from the category with few samples i As the base sample, it is used to synthesize new samples, and x is calculated by the feature fusion formula. i The k nearest neighbor value of

[0043] From x i A sample point x is randomly selected from the k nearest neighbors i,j As auxiliary samples, it is used to synthesize new samples, and the sampling rate is set to N, x i With x i,j Perform difference operation to synthesize N new samples, using the formula:

[0044] X=x t +rand(0,1)*|x i -x ii |

[0045] Where X is the synthetic data, x i , x i,j They are base sample points and auxiliary sample points respectively.

[0046] The invention has accurate and efficient modeling, high recognition accuracy, and strong overall practicality. Figure 2 As shown, after the model training is completed, the fault data collected by the 6 sensors of the experimental platform are classified. The results show that the model classification effect is excellent, and the following confusion matrix is ​​drawn for the classification results. It can be seen that only a small number of samples in the 6 categories of sample data are misjudged, and the recognition accuracy meets the industrial needs. It shows that the equipment fault diagnosis method proposed in the present invention has certain practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The following is a brief description of the contents expressed in each figure in the specification of the present invention:

[0048] Figure 1 It is a flow chart of the rotating machinery fault diagnosis method;

[0049] Figure 2 A training example diagram for testing the rotating machinery fault diagnosis method;

[0050] Figure 3 Schematic diagram of the feature extraction network structure; DETAILED DESCRIPTION

[0051] The following is a further detailed description of the specific implementation methods of the present invention, such as the shape, structure, relative position and connection relationship between the various components involved, the function and working principle of each part, the manufacturing process and operation method, etc., through the description of the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0052] In view of the shortcomings of the current equipment fault diagnosis method, this paper proposes the following solutions:

[0053] 1. Build an algorithm model through a neural network and train the model. The trained model no longer requires diagnostic analysts to perform diagnostic care, thus realizing intelligent operation and maintenance.

[0054] 2. By constructing a branch network to extract the characteristic information of data of different fault types, three branch networks are constructed to extract features of vibration data respectively, and then the branch features are fused to more fully extract the fault features.

[0055] 3. A smote data enhancement algorithm module is added to the neural network to effectively solve the problem of insufficient real fault data, fully train the model, and improve the accuracy of the algorithm model in equipment fault diagnosis.

[0056] Specifically, if Figure 1 As shown in the figure, the equipment fault diagnosis process is divided into the collection and processing of vibration data, the construction of algorithm model, and the identification of fault type.

[0057] The method for diagnosing faults of rotating mechanical equipment comprises the following steps:

[0058] Step 1: Collect diagnostic data from each measuring point, pre-process and divide the data to obtain a test set;

[0059] The methods for preprocessing and dividing data include:

[0060] (1) Data preprocessing and partitioning

[0061] Due to the complexity of the real production environment, the raw data collected by the sensor contains a lot of noise interference or some data is missing, and it is necessary to ensure a certain amount of data to verify the effect of model training. Therefore, preprocessing and dividing the data becomes a necessary step;

[0062] (1.1) Data preprocessing

[0063] Due to the reasons of field equipment and sensor hardware, the collected raw time series data may have missing data at a certain time point. The missing of high-frequency sensor data density data will reduce the accuracy of the model, and most of the missing data is manifested as single data missing. For the continuous missing data segment, it means that the collected data has been distorted and cannot be directly discarded for model training. Based on the above considerations, formula (1) is used to fill the missing data.

[0064]

[0065] In the above formula, x i is the vibration value at time i in the time series signal, x i-1 With x i+1 are the vibration values ​​before and after respectively.

[0066] (1.2) Data partitioning

[0067] Data division: To accurately test the recognition accuracy of the model, it is necessary to ensure that the training and validation data do not overlap. Therefore, the entire time series signal is divided into two segments according to the total number of vibration points in the ratio of 8:2. The first segment is the training set and the second segment is the validation set. The number of samples d in each segment is calculated according to formula (2).

[0068]

[0069] In the above formula, l is the total length of the signal, m is the step size when splitting, and the calculation result is rounded down. Because the data will be smote-expanded in subsequent operations, there is no need to reuse the data when splitting. In this step, the size of m should be equal to the input length of the first branch of the subsequent feature extraction network.

[0070] Step 2: Use the feature extraction model to process the test set;

[0071] Step 3: Obtain high-dimensional features of the test set;

[0072] Step 4: Use the classification network model to process the high-dimensional features of the test set;

[0073] The network model is divided into two models, one is the feature extraction network and the other is the classification network. The classification network can be improved by using the currently common resnet18 network.

[0074] The preferred feature extraction network of the present invention is as follows:

[0075] (4.1) Feature extraction network

[0076] The feature extraction network adopts a branch network structure. Here, three branches are used for extraction and then fusion, which maximizes the integrity of the data. The structure is shown as follows: Figure 3 As shown;

[0077] For a certain type of fault data, input the feature extraction network for data expansion. The relevant experiments of the present invention involve 6 types of data, which are input into the network to obtain 6 types of data. Each type of data is divided into a data set of length 2048 and then further divided. X1, X2, and X3 are respectively input into three branches for convolution calculation to extract data features to obtain high-dimensional feature data, and then the high-dimensional data is feature fused, and the fused data is smote data expanded. Smote expansion of high-dimensional feature data.

[0078] (4.2) Classification Network

[0079] The data packet expanded by smote contains various fault features. The relevant experiments of the present invention expanded 6 types of fault data, and sent these 6 types of data to the classification network model to train the classification network model. The classification network uses an improved resnet-18 network. Because the resnet-18 network has a deeper layer, the present invention compresses the model to obtain a lightweight network model. The compressed model has 5 layers and includes two layers of residual structure. The lightweight network structure accelerates the computing performance and is more convenient for deployment on edge computing devices.

[0080] Step 5: Output the fault type identification result.

[0081] In step 1 above, when processing data, it is necessary to build an algorithm model. The steps for building the algorithm model include:

[0082] 1) Collect diagnostic data from each measuring point, pre-process and divide the data to obtain a training set. The training set and test set are processed in the same way.

[0083] 2) Processing the training set using the training feature extraction network;

[0084] The three branches of the feature extraction network calibrate the vibration signal to a length of 2048 and then split it. The single-segment signal after segmentation is used as the input X1 of the first branch of the feature extraction network. X1 is then equally split to obtain two vibration signals with a length of 1024 and used as the input X2 and X3 of the second and third branches respectively.

[0085] X1 is a single-segment signal obtained by segmenting a whole-length signal. The origin of X1 refers to (1.2) Data Division. X1 is then divided into two equal parts, X2 and X3. X1, X2, and X3 are respectively input into the first, second, and third branches of the feature extraction network. The results obtained from the three feature extraction layers are then subjected to feature fusion.

[0086] The feature extraction network concatenates x2 and x3, and then fuses the concatenated result with x1.

[0087] Feature fusion formula: x 2，3 =x2⊙x3

[0088] Add the features of corresponding positions to get the complete data features.

[0089] x 2，3 =x2☉x3

[0090]

[0091] Where ⊙ represents feature concatenation, x2 and x3 are expanded separately, and then the two are concatenated in sequence. The length of the concatenated signal should be the same as the original signal length. It means to sum the corresponding elements.

[0092] The advantages of this step method are: the vibration signal contains the status information of the equipment during operation, and feature extraction of the signal can obtain the operation information to realize equipment fault diagnosis, but the neural network needs to build multiple layers to fully extract the information to extract fault features, and too many computing resources will be consumed when the neural network structure is too complex. Considering that a considerable part of the production equipment that needs to be diagnosed is highly remote equipment, it is more appropriate to use edge computing equipment. The computing power of edge devices is limited and it is not suitable to use a deeper network structure. Therefore, in order to fully extract the vibration signal, the signal X1 is divided into X2 and X3, and then input into three networks with shallower network layers for feature extraction, and then feature fusion is performed.

[0093] 3) Calculate the loss value after processing;

[0094] Loss function settings:

[0095] In order to ensure that data features are fully extracted, several epochs need to be set to extract features. Too many cycles will make the data too redundant and cause overfitting of network training. Therefore, a suitable loss function should be selected to adjust the progress of data extraction. In this invention, the following loss function is selected to calculate the loss value:

[0096]

[0097] loss=max(d,0)(6)

[0098] Where d represents the distance between two sets of data, which is also calculated using the Euclidean distance (5). λ is an adjustment factor that can adjust the degree of vibration data extraction. It is adjusted according to different data. In this model, λ = 0.2. When d in formula (6) is 0 or a negative value, data extraction is stopped.

[0099] 4) Determine whether the training is completed. If so, save the training weights and obtain a feature extraction model based on the data. If not, repeat the previous steps.

[0100] 5) Process the training set using the feature extraction model;

[0101] 6) Obtain high-dimensional features of the training set;

[0102] 7) Use the smote model to expand data;

[0103] The specific operation process is shown in (4) Network model structure.

[0104] The principle of Smote expansion data is shown in (3.2) Vibration data processing;

[0105] The smote algorithm is introduced into the model to achieve sample expansion and sample balance by randomly synthesizing sample type data with a small number of samples and randomly reducing sample type data with a large number of samples. New sample points are inserted between sample points in minority class samples to expand the data. The synthesized data and the original data are combined into a new data set. The distance between sample points of minority class data is measured by Euclidean distance.

[0106] The Euclidean distance:

[0107]

[0108] Where a and b are sample points. i With b i is the i-th eigenvalue of two sample points in the feature space, and the data dimension is n.

[0109] Select sample point x from the category with few samples i As the base sample, it is used to synthesize new samples, and x is calculated by the feature fusion formula. i The k nearest neighbor value of

[0110] From x i A sample point x is randomly selected from the k nearest neighbors i,j As auxiliary samples, it is used to synthesize new samples, and the sampling rate is set to N, x i With x i,j Perform difference operation to synthesize N new samples, using the formula:

[0111] X=x i +rand(0,1)*|x i -x ij |

[0112] Where X is the synthetic data, x i , x i,j They are base sample points and auxiliary sample points respectively.

[0113] 8) Training to obtain a classification network model;

[0114] 9) Determine whether the training is completed. If not, return to the previous step. If so, save the training weights to obtain the classification network model.

[0115] 4. The method for fault diagnosis of rotating mechanical equipment according to claim 3 is characterized in that: in the step 2), the feature extraction network is composed of 3 branches, and the 3 branches perform feature extraction on the same vibration signal, and the input of each branch is X1, X2, and X3, and feature extraction is performed to obtain x1, x2, and x3.

[0116] In order to simulate the equipment operating environment and verify the reliability of the model, an experimental platform was built to collect vibration data. Six acceleration sensors were installed on the equipment, corresponding to six measuring points. Each measuring point can collect normal and fault data, and up to 12 types of data can be collected;

[0117] The sensor should be installed close to the bearing, or installed on the component that needs to be monitored and diagnosed to collect vibration data. It should be installed as close as possible to ensure the accuracy of the collected data. There are no special requirements for the installation steps.

[0118] For example, six test points are set up, and the output shaft of the test point is connected to the oil tank (of the oil supply circulation device) through multiple brackets. The power is led out to the drive shaft branch next to the output shaft through a belt on the drive shaft. The drive shaft branch drives the driven shaft to rotate, and the driven shaft is also supported by the bracket. The key points of the test point are the motor and the drive shaft, so the sensors of the test point are fixed on the motor, as well as the drive shaft branch, the drive shaft bracket, and the driven shaft bracket.

[0119] After the model training is completed, the fault data collected by the six sensors on the test bench are classified, such as Figure 2 As shown, the results show that the model classification effect is excellent, and the following confusion matrix is ​​drawn for the classification results. It can be seen that only a small number of samples in the 6 categories of sample data are misjudged, and the recognition accuracy meets industrial needs. This shows that the equipment fault diagnosis method proposed in the present invention has certain practical value.

[0120] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for diagnosing faults of rotating mechanical equipment, characterized in that: The following steps are involved: Step 1: Collect diagnostic data from each measuring point, pre-process and divide the data to obtain a test set; Step 2: Use the feature extraction model to process the test set; Step 3: Obtain high-dimensional features of the test set; Step 4: Use the classification network model to process the high-dimensional features of the test set; Step 5: Output the fault type identification result.

2. The method for diagnosing faults of rotating machinery equipment according to claim 1, characterized in that: The data in step 1 is preprocessed and divided, including: Data preprocessing: Adoption Complete missing data; In the above formula, x i is the vibration value at time i in the time series signal, x i-1 With x i+1 are the vibration values ​​before and after respectively; Data partitioning: A whole time series signal is divided into two segments according to the total number of vibration points in the ratio of 8:

2. The first segment is the training set and the second segment is the validation set. The number of samples d in each segment can be obtained according to the formula Perform calculations; Among them, l is the total length of the signal, m is the step size of the movement during segmentation, and the calculation result is rounded down. The size of m should be equal to the input length of the first branch of the subsequent feature extraction network.

3. The method for diagnosing faults of rotating machinery according to claim 2, characterized in that: Also includes: Steps to build the algorithm model: 1) Collect diagnostic data from each measuring point, pre-process and divide the data to obtain a training set; 2) Processing the training set using the training feature extraction network; 3) Calculate the loss value after processing; 4) Determine whether the training is completed. If so, save the training weights and obtain a feature extraction model based on the data. If not, repeat the previous steps.

4. The method for diagnosing faults of rotating machinery according to claim 3, characterized in that: The step of constructing the algorithm model also includes: 5) Process the training set using the feature extraction model; 6) Obtain high-dimensional features of the training set; 7) Use the smote model to expand data; 8) Training to obtain a classification network model; 9) Determine whether the training is completed. If not, return to the previous step. If so, save the training weights to obtain the classification network model.

5. The method for diagnosing faults of rotating machinery according to claim 4, characterized in that: In the step 2), the feature extraction network is composed of three branches, and the three branches extract features from the same vibration signal. The input of each branch is X1, X2, and X3, and feature extraction is performed to obtain x1, x2, and x3.

6. The method for diagnosing faults of rotating machinery according to claim 5, characterized in that: The three branches of the feature extraction network calibrate the vibration signal to a length of 2048 and then split it. The single-segment signal after splitting is used as the input X1 of the first branch of the feature extraction network. X1 is then split equally to obtain two vibration signals with a length of 1024 and used as the inputs X2 and X3 of the second and third branches respectively.

7. The method for diagnosing faults of rotating machinery according to claim 6, characterized in that: The feature extraction network concatenates x2 and x3, and then fuses the concatenated result with x1. Feature fusion formula: x 2，3 =x2⊙x3 Where ⊙ represents feature concatenation, x2 and x3 are expanded separately, and then the two are concatenated in sequence. The length of the concatenated signal should be the same as the original signal length. It means to sum the corresponding elements.

8. The method for diagnosing faults of rotating machinery according to claim 4 or 7, characterized in that: In step 7), the smote algorithm is introduced into the model to achieve sample expansion and sample balance by randomly synthesizing sample type data with a small number of samples and randomly reducing sample type data with a large number of samples. New sample points are inserted between sample points in the minority class samples to expand the data. The synthesized data and the original data are combined into a new data set. The distance between the sample points of the minority class data is measured by the Euclidean distance. The Euclidean distance: Where a and b are sample points. i With b i is the i-th eigenvalue of two sample points in the feature space, and the data dimension is n.

9. The method for diagnosing faults of rotating machinery according to claim 8, characterized in that: Select sample point x from the category with few samples i As the base sample, it is used to synthesize new samples, and x is calculated by the feature fusion formula. i The k nearest neighbor value of Randomly select a sample point x from the k nearest neighbors of xi i,j As auxiliary samples, it is used to synthesize new samples, and the sampling rate is set to N, x i With x i,j Perform difference operation to synthesize N new samples, using the formula: X=x i +rand(0,1)*|x i -x ij | Where X is the synthetic data, x i , x i,j They are base sample points and auxiliary sample points respectively.

Images