A multi-element parallel neural network mechanical equipment fault diagnosis method
By using a multivariate parallel neural network structure and the EfficientNet network, combined with short-time Fourier transform and Manhattan distance discrimination, the problem of insufficient data in machine learning is solved, and efficient and accurate fault diagnosis of mechanical equipment is achieved.
Patent Information
- Application Number
- CN202311699666.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-12
- Publication Date
- 2025-11-28
- Estimated Expiration
- 2043-12-12
AI Technical Summary
In existing technologies, machine learning suffers from insufficient data in mechanical fault diagnosis, making it difficult to effectively train deep neural networks. Furthermore, traditional manual diagnosis is time-consuming, labor-intensive, and economically inefficient.
A multi-parallel neural network structure is adopted, combining short-time Fourier transform and EfficientNet network. Vibration data is collected through multiple sensors, feature extraction and loss function optimization are performed, and Manhattan distance is used to identify the fault category.
It improves data reuse rate and feature extraction capability, enabling rapid and accurate fault diagnosis of mechanical equipment, and enhancing diagnostic efficiency and accuracy.
Smart Images
Figure CN117743786B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of mechanical equipment fault diagnosis, and in particular to a multi-element parallel neural network mechanical equipment fault diagnosis method. BACKGROUND
[0002] China is a manufacturing power, and it is particularly important to ensure the good running state of mechanical equipment and monitor the health state of equipment in real time in production and manufacturing. The traditional mechanical fault diagnosis adopts the manual discrimination method, but this method is time-consuming and laborious and has low economic benefits. With the development of artificial intelligence, algorithm models are gradually applied in fault diagnosis, and intelligent mechanical fault diagnosis is realized.
[0003] Machine learning is first applied in fault diagnosis, however, machine learning performs poorly in nonlinear changes. Since mechanical fault information is hidden in vibration signals, the network model needs to have strong feature extraction capability to extract the features, so neural networks have good application prospects in the field of fault diagnosis. In the field of artificial intelligence, neural network models have developed rapidly in recent years. The current mainstream deep neural network models include convolutional neural networks, recurrent neural networks and long short-term memory networks. The above network models have strong nonlinear processing capability, but they need a large amount of sample data for model training. However, in actual production, there is not enough effective fault data to effectively train the deep neural network. SUMMARY
[0004] The present application relates to the technical field of mechanical equipment fault diagnosis, and in particular to a multi-element parallel neural network mechanical equipment fault diagnosis method.
[0005] A multi-element parallel neural network mechanical equipment fault diagnosis method comprises the following steps:
[0006] S1, a plurality of measuring points are set on the experimental equipment, and a sensor is installed on each measuring point for collecting vibration data;
[0007] S2, the vibration data is processed by using short-time Fourier transform, and the data is divided into a training set and a test set;
[0008] 3. Randomly take base samples, positive samples and negative samples from the training set to form a sample pair, and input the sample pair into a feature extraction network for training to obtain low-dimensional features of the sample pair;
[0009] S4, a loss function is set to calculate the loss value of the sample pair, and the network model parameters are optimized by back propagation;
[0010] S5, the test set sample is input into the network model to obtain low-dimensional features of the test set;
[0011] S6, transform the test set low-dimensional features into test set feature mean values, then compare the similarity with the training set feature mean values, then determine the fault category and output.
[0012] In the multi-element parallel neural network mechanical equipment fault diagnosis method, the data collected by each sensor in step S1 includes normal data and fault data of the measurement point.
[0013] In the multi-element parallel neural network mechanical equipment fault diagnosis method, in step S2, the short-time Fourier transform is calculated based on the Fourier transform with windowing, and the window function is selected as the Hanning window.
[0014]
[0015] Assuming that τ is any point on the time-frequency data, the windowed signal is:
[0016] y t = x t ·w(t-τ)
[0017] Performing short-time Fourier transform along the direction set by τ, S (w,τ) is the transformed result:
[0018]
[0019] In the multi-element parallel neural network mechanical equipment fault diagnosis method, in step S3, the base sample is a randomly selected sample in the training set, the positive sample refers to a sample with the same fault state as the base sample, and the negative sample refers to a sample with a different fault state from the base sample.
[0020] In the multi-element parallel neural network mechanical equipment fault diagnosis method, in step S3, the feature extraction network is provided with a sub-network structure, and EfficientNet network is used for vibration signal feature extraction in each module of the sub-network. After 4 times of feature extraction, a full connection layer is used to obtain low-dimensional features.
[0021] In the multi-element parallel neural network mechanical equipment fault diagnosis method, in step S4, the loss function defined by the model is as follows:
[0022]
[0023] Where s1 is the low-dimensional feature of the base sample, s2 is the low-dimensional feature of the positive sample, s3 is the low-dimensional feature of the negative sample, 0≤τ≤1 is a threshold set artificially, D represents the Manhattan distance between two n-dimensional feature vectors, and the distance calculation formula is as follows:
[0024]
[0025] When Loss takes The model is then optimized through backpropagation.
[0026] In the above-mentioned multi-parallel neural network fault diagnosis method for mechanical equipment, in step S4, the optimization of the model is stopped when the loss value does not meet the set threshold.
[0027] Compared with existing technologies, the advantages of this invention are:
[0028] 1. This invention designs a multi-parallel neural network structure based on convolutional neural networks. Vibration data with fault characteristics are input into sub-networks with similar three-layer network structures for feature extraction, thereby improving the data reuse rate.
[0029] 2. An EfficientNet network module was added to the sub-network. Compared with other network modules, it has stronger feature extraction capabilities and faster computing speed. In addition, Manhattan distance is used to determine the distance between the output results of different sub-networks, which facilitates the classification and output of different faults. Attached Figure Description
[0030] Figure 1 This is a schematic diagram of the process structure of the present invention.
[0031] Figure 2 This is a schematic diagram of data partitioning according to the present invention.
[0032] Figure 3 This is a schematic diagram of the feature extraction network in this invention.
[0033] Figure 4 This is a schematic diagram of the network model in this invention.
[0034] Figure 5 This is a diagram illustrating the data obfuscation matrix in an embodiment of the present invention. Detailed Implementation
[0035] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0036] Reference Figures 1-3 As shown, a multi-parallel neural network-based fault diagnosis method for mechanical equipment includes the following steps:
[0037] S1. Set up multiple measuring points on the experimental equipment, and install a sensor at each measuring point to collect vibration data;
[0038] S2. The vibration data is processed using short-time Fourier transform, and the data is divided into training set and test set;
[0039] S3. Randomly select base samples, positive samples, and negative samples from the training set to form a sample pair. Input the sample pair into the feature extraction network for training to obtain low-dimensional features of the sample pair.
[0040] S4. Set the loss function, calculate the loss value of the sample pair, and backpropagate to optimize the network model parameters;
[0041] S5. Input the test set samples into the network model to obtain the low-dimensional features of the test set;
[0042] S6. Transform the low-dimensional features of the test set into the mean of the test set features, then compare the similarity with the mean of the training set features, and finally determine the fault category and output it.
[0043] In step S1, the data collected by each sensor includes normal data and fault data at the measurement point.
[0044] In step S2, the short-time Fourier transform is performed by windowing based on the Fourier transform, and the Hanning window is selected as the window function.
[0045]
[0046] Assuming τ is any point in the time-frequency data, the windowed signal is:
[0047] y t =x t ·w(t-τ)
[0048] Perform a short-time Fourier transform along the direction set by τ, S (w,τ) The result after transformation:
[0049]
[0050] In step S3, the base sample is a sample randomly selected from the training set. The positive sample refers to a sample with the same fault state as the base sample, and the negative sample refers to a sample with a different fault state than the base sample. The feature extraction network is set with a sub-network structure. In each module of the sub-network, the EfficientNet network is used to extract vibration signal features from the sample pairs. After four feature extractions, a fully connected layer is used to obtain low-dimensional features.
[0051] In step S4, the loss function defined in this model is as follows:
[0052]
[0053] Wherein, s1 is the low-dimensional feature of the base sample, s2 is the low-dimensional feature of the positive sample, s3 is the low-dimensional feature of the negative sample, 0≤τ≤1 is a threshold value artificially set, D represents the Manhattan distance between two n-dimensional feature vectors, and the distance calculation formula is as follows:
[0054]
[0055] When the loss takes The model is optimized by back propagation, and the optimization of the model is stopped when the loss value does not satisfy the set threshold value, wherein the framework of the network model and the feature extraction network is the same, but the weight parameters are different, the feature extraction network does not have weight parameters at the beginning, the network training is performed using the feature extraction network, and the weight parameters are gradually generated in the training process, and the network with complete weight parameters generated after training is the final network for fault identification, which is also called a network model. The weight parameters generated after training are input into the network model to form a complete network model, and the complete network model can perform fault identification.
[0056] In step S6, the similarity determination uses the Manhattan distance formula to determine, which can calculate the 'distance' between different low-dimensional feature means. The method of converting low-dimensional features into feature means is the method of solving vector means. The feature mean refers to the mean of a set of samples (feature vectors). Assuming that the data samples of a certain fault type are n, the feature mean of this fault type is When the feature is d-dimensional:
[0057]
[0058] The network model solves all types of faults in the training set, and each fault data corresponds to a feature mean. If an unknown test set sample is input into the network model, the sample is closest to which fault in terms of similarity score, that is, the Manhattan distance is closest, and it is determined that the sample is which fault.
[0059] Embodiment
[0060] The data used by the algorithm model of the application is vibration data collected by a self-built experimental platform. Six measuring points are distributed on the experimental equipment, and a sensor is installed on each measuring point to collect vibration acceleration. The data collected by each sensor includes normal data and fault data of the measuring point, and a total of 12 types of data.
[0061] When verifying the identification performance of the model, 7 groups of type data are used for experiments, including fault data of 6 measuring points and normal data of measuring point 1. 60 samples of each type of data are selected, wherein 0 represents normal data, and 1-6 represents fault data of 1-6 measuring points. The confusion matrix obtained in the experiment is shown inFigure 5 As shown in the table, it can be seen that the recognition rate of the network model is high, and the fault data can be effectively recognized.
[0062] From the common technical knowledge, the application can be realized by other embodiments without departing from the spirit or essential characteristics thereof. Therefore, the above disclosed embodiments are only examples and are not the only ones. All changes within the scope of the application or within the scope equivalent to the application are included in the application.
Claims
1. A multi-parallel neural network-based fault diagnosis method for mechanical equipment, characterized in that: Includes the following steps: S1. Set up multiple measuring points on the experimental equipment, and install a sensor at each measuring point to collect vibration data; S2. The vibration data is processed using short-time Fourier transform, and the data is divided into training set and test set; S3. Randomly select base samples, positive samples, and negative samples from the training set to form a sample pair. Input the sample pair into the feature extraction network for training to obtain low-dimensional features of the sample pair. S4. Set the loss function, calculate the loss value of the sample pair, and backpropagate to optimize the network model parameters; S5. Input the test set samples into the network model to obtain the low-dimensional features of the test set; S6. Transform the low-dimensional features of the test set into the mean of the test set features, then compare the similarity with the mean of the training set features, and then determine the fault category and output it. In step S4, the loss function defined in this model is as follows: in, Low-dimensional features of the base samples These are low-dimensional features of positive samples. These are low-dimensional features of negative samples. A threshold set by an individual. Indicates two The Manhattan distance between 3D feature vectors is calculated using the following formula: when Pick The model is then optimized through backpropagation.
2. The multi-parallel neural network fault diagnosis method for mechanical equipment according to claim 1, characterized in that: In step S1, the data collected by each sensor includes normal data and fault data at the measurement point.
3. The multi-parallel neural network fault diagnosis method for mechanical equipment according to claim 1, characterized in that: In step S2, the short-time Fourier transform is performed by windowing calculation based on the Fourier transform, and the window function is... Choose Hanning windows; For any point in the time-frequency data, the windowed signal is: Along Perform a short-time Fourier transform in the given direction. The result after transformation: 。 4. The multi-parallel neural network fault diagnosis method for mechanical equipment according to claim 1, characterized in that: In step S3, the base sample is a sample randomly selected from the training set, the positive sample refers to a sample with the same fault state as the base sample, and the negative sample refers to a sample with a different fault state than the base sample.
5. The multi-parallel neural network fault diagnosis method for mechanical equipment according to claim 1, characterized in that: In step S3, the feature extraction network is set with a sub-network structure. In each module of the sub-network, the EfficientNet network is used to extract vibration signal features. After four feature extractions, a fully connected layer is applied to obtain low-dimensional features.
6. The multi-parallel neural network fault diagnosis method for mechanical equipment according to claim 1, characterized in that: In step S4, the optimization of the model is stopped when the loss value does not meet the set threshold.