Deep learning method for multi-source vibration identification

Abstract

The invention discloses a deep learning method for multi-source vibration recognition. The deep learning method comprises the steps of 1, testing and collecting an original vibration data set; step 2, data segmentation: randomly segmenting the collected original vibration data according to a set sample length; 3, performing continuous wavelet transform on the data, and converting one-dimensional vibration data into a two-dimensional image; step 4, dividing the data set; step 5, building a ResNet deep learning model; step 6, training the ResNet model, and performing parameter optimization according to a training result; 7, the test set is imported into the trained model for testing so as to obtain a classification result and test accuracy, and if the test accuracy is 95% or above generally, it is regarded that the target accuracy is reached; and 8, putting the prepared to-be-identified multi-source vibration data into the trained model for identification, wherein the identification comprises the application of continuous wavelet transform, the identification of a two-dimensional image and the inverse process of conversion to one-dimensional data.

Description

Technical Field

[0001] The present invention relates to the field of rotation control technology, and more specifically to a deep learning method for multi-source vibration recognition, which can be applied to strong and micro vibration control fields such as construction engineering, industrial engineering, and semiconductors. Background Art

[0002] In recent years, the application of engineering vibration control technology in modern industrial engineering, scientific and technological research and development, and bottleneck research has become more and more extensive. When carrying out vibration control, especially the design of active and semi-active control systems, the identification of vibration source components is directly related to the design accuracy of the control system and the effective performance of vibration control. In previous practical projects, engineers often rely on their own engineering experience or Fourier transform to perform spectral analysis on all or part of the collected time-course data to obtain frequency components. This approach cannot accurately obtain the detailed components of the vibration source, and the results are relatively rough and unreliable. It is necessary to propose a more accurate and effective multi-source vibration component identification method. Summary of the invention

[0003] In view of the above problems, the present invention proposes a deep learning method for multi-source vibration identification. The inventive concept of this design method is to closely combine continuous wavelet transform and deep learning method, convert one-dimensional vibration data into two-dimensional time-frequency diagram, use deep learning method to extract two-dimensional image features, and realize accurate identification of vibration data based on large model training.

[0004] More specifically, according to one aspect of the present invention, a deep learning method for multi-source vibration recognition is provided, comprising:

[0005] Step 1: Test and collect the original vibration data set. The original vibration data type is generally not less than 20 groups, and the length of a single data is generally not less than 60,000;

[0006] Step 2, data segmentation, including randomly segmenting the collected original vibration data according to the set sample length;

[0007] Step 3, performing continuous wavelet transform on the data to convert the one-dimensional vibration data into a two-dimensional image;

[0008] Step 4: Divide the data set into training set and test set;

[0009] Step 5: Build the ResNet deep learning model and set parameters;

[0010] Step 6: Train the ResNet model, optimize the parameters according to the training results, and then save the trained model;

[0011] Step 7: Import the test set into the trained model and test it to obtain the classification results and test accuracy. Generally, the test accuracy is above 95%, which is considered to have reached the target accuracy.

[0012] Step 8: Input the prepared multi-source vibration data to be identified into the trained model for identification, including the application of continuous wavelet transform and the inverse process of two-dimensional image identification and conversion to one-dimensional data.

[0013] According to an embodiment of the present invention, step one also includes the consistency of sampling frequency of each group of vibration data, the richness of each group of vibration data and samples, that is, they should basically cover all potential components of the vibration signal to be identified, and the data samples are not repetitive.

[0014] According to an embodiment of the present invention, the step three should also include that when using continuous wavelet transform, the selection of wavelet basis should be optimized according to the characteristics of vibration data.

[0015] According to an embodiment of the present invention, step five includes selecting a shallow or deep ResNet, preferably ResNet50.

[0016] According to an embodiment of the present invention, step six further comprises using T-SNE technology to visualize data of samples before and after training.

[0017] According to an embodiment of the present invention, step seven further comprises using a classification confusion matrix to evaluate the accuracy of predicted samples and true samples.

[0018] According to an embodiment of the present invention, step eight further includes obtaining identification results of the vibration data to be identified in different time periods and identifying the stripping of multi-source vibration components. The stripping process is preferably for vibration data at every 3s interval.

[0019] According to another aspect of the present invention, there is provided a deep learning device for multi-source vibration recognition, comprising:

[0020] The original vibration data set collection module tests and collects the original vibration data set. The original vibration data type is generally not less than 20 groups, and the length of a single data is generally not less than 60,000;

[0021] The data segmentation module includes randomly segmenting the collected original vibration data according to the set sample length;

[0022] A data transformation module is used to perform continuous wavelet transformation on the data to convert the one-dimensional vibration data into a two-dimensional image;

[0023] The data set partitioning module is used to partition the data set into a training set and a test set;

[0024] ResNet deep learning model building module, used to build the ResNet deep learning model and set parameters;

[0025] ResNet model training module, used to train the ResNet model, optimize parameters based on the training results, and then save the trained model;

[0026] ResNet model testing module, used to import the test set into the trained model for testing to obtain classification results and test accuracy. Generally, a test accuracy of more than 95% is considered to have reached the target accuracy; and

[0027] The multi-source vibration data recognition module is used to input the prepared multi-source vibration data to be recognized into the trained model for recognition, including the application of continuous wavelet transform and the inverse process of two-dimensional image recognition and conversion to one-dimensional data.

[0028] According to another aspect of the present invention, there is also provided an electronic device, comprising: a memory and one or more processors;

[0029] The memory is used to store one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in the present invention.

[0030] The present invention proposes a deep learning method for multi-source vibration recognition. Based on a large number of samples and vibration data collection, the method converts one-dimensional vibration data into a two-dimensional time-frequency graph based on continuous wavelet transform, and then selects and designs a suitable ResNet deep learning network as needed to automatically extract the converted two-dimensional image features, and finally completes classification through the SoftMax layer of ResNet to realize vibration signal recognition. The method is based on artificial intelligence and deep learning methods, has high efficiency, high accuracy, good intelligence, and can eliminate the influence of prior knowledge and expert experience.

[0031] Based on the following detailed description of specific embodiments of the present invention in conjunction with the accompanying drawings, those skilled in the art will become more aware of the above and other objects, advantages and features of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic flow chart of a deep learning method for multi-source vibration recognition according to an embodiment of the present invention;

[0033] Figure 2 A schematic structural diagram of a deep learning device for multi-source vibration recognition according to an embodiment of the present invention; and

[0034] Figure 3Schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. The shown contents are used to fully illustrate the contents of the present invention, but are not used to limit the present invention.

[0036] It should be understood that the ResNet deep learning model, continuous wavelet transform, T-SNE technology and classification confusion matrix involved in the present invention are known in themselves. Therefore, the present invention focuses on how to combine and apply the above-mentioned various tools or theories to design the process of the deep learning method for multi-source vibration identification of the present invention.

[0037] Figure 1 FIG. 1 is a flow chart of a deep learning method for multi-source vibration recognition according to an embodiment of the present invention. Figure 1 , the deep learning method for multi-source vibration recognition according to an embodiment of the present invention may include:

[0038] First, widely collect engineering vibration signals containing various components and build a sample database. More specifically, a wide range of multi-source vibration components are used as the background database for purposeful collection, which can include rail transit, ground pulses, power equipment, wind, pedestrians, construction, etc., which almost cover all possible vibration signals encountered in daily industrial engineering or scientific research projects. Ensure the consistency of the sampling frequency of each group of vibration data, the richness of each group of vibration data and samples, that is, it should basically cover all the potential components of the vibration signal to be identified, and it is not recommended that the data samples have repetitiveness. The original vibration data type should generally not be less than 20 groups, and the length of a single data should generally not be less than 60,000.

[0039] The data is randomly segmented, including randomly segmenting the collected original vibration data according to the set sample length.

[0040] Perform continuous wavelet transform on the data to convert the one-dimensional vibration data into a two-dimensional image. When designing continuous wavelet transform, the selection of wavelet basis should basically match the characteristics of the vibration signal to be processed, that is, the selection of wavelet basis should be optimized according to the characteristics of vibration data. Then divide the data into training set and test set according to a certain ratio, which can generally be divided in a ratio of 7:3.

[0041] After that, you can build the ResNet model and set the parameters. ResNet models can generally be divided into shallow and deep models. According to the requirements of engineering vibration signal recognition, the complexity of network depth and the ability to handle practical problems, ResNet50 can generally be selected.

[0042] After the model is built, the above training set can be used for model training, which can include the setting and optimization of key parameters, such as the initial learning rate, number of trainings, etc.; it also includes continuously optimizing the selected key training parameters according to the training results, such as the initial learning rate, etc. After the parameter optimization verification converges, the training is completed, and the model can be stored and called after training.

[0043] The trained model can then be tested using the test set. That is, the test set is imported into the trained model for testing. This process includes using T-SNE technology to visualize the data of the samples before and after training, and using the classification confusion matrix to evaluate the accuracy of the predicted samples and the real samples. The test obtains the classification results and the test accuracy. Generally, the test accuracy is above 95%, which is considered to have reached the target accuracy.

[0044] Then the multi-source vibration data to be identified is fed into the trained model for identification and evaluated based on the classification and identification accuracy. For example, the design process can include the application of continuous wavelet transform and the inverse process of two-dimensional image identification and conversion to one-dimensional data, which can realize a time domain vibration signal, identify vibration components at a certain time interval, and strip the identified multi-source vibration components. The stripping process is preferably for each vibration data with an interval of 3s.

[0045] The deep learning method for multi-source vibration identification of the implementation scheme of the present invention is cleverly designed and has significant key points. It can start from the mechanism of vibration identification and realize vibration identification from the perspective of artificial intelligence through deep learning based on large sample and large model data. In addition, the method also cleverly introduces continuous wavelet transform to build a bridge between one-dimensional vibration data and two-dimensional time-frequency images, closely linking vibration identification with the advantages of deep learning, and ensuring the efficiency of the method. The invention can grasp the powerful advantages of deep learning in image recognition, while grasping the key to converting vibration signals into images. It can find breakthroughs in traditional data analysis and processing, and connect them with deep learning, thereby innovating traditional methods into the field of artificial intelligence. The algorithm is powerful and has strong applicability, which has important guiding significance for practical engineering.

[0046] Figure 2 Schematic diagram of a deep learning device for multi-source vibration recognition according to an embodiment of the present invention. Figure 2As shown, the device includes: an original vibration data set collection module 210, which is used to test and collect original vibration data sets. The original vibration data type is generally not less than 20 groups, and the length of a single data is generally not less than 60,000; a data segmentation module 220, which is used to segment the data, including randomly segmenting the collected original vibration data according to the set sample length; a data transformation module 230, which is used to perform continuous wavelet transformation on the data to convert the one-dimensional vibration data into a two-dimensional image; a data set division module 240: used to divide the data set into a training set and a test set; a ResNet deep learning model building module 250, which is used to build a ResNet deep learning model. et deep learning model, and set parameters; ResNet model training module 260, used to train the ResNet model, optimize parameters according to the training results, and then save the trained model; ResNet model testing module 270, used to import the test set into the trained model for testing to obtain classification results and test accuracy. Generally, the test accuracy is above 95%, which is considered to have reached the target accuracy; multi-source vibration data recognition module 280, used to import the prepared multi-source vibration data to be identified into the trained model for identification, including the application of continuous wavelet transform and the inverse process of two-dimensional image recognition and conversion to one-dimensional data.

[0047] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention is shown in FIG. Figure 3 As shown, the electronic device includes a processor 310, a memory 320, an input device 330 and an output device 340; the number of the processor 310 in the electronic device can be one or more. Figure 3 A processor 310 is taken as an example; the processor 310, the memory 320, the input device 330 and the output device 340 in the electronic device can be connected via a bus or other means. Figure 3 The example of connecting through bus is taken in the following.

[0048] The memory 320, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the deep learning method for multi-source vibration recognition in the embodiment of the present invention (for example, the original vibration data set collection module 210; the data segmentation module 220; the data transformation module 230; the data set division module 240; the ResNet deep learning model building module 250; the ResNet model training module 260; the ResNet model testing module 270; the multi-source vibration data recognition module 280). The processor 310 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory 320, that is, realizes the above-mentioned deep learning method for multi-source vibration recognition.

[0049] The present invention can achieve beneficial technical effects:

[0050] (1) Identifying multi-source vibration components in an efficient, accurate and intelligent manner is an important application scenario of modern artificial intelligence technology.

[0051] (2) Intelligent identification of vibration components based on large sample training is a representative work that promotes the development of traditional expert prior knowledge and is a method with important contemporary significance.

[0052] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the embodiments herein, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A deep learning method for multi-source vibration recognition, characterized in that: include: Step 1: Test and collect the original vibration data set. The original vibration data type is generally not less than 20 groups, and the length of a single data is generally not less than 60,000; Step 2, data segmentation, including randomly segmenting the collected original vibration data according to the set sample length; Step 3, performing continuous wavelet transform on the data to convert the one-dimensional vibration data into a two-dimensional image; Step 4: Divide the data set into training set and test set; Step 5: Build the ResNet deep learning model and set parameters; Step 6: Train the ResNet model, optimize the parameters according to the training results, and then save the trained model; Step 7: Import the test set into the trained model and test it to obtain the classification results and test accuracy. Generally, the test accuracy is above 95%, which is considered to have reached the target accuracy. Step 8: Input the prepared multi-source vibration data to be identified into the trained model for identification, including the application of continuous wavelet transform and the inverse process of two-dimensional image identification and conversion to one-dimensional data.

2. A deep learning method for multi-source vibration recognition according to claim 1, characterized in that: The step one also includes the consistency of sampling frequency of each group of vibration data, and the richness of each group of vibration data and samples, that is, they should basically cover all potential components of the vibration signal to be identified, and the data samples should not be repetitive.

3. A deep learning method for multi-source vibration recognition according to claim 1, characterized in that: The step three should also include that when using continuous wavelet transform, the selection of wavelet basis should be optimized according to the characteristics of vibration data.

4. A deep learning method for multi-source vibration recognition according to claim 1, characterized in that: The step five includes selecting a shallow or deep ResNet, preferably ResNet50.

5. The deep learning method for multi-source vibration recognition according to claim 1, characterized in that: The step six also includes the use of T-SNE technology to visualize the data of the samples before and after training.

6. A deep learning method for multi-source vibration recognition according to claim 1, characterized in that: The step seven also includes the use of a classification confusion matrix to evaluate the accuracy of predicted samples and true samples.

7. A deep learning method for multi-source vibration recognition according to claim 1, characterized in that: The step eight also includes obtaining identification results of the vibration data to be identified in different time periods and identifying the stripping of multi-source vibration components. The stripping process is preferably performed on vibration data at intervals of 3s.

8. A deep learning device for multi-source vibration recognition, characterized in that: include: The original vibration data set collection module tests and collects the original vibration data set. The original vibration data type is generally not less than 20 groups, and the length of a single data is generally not less than 60,000; The data segmentation module includes randomly segmenting the collected original vibration data according to the set sample length; A data transformation module is used to perform continuous wavelet transformation on the data to convert the one-dimensional vibration data into a two-dimensional image; The data set partitioning module is used to partition the data set into a training set and a test set; ResNet deep learning model building module, used to build the ResNet deep learning model and set parameters; ResNet model training module, used to train the ResNet model, optimize parameters based on the training results, and then save the trained model; The ResNet model testing module is used to import the test set into the trained model for testing to obtain the classification results and test accuracy. Generally, the test accuracy is above 95%, which is considered to have reached the target accuracy. as well as The multi-source vibration data recognition module is used to input the prepared multi-source vibration data to be recognized into the trained model for recognition, including the application of continuous wavelet transform and the inverse process of two-dimensional image recognition and conversion to one-dimensional data.

9. An electronic device, characterized in that: include: memory and one or more processors; The memory is used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

Images