A Deep Learning-Based Method and System for Extracting Feature Sequences from Periodic Signal Data
By dividing periodic signal data into multiple segments and converting it into a two-dimensional graph, and using convolutional neural networks and Grad-CAM to extract feature sequences, the problems of inaccurate signal feature extraction and noise processing in existing technologies are solved, achieving more efficient signal analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA ELECTRIC POWER RESEARCH INSTITUTE CO LTD
- Filing Date
- 2021-01-20
- Publication Date
- 2026-07-31
AI Technical Summary
Existing signal analysis methods, such as empirical mode decomposition, suffer from mode aliasing problems. Noise factors affect the accuracy and efficiency of signal feature extraction, and convolutional neural networks have difficulty processing signals mixed with noise.
Periodic signal data is divided into multiple segments according to time windows and converted into a two-dimensional data signal map. A convolutional neural network model is used to locate feature regions, and feature sequences are extracted through Grad-CAM and binarization operations to optimize the signal processing flow.
It improves the accuracy and efficiency of signal processing, reduces the difficulty of identifying noise signals, and enhances the generalization ability and robustness of signal analysis.
Smart Images
Figure CN113657139B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of load forecasting, and in particular relates to a method and system for extracting feature sequences from periodic signal data based on deep learning. Background Technology
[0002] Signal analysis and processing involves removing spurious signals and interference from test signals to obtain the desired useful information. Generally, the process of studying the composition and eigenvalues of a signal is called signal analysis, while the process of performing necessary transformations on the signal to obtain the required information is called signal processing. These two processes are interrelated. Existing signal analysis methods can handle various types of complex nonlinear signals. For example, Empirical Mode Decomposition (EMD) is an adaptive time-frequency signal analysis method that can be widely used for analyzing nonlinear and non-stationary processes. It decomposes the signal into several intrinsic eigenfunctions (IMFS) and then extracts signal data features by analyzing these eigenfunctions. However, this type of method has significant limitations. It relies on expert experience and is often subject to mode aliasing due to noise, resulting in low accuracy and efficiency in signal feature extraction.
[0003] In recent years, machine learning (ML) theory has developed rapidly, aiming to simulate human learning in real time by using computers as tools and effectively improve learning efficiency by dividing existing content into knowledge structures. Deep learning (DL) is a new research direction in the field of machine learning. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition that far surpass previous related technologies. Deep learning methods can automatically learn features and make classification decisions, significantly improving both accuracy and efficiency. Convolutional neural networks are one of the most popular deep learning models in recent years. Signals obtained through testing are often mixed with various noises. The source of noise may be due to imperfections in the testing device itself, or it may be due to other input sources mixed into the system that cannot process the signal. Moreover, existing convolutional neural networks cannot handle signals mixed with various noise data, leading to the difficulty of noise signal recognition. Summary of the Invention
[0004] To overcome the shortcomings of the existing technology, this invention proposes a method for extracting feature sequences from periodic signal data based on deep learning, comprising:
[0005] S1: Divide the periodic signal data into multiple segments according to the preset time window;
[0006] S2: The multiple signal data segments are converted and combined to obtain a two-dimensional data signal graph;
[0007] S3: Input the two-dimensional data signal image into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal image;
[0008] S4: Determine the feature sequence of the signal data based on the plurality of feature regions;
[0009] The convolutional neural network model is trained based on a two-dimensional data signal graph with feature labels.
[0010] Preferably, determining the feature sequence of the signal data based on the plurality of feature regions includes:
[0011] Binarization is performed based on the multiple feature regions;
[0012] The binarization operation is used to trace the multiple feature regions to obtain multiple feature sequences of the signal data.
[0013] Preferably, the step of converting and combining the multiple signal data segments to obtain a two-dimensional data signal graph includes:
[0014] The multiple signal data segments are transformed using short-time Fourier transform or wavelet transform, and then combined and arranged into a matrix of a preset size to obtain a two-dimensional data signal graph corresponding to the signal data.
[0015] Preferably, the establishment of the convolutional neural network model includes:
[0016] The convolutional neural network model is trained and obtained by using a two-dimensional data signal map with feature labels as input and the feature regions in the two-dimensional data signal map with feature labels as output.
[0017] Preferably, the acquisition of a two-dimensional data signal map with feature markers includes:
[0018] Acquire historical periodic signal data and label the historical periodic signal data according to its features;
[0019] The historical periodic signal data is transformed and combined into a two-dimensional data signal graph with feature labels.
[0020] Preferably, locating multiple feature regions from the two-dimensional data signal image includes:
[0021] Multiple feature regions are located from the two-dimensional data signal image using Grad-CAM.
[0022] Preferably, after inputting the two-dimensional data signal image into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal image, and before determining the feature sequence of the signal data based on the multiple feature regions, the method further includes:
[0023] The two-dimensional data signal map is trained multiple times in the convolutional neural network model until multiple feature regions located from the two-dimensional data signal map are the same.
[0024] Preferably, after determining the feature sequence of the signal data based on the plurality of feature regions, the method further includes:
[0025] By repeatedly executing steps S1 to S4 with different preset lengths, feature sequences under multiple preset time windows can be obtained.
[0026] Select the identical feature sequence from all feature sequences as the optimal feature sequence;
[0027] The optimal feature sequence is set as the feature sequence corresponding to the periodic signal data. Based on the same inventive concept, the present invention also provides a deep learning-based system for extracting feature sequences from periodic signal data, comprising: a partitioning module, a transformation and combination module, a training and localization module, and a sequence determination module;
[0028] The segmentation module is used to divide periodic signal data into multiple segments according to a preset time window;
[0029] The conversion and combination module is used to convert and combine the multiple signal data segments to obtain a two-dimensional data signal diagram;
[0030] The training and localization module is used to input the two-dimensional data signal map into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal map.
[0031] The sequence determination module is used to determine the feature sequence of the signal data based on the plurality of feature regions;
[0032] The convolutional neural network model is trained based on a two-dimensional data signal graph with feature labels.
[0033] Preferably, the sequence determination module includes: a binarization submodule and a tracing submodule;
[0034] The binarization submodule is used to perform binarization operations based on the multiple feature regions.
[0035] The tracing submodule is used to trace the multiple feature regions according to the binarization operation to obtain multiple feature sequences of the signal data.
[0036] Compared with the closest existing technology, the present invention has the following beneficial effects:
[0037] 1. This invention implements a method and system for extracting feature sequences from periodic signal data based on deep learning, comprising: dividing periodic signal data into multiple segments according to a preset time window; converting and combining the multiple segments of signal data to obtain a two-dimensional data signal map; inputting the two-dimensional data signal map into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal map; and determining the feature sequence of the signal data based on the multiple feature regions; wherein, the convolutional neural network model is trained based on the two-dimensional data signal map with feature labels; this invention provides a signal recognition method that improves the problem that existing convolutional neural networks cannot handle signals mixed with various noise data, thereby improving the accuracy and efficiency of signal processing;
[0038] 2. This invention can optimize and improve the overall system, thereby enhancing generalization ability and robustness, and addressing the limitations and high degree of human intervention in signal analysis. Attached Figure Description
[0039] Figure 1 A schematic diagram of the process for extracting feature sequences from periodic signal data based on deep learning, provided by the present invention;
[0040] Figure 2 This is a diagram illustrating the effect of converting a one-dimensional periodic signal into a two-dimensional matrix in an embodiment of the present invention.
[0041] Figure 3 This is a schematic diagram of the structure of the convolutional neural network in an embodiment of the present invention;
[0042] Figure 4 This is a feature region map obtained using Grad-CAM and a binarized result image in an embodiment of the present invention;
[0043] Figure 5 This is an illustration of the effect of tracing a binarized graph to points on a one-dimensional sequence in an embodiment of the present invention.
[0044] Figure 6 This is the overall flowchart in the embodiments of the present invention;
[0045] Figure 7 This invention provides a schematic diagram of a system for extracting feature sequences from periodic signal data based on deep learning. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0047] Example 1:
[0048] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0049] like Figure 1 As shown, an embodiment of the present invention provides a method for extracting feature sequences from periodic signal data based on deep learning, comprising:
[0050] S1: Divide the periodic signal data into multiple segments according to the preset time window;
[0051] S2: The multiple signal data segments are converted and combined to obtain a two-dimensional data signal graph;
[0052] S3: Input the two-dimensional data signal image into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal image;
[0053] S4: Determine the feature sequence of the signal data based on the plurality of feature regions;
[0054] The convolutional neural network model is trained based on a two-dimensional data signal graph with feature labels.
[0055] To address the shortcomings of existing signal processing methods, this invention proposes a method for finding the feature sequences of periodic signals. The technical solution adopted in this invention is as follows:
[0056] First, various periodic signals are acquired and labeled. Then, these one-dimensional signal sequences (time series with time M) are transformed into two-dimensional matrices. A convolutional neural network (CNN) is then used to transform the signal analysis problem into an image processing problem (signal-image conversion). A large number of labeled periodic signal samples are used for training the CNN. Grad-CAM is used to visualize the feature regions of the images; these feature regions are the basis for the CNN to classify these images. Therefore, based on these regions, binarization allows us to trace the feature sequences of the one-dimensional signals. Finally, we change the size of the two-dimensional matrix (i.e., the image size) and repeat the above operations to obtain the feature sequences of the same one-dimensional signal again. We then filter out sequences that appear simultaneously in both preceding and following feature sequences; these filtered sequences are the feature sequences we are looking for.
[0057] The convolutional neural network classification module transforms the labeled one-dimensional signal into a two-dimensional matrix and then trains the convolutional neural network.
[0058] The feature visualization module uses the Grad-CAM method to obtain regions that serve as the classification basis for a convolutional neural network, and ensures that the arrangement of the obtained feature regions is similar to the texture in the original image by training the neural network multiple times.
[0059] The thresholding module is used to binarize the obtained image, select an appropriate threshold, and ensure that the number of pixels with a pixel value of 255 accounts for about one-quarter of the total number of pixels in the binarized image.
[0060] The mathematical operation module is used to trace the points with a pixel value of 255 in the image to a one-dimensional signal, and by changing the image size and repeating the above method, different feature sequences of the same signal source can be obtained.
[0061] This invention proposes a method for noise detection and diagnosis in substation inspections using periodic signals, comprising the following steps:
[0062] S1: Divide the periodic signal data into multiple segments according to the preset time window;
[0063] Step (1) Obtain several periodic signals and label them for subsequent learning and training of deep convolutional neural networks.
[0064] Step (2): Each periodic signal is divided into sampling points of a specified length.
[0065] S2: The multiple signal data segments are converted and combined to obtain a two-dimensional data signal graph;
[0066] Short-time Fourier transform (SFT) or wavelet transform is used to convert one-dimensional signals into a matrix of a preset size, arranged as a time sequence. Alternatively, multiple array segments can be arranged into a preset matrix as a time sequence through preset mapping conditions, and then transformed into a two-dimensional matrix. (SFT represents the signal characteristics at a certain moment through a segment of signal within a time window. The window length determines the time and frequency resolution of the spectrum; the longer the window, the longer the intercepted signal, resulting in higher frequency resolution but lower time resolution. SFT first multiplies a function with a window function, then performs a one-dimensional Fourier transform, obtaining a series of Fourier transform results through window function sliding, and then arranges these results into a two-dimensional time-frequency graph. Wavelet transform can automatically adjust the window size according to frequency, enabling multi-resolution analysis.) Considering the finiteness of the original measurement data, overlapping sampling of a certain length is used to increase the number of samples during sample construction. A convolutional neural network is trained using a two-dimensional matrix as training samples. The convolutional neural network includes convolutional layers, pooling layers, fully connected layers, and an output layer.
[0067] S3: Input the two-dimensional data signal image into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal image;
[0068] Step (3) uses Grad-CAM to visualize the image and obtain the region that the convolutional neural network uses as the classification basis.
[0069] Step (4) train the convolutional neural network multiple times using the same training sample to ensure that the feature regions obtained by Grad-CAM have a high similarity to the texture of the original image.
[0070] Step (5): Binarize the feature regions obtained on the two-dimensional image, select an appropriate threshold, and ensure that the number of white pixels in the binarized image accounts for about one-quarter of the total number of pixels in the image.
[0071] S4: Determine the feature sequence of the signal data based on the plurality of feature regions;
[0072] Step (6): Based on the feature regions obtained on the two-dimensional image, trace back to the one-dimensional signal to obtain several feature sequences of the one-dimensional signal.
[0073] Step (7): For the same periodic signal, change the size of its two-dimensional matrix and repeat steps 2 to 6 to filter out the sequences that are also feature sequences, which are the feature sequences of the periodic signal to be obtained.
[0074] Following these seven steps can help improve the safety of non-contact noise sensors, reduce the complexity of processing, and greatly enhance the detection and diagnostic capabilities of the equipment.
[0075] In addition to noise signals, the above steps can also be performed on pulse signals and periodic harmonic signals in the power grid.
[0076] Figure 2 The image shows the effect of converting a one-dimensional periodic signal into a two-dimensional matrix. This can convert time-series periodic signals into two-dimensional matrices for training convolutional neural networks.
[0077] Figure 3 The diagram shows a convolutional neural network structure used for deep learning, including an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional and pooling layers are responsible for automatic feature learning and extraction, while the output layer, combined with a classifier, can obtain the types and states of a two-dimensional matrix.
[0078] Figure 4The image shows the feature region map obtained using Grad-CAM and the binarized result. The Grad-CAM algorithm is a feature visualization method that allows us to obtain regions that can be used as classification criteria by CNN.
[0079] Figure 5 To trace the binarized graph back to points on a one-dimensional sequence.
[0080] Figure 6 This is the overall flowchart.
[0081] Example 2:
[0082] Based on the same inventive concept, this invention also provides a system for extracting feature sequences from periodic signal data based on deep learning, such as... Figure 7 As shown, it includes: a partitioning module, a transformation and combination module, a training and localization module, and a sequence determination module;
[0083] The segmentation module is used to divide periodic signal data into multiple segments according to a preset time window;
[0084] The conversion and combination module is used to convert and combine the multiple signal data segments to obtain a two-dimensional data signal diagram;
[0085] The training and localization module is used to input the two-dimensional data signal map into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal map.
[0086] The sequence determination module is used to determine the feature sequence of the signal data based on the plurality of feature regions;
[0087] The convolutional neural network model is trained based on a two-dimensional data signal graph with feature labels.
[0088] Preferably, the sequence determination module includes: a binarization submodule and a tracing submodule;
[0089] The binarization submodule is used to perform binarization operations based on the multiple feature regions.
[0090] The tracing submodule is used to trace the multiple feature regions according to the binarization operation to obtain multiple feature sequences of the signal data.
[0091] Preferably, the conversion and combination module includes: a preset conversion submodule;
[0092] The preset conversion submodule is used to convert the multiple signal data segments using short-time Fourier transform or wavelet transform, and combine them into a matrix of preset size to obtain a two-dimensional data signal graph corresponding to the signal data.
[0093] The establishment of the convolutional neural network model includes:
[0094] The convolutional neural network model is trained and obtained by using a two-dimensional data signal map with feature labels as input and the feature regions in the two-dimensional data signal map with feature labels as output.
[0095] Acquisition of two-dimensional data signal maps with feature markers includes:
[0096] Acquire historical periodic signal data and label the historical periodic signal data according to its features;
[0097] The historical periodic signal data is transformed and combined into a two-dimensional data signal graph with feature labels.
[0098] The training localization module includes: a localization submodule;
[0099] The positioning submodule is used to locate multiple feature regions from the two-dimensional data signal map using Grad-CAM.
[0100] The system also includes: a multiple training module;
[0101] The multiple training module is used to train the two-dimensional data signal map multiple times in the convolutional neural network model until the multiple feature regions located from the two-dimensional data signal map are the same.
[0102] The system also includes: a repetitive execution module, a sequence acquisition module, and a sequence setting module;
[0103] The repetitive execution module is used to repeatedly execute a deep learning-based method for extracting periodic signal data features by setting different preset lengths to obtain feature sequences under multiple preset time windows.
[0104] The sequence acquisition module is used to select the same feature sequence from all feature sequences as the optimal feature sequence;
[0105] The sequence module is configured to set the optimal feature sequence as the feature sequence corresponding to the periodic signal data.
[0106] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.
Claims
1. A method for extracting a sequence of features of periodic signal data based on deep learning, characterized in that, include: S1: Divide the periodic signal data into multiple segments according to the preset time window; S2: The multiple signal data segments are converted and combined to obtain a two-dimensional data signal graph; S3: Input the two-dimensional data signal image into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal image; S4: Determine the feature sequence of the signal data based on the plurality of feature regions; The convolutional neural network model is trained based on a two-dimensional data signal graph with feature labels; the periodic signal data includes: substation inspection noise signal, power grid pulse signal, or power grid harmonic signal. The process of converting and combining the multiple signal data segments to obtain a two-dimensional data signal image includes: The multiple signal data segments are transformed using short-time Fourier transform or wavelet transform, and then combined and arranged into a matrix of a preset size to obtain a two-dimensional data signal graph corresponding to the signal data. Locating multiple feature regions from the two-dimensional data signal image, including: Multiple feature regions are located from the two-dimensional data signal image using Grad-CAM techniques; Determining the feature sequence of the signal data based on the plurality of feature regions includes: Binarization is performed based on the multiple feature regions; The binarization operation is used to trace the multiple feature regions to obtain multiple feature sequences of the signal data; After determining the feature sequence of the signal data based on the plurality of feature regions, the method further includes: By repeatedly executing steps S1 to S4 with different preset lengths, feature sequences under multiple preset time windows can be obtained. Select the identical feature sequence from all feature sequences as the optimal feature sequence; The optimal feature sequence is set as the feature sequence corresponding to the periodic signal data.
2. The method of claim 1, wherein, The establishment of the convolutional neural network model includes: The convolutional neural network model is trained and obtained by using a two-dimensional data signal map with feature labels as input and the feature regions in the two-dimensional data signal map with feature labels as output.
3. The method of claim 2, wherein, Acquisition of two-dimensional data signal maps with feature markers includes: Acquire historical periodic signal data and label the historical periodic signal data according to its features; The historical periodic signal data is transformed and combined into a two-dimensional data signal graph with feature labels.
4. The method of claim 1, wherein, After inputting the two-dimensional data signal image into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal image, and before determining the feature sequence of the signal data based on the multiple feature regions, the method further includes: The two-dimensional data signal map is trained multiple times in the convolutional neural network model until multiple feature regions located from the two-dimensional data signal map are the same. 5.A deep learning based periodic signal data feature sequence extraction system, characterized in that, include: The module consists of a partitioning module, a transformation and combination module, a training and localization module, and a sequence determination module. The segmentation module is used to divide periodic signal data into multiple segments according to a preset time window; The conversion and combination module is used to convert and combine the multiple signal data segments to obtain a two-dimensional data signal diagram; The training and localization module is used to input the two-dimensional data signal map into a pre-established convolutional neural network model to locate multiple feature regions from the two-dimensional data signal map. The sequence determination module is used to determine the feature sequence of the signal data based on the plurality of feature regions; The convolutional neural network model is trained based on a two-dimensional data signal graph with feature labels; the periodic signal data includes: substation inspection noise signal, power grid pulse signal, or power grid harmonic signal. The conversion and combination module includes: a preset conversion submodule; The preset conversion submodule is used to convert the multiple signal data segments using short-time Fourier transform or wavelet transform, and combine them into a matrix of preset size to obtain a two-dimensional data signal graph corresponding to the signal data. The training localization module includes: a localization submodule; The positioning submodule is used to locate multiple feature regions from the two-dimensional data signal image using Grad-CAM. The sequence determination module includes: a binarization submodule and a tracing submodule; The binarization submodule is used to perform binarization operations based on the multiple feature regions; The tracing submodule is used to trace the multiple feature regions according to the binarization operation to obtain multiple feature sequences of the signal data; The system also includes: a repetitive execution module, a sequence acquisition module, and a sequence setting module; The repetitive execution module is used to repeatedly execute a deep learning-based method for extracting periodic signal data features by setting different preset lengths to obtain feature sequences under multiple preset time windows. The sequence acquisition module is used to select the same feature sequence from all feature sequences as the optimal feature sequence; The sequence module is configured to set the optimal feature sequence as the feature sequence corresponding to the periodic signal data.