A method for extracting schumann resonance time-frequency map features based on convolutional autoencoder

Abstract

This invention discloses a method for extracting Schumann resonance time-frequency image features based on a convolutional autoencoder. The training process of the convolutional autoencoder involves: first, merging, cropping, grayscale conversion, and standardization of the original time-frequency image set to obtain grayscale time-frequency images with consistent width and height; then, designing the encoder as a UNET model structure and the decoder as a mapping function satisfying conditional constraints; and using the processed grayscale time-frequency images as both the training set and labels, inputting the time-frequency images into the autoencoder for iterative training, with gradient updates to the autoencoder parameters. This method allows for the calculation of the first and second derivatives of the time-frequency image, enabling observation of the direction and rate of change of unit pixels, thus facilitating the analysis of time-frequency image data and providing a solution for time-frequency image data analysis.

Description

Technical Field

[0001] This invention relates to a method for feature mapping of Schumann resonance time-frequency graphs using a convolutional autoencoder, and more particularly to a method for analyzing changes in time-frequency graphs, belonging to the fields of image processing and data analysis methods for Schumann resonances. Background Technology

[0002] On the one hand, due to the inherent nature of the original time-domain signal, the anomalous changes reflected in the time-frequency plot may be minute, resulting in the problem that the anomalies observed in the time-frequency plot are not obvious. On the other hand, the time-frequency plot itself and the pixels within it are discrete quantities; there is no method to simultaneously differise time and frequency, meaning it is impossible to observe the amplitude changes of the time-frequency plot within a unit time and unit frequency, nor is it possible to directly calculate data features such as the direction and rate of change of each pixel in the time-frequency plot. Furthermore, in current cases of Schumann resonance time-frequency plot data analysis, most analytical methods are based on directly observing changes in data before and after specific events (such as earthquakes and geomagnetic storms) and extracting and analyzing anomalies through mathematical statistical methods. These analytical methods and feature extraction are relatively simple and direct, lacking complexity and interpretability, and cannot fully analyze anomalies in Schumann resonance signals. Summary of the Invention

[0003] To address the aforementioned problems in existing technologies, this invention aims to provide a Schumann resonance time-frequency map feature extraction method based on a convolutional autoencoder. By selecting a suitable encoder model and designing a decoder that meets the constraints, the method can perform first-order and second-order matrix differentiation on the time-frequency map for unit pixel changes, thereby obtaining the variation law of the time-frequency map.

[0004] To achieve the objectives of this invention, the technical solution adopted is as follows:

[0005] A method for extracting Schumann resonance time-frequency map features based on a convolutional autoencoder, the method comprising the following steps:

[0006] Step 1: Perform image merging, high-resolution cropping, grayscale conversion, and standardization on the original time-frequency image set;

[0007] Step 2: Select a suitable model structure as the encoder based on the specific problem;

[0008] Step 3: Design a decoder function that satisfies the constraints;

[0009] Step 4: Fuse the encoder and decoder designed in Step 2 and Step 3 to obtain a convolutional autoencoder. Input the time-frequency image processed in Step 1 into the convolutional autoencoder for calculation. Use the input time-frequency image as both training sample and sample label. Calculate the mean square error of the model input and output as the loss value. Iterate the model based on the loss value.

[0010] Step 5: Repeat step 4 until the maximum number of iterations of the algorithm is reached.

[0011] As an improvement of the present invention, the image preprocessing method in step 1 is as follows:

[0012] The original time-frequency graph has a resolution of 640*560, covering a time range of 0–24 hours (from left to right) and a frequency range of 0–50Hz (from bottom to top). Both the unit time and unit frequency in the time-frequency graph are equally spaced, meaning the time resolution is 2.25 minutes and the frequency resolution is 0.089Hz. First, the seven time-frequency graphs are arranged in rows, then cropped using a 560*560 window, resulting in eight 560*560 time-frequency graphs. At this point, each image is reduced to 21 hours in time scale while remaining unchanged in frequency scale. Finally, the images are converted to grayscale, and the pixels of the cropped images are standardized with a base value of 1.

[0013] As an improvement of this invention, in step 2, a suitable model structure is selected as the encoder based on the specific problem. Considering factors such as the encoder's output dimension, the decoder's structure, the type of input data, and the implementation goal, the UNET model is selected as the main structure of the encoder. Features are extracted from the time-frequency graph through downsampling of UNET, and the shape of the time-frequency graph is restored through upsampling. The advantage of choosing UNET is that it can extract features at different levels in the time-frequency graph, such as amplitude features (brightness, pixel size), frequency features, etc., and integrate them through feature overlay.

[0014] As an improvement to the present invention, in step 3, a decoder function that satisfies the constraints is designed:

[0015] To explore the decoder structure in an autoencoder, let's assume the original image matrix input to the model is Y, and the image matrix output by the convolutional autoencoder is H. Generally speaking, H approximates Y, and for a single element h in the image matrix H... ij In terms of h ij It also includes information on time, frequency, and amplitude.

[0016] Now assume that there exist functions f and x ij As shown in Equation 1, the following condition is met:

[0017] h ij =f(x) ij i∈[0,560],j∈[0,560],x ij ∈R,h ij ∈R(1)

[0018] Then x ijProvided that H is sufficiently close to Y, it can be considered that H also contains information about time, frequency, and amplitude.

[0019] Therefore, in order to obtain a single element h ij To determine the changes in [the data], the following constraints need to be established:

[0020]

[0021] Therefore, we can assume the function f is:

[0022]

[0023] It satisfies:

[0024]

[0025] Meanwhile, considering that the actual calculations involve matrix operations, we can set x as follows: ij The resulting mapping matrix is ​​X, and the square matrix with all elements equal to 1 is... Then we have:

[0026]

[0027] Among them, matrix multiplication is the Hadamard product, which is the multiplication of elements at corresponding positions, and f(X) is the decoder in the autoencoder.

[0028] As an improvement of the present invention, in step 4, the encoder and decoder are fused, and the time-frequency graph processed in step 1 is input into the convolutional autoencoder for calculation. The specific process is as follows:

[0029] First, connect the UNET encoder and the decoder function that meets the constraints to obtain the convolutional autoencoder. Then, input the preprocessed image matrix Y into the convolutional autoencoder to obtain the output matrix H. Calculate the mean squared error MSE(Y,H) and iterate the convolutional autoencoder model using the gradient method based on the mean squared error MSE(Y,H).

[0030] As an improvement of the present invention, in step 5, step 4 is repeated until the maximum number of iterations of the algorithm is reached, the iteration number t is initialized to 0, and the maximum iteration number t is set. max =100, after completing step 4, t = t + 1, repeat until t = t max The algorithm can be represented as follows:

[0031]

[0032]

[0033] Compared with other methods, the method proposed in this invention has the following beneficial effects:

[0034] First, by calculating the convolutional autoencoder, the information of the time-frequency graph is enhanced, highlighting the changes in the data. At the same time, the feature enhancement in the time-frequency graph is achieved through the design of the decoder, making the pixel changes in the time-frequency graph more sensitive in numerical representation.

[0035] Second, by designing a decoder that satisfies the constraints, the time-frequency... Figure 1 It is second-order and second-differentiable, and its value range is over the entire real number domain.

[0036] Third, since the mapped matrix is ​​a numerical matrix with consistent width and height, matrix features such as determinant, rank, and trace can be calculated. This allows for targeted analysis of the research problem based on data changes, providing more feature extraction capabilities for the analysis of Schumann resonance time-frequency diagram data. To a certain extent, this provides a solution to the current research bottleneck of Schumann resonance data and offers an effective and interpretable feature extraction method for the changes in Schumann resonance time-frequency diagram data. Attached Figure Description

[0037] Figure 1 This is a flowchart of a Schumann resonance time-frequency map feature extraction method based on a convolutional autoencoder according to the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] Example 1: See Figure 1 A method for extracting Schumann resonance time-frequency map features based on a convolutional autoencoder, the method comprising the following steps:

[0040] Step 1: Perform image merging, high-resolution cropping, grayscale conversion, and standardization on the original time-frequency image set;

[0041] Step 2: Select a suitable model structure as the encoder based on the specific problem;

[0042] Step 3: Design a decoder function that satisfies the constraints;

[0043] Step 4: Fuse the encoder and decoder designed in Step 2 and Step 3 to obtain a convolutional autoencoder. Input the time-frequency image processed in Step 1 into the convolutional autoencoder for calculation. Use the input time-frequency image as both training sample and sample label. Calculate the mean square error of the model input and output as the loss value. Iterate the model based on the loss value.

[0044] Step 5: Repeat step 4 until the maximum number of iterations of the algorithm is reached.

[0045] In step 1, the image preprocessing method is as follows: The original time-frequency image has a resolution of 640*560, covering a time range of 0–24 hours (from left to right) and a frequency range of 0–50Hz (from bottom to top). The time and frequency units in the time-frequency image are equally spaced, meaning the time resolution is 2.25 minutes and the frequency resolution is 0.089Hz. First, the seven time-frequency images are arranged in rows, then cropped using a 560*560 window, resulting in eight 560*560 time-frequency images. At this point, the time scale of a single image is reduced to 21 hours, while the frequency scale remains unchanged. Finally, the image is converted to grayscale, and the pixels of the cropped image are standardized with a base of 1.

[0046] In step 2, a suitable model structure is selected as the encoder based on the specific problem. Considering factors such as the encoder's output dimension, the decoder's structure, the type of input data, and the implementation goal, the UNET model is chosen as the main structure of the encoder. UNET is used for downsampling to extract features from the time-frequency graph, and upsampling is used to restore the shape of the time-frequency graph. The advantage of choosing UNET is that it can extract features at different levels in the time-frequency graph, such as amplitude features (brightness, pixel size) and frequency features, and integrate them through feature overlay.

[0047] In step 3, a decoder function that satisfies the constraints is designed. To explore the decoder structure in the autoencoder, let the original image matrix input to the model be Y, and the image matrix output by the convolutional autoencoder be H. Generally speaking, H approximates Y, and for a single element h in the image matrix H... ij In terms of h ij It also includes information on time, frequency, and amplitude.

[0048] Now assume that there exist functions f and x ij As shown in Equation 1, the following condition is met:

[0049] h ij =f(x) ij i∈[0,560],j∈[0,560],x ij ∈R,h ij ∈R(1)

[0050] Then x ij Provided that H is sufficiently close to Y, it can be considered that H also contains information about time, frequency, and amplitude.

[0051] Therefore, in order to obtain a single element h ij To determine the changes in [the data], the following constraints need to be established:

[0052]

[0053] Therefore, we can assume the function f is:

[0054]

[0055] It satisfies:

[0056]

[0057] Meanwhile, considering that the actual calculations involve matrix operations, we can set x as follows: ij The resulting mapping matrix is ​​X, and the square matrix with all elements equal to 1 is... Then we have:

[0058]

[0059] Among them, matrix multiplication is the Hadamard product, which is the multiplication of elements at corresponding positions, and f is the decoder.

[0060] In step 4, the encoder and decoder are fused. The time-frequency image processed in step 1 is input into the convolutional autoencoder for calculation. The specific process is as follows: First, the UNET encoder and the decoder function that meets the constraints are connected to obtain the convolutional autoencoder. Then, the preprocessed image matrix Y is input into the convolutional autoencoder to obtain the output matrix H. The mean square error MSE(Y,H) is calculated, and the convolutional autoencoder model is iterated using the gradient method based on the mean square error MSE(Y,H).

[0061] In step 5, repeat step 4 until the maximum number of iterations of the algorithm is reached. Initialize the iteration number t = 0 and set the maximum iteration number t. max =100, after completing step 4, t = t + 1, repeat until t = t max The algorithm can be represented as follows:

[0062]

[0063] It should be noted that the above embodiments are merely best practice examples of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or alternatives made based on the above technical solutions are all within the scope of protection of the present invention.

Claims

1. A method for extracting Schumann resonance time-frequency map features based on a convolutional autoencoder, characterized in that, The method includes the following steps: Step 1: Image preprocessing, the original time-frequency image set is merged, cropped while maintaining high invariance, grayscale converted, and standardized; Step 2: Select a suitable model structure as the encoder based on the specific problem; Step 3: Design a decoder function that satisfies the constraints; Step 4: Fuse the encoder and decoder designed in Step 2 and Step 3 to obtain a convolutional autoencoder. Input the time-frequency image processed in Step 1 into the convolutional autoencoder for calculation. Use the input time-frequency image as both training sample and sample label. Calculate the mean square error of the model input and output as the loss value. Iterate the model based on the loss value. Step 5: Repeat step 4 until the maximum number of iterations of the algorithm is reached; In step 3, the decoder function that satisfies the constraints is designed as follows: Let the original image matrix input to the model be... The image matrix output by the convolutional autoencoder is ,have Approaching And for image matrices single element In other words, It also includes information on time, frequency, and amplitude. Now assume there exists a function and As shown in Equation 1, the following condition is met: (1) but exist close enough to Under the premise that, it can be regarded as It also includes information on time, frequency, and amplitude. Therefore, in order to obtain a single element To determine the changes in [the data], the following constraints need to be established: (2) Therefore, we can set a function for: (3) It satisfies: (4) Meanwhile, considering that the actual calculations involve matrix operations, we can assume that... The resulting mapping matrix is A square matrix with all elements being 1 is Then we have: (5) In this context, matrix multiplication is called the Hadamard product, which involves multiplying corresponding elements. This is the decoder in the autoencoder.

2. The method for extracting Schumann resonance time-frequency map features based on a convolutional autoencoder according to claim 1, characterized in that, In step 1, the image preprocessing method is as follows: The original time-frequency image has a resolution of 640*560, covering a time range of 0-24 hours and a frequency range of 0-50Hz. The time and frequency units in the time-frequency image are equally spaced, i.e., the time resolution is 2.25 minutes and the frequency resolution is 0.089Hz. First, the 7 time-frequency images are arranged in rows, and then cropped with a 560*560 window, resulting in 8 560*560 time-frequency images. At this point, the time scale of a single image is reduced to 21 hours, while the frequency scale remains unchanged. Finally, the image is grayscaled, and the pixels of the cropped image are standardized with 1 as the baseline.

3. The method for extracting Schumann resonance time-frequency map features based on a convolutional autoencoder according to claim 1, characterized in that, In step 2, a suitable model structure is selected as the encoder according to the specific problem, as follows: the UNET model is selected as the main structure of the encoder. The time-frequency graph is extracted by downsampling of UNET, and the shape of the time-frequency graph is restored by upsampling. The advantage of selecting UNET is that UNET can extract features at different levels in the time-frequency graph and integrate them by feature superposition.

4. The method for extracting Schumann resonance time-frequency map features based on a convolutional autoencoder according to claim 1, characterized in that, In step 4, the encoder and decoder are fused. The time-frequency image processed in step 1 is input into the convolutional autoencoder for calculation. Specifically, the UNET encoder and the decoder function that meets the constraints are first connected to obtain the convolutional autoencoder. Then, the preprocessed image matrix is... The input is fed into a convolutional autoencoder to obtain the output matrix. Calculate the mean squared error (MSE). , ), and based on the mean squared error (MSE) , The gradient method is used to iterate the convolutional autoencoder model.

5. The method for extracting Schumann resonance time-frequency map features based on a convolutional autoencoder according to claim 1, characterized in that: Repeat step 4 until the maximum number of iterations of the algorithm is reached, then initialize the number of iterations. Set the maximum number of iterations. After completing step 4 of the calculation Repeat until The algorithm can be expressed as follows: 。

Images