Broadband random harmonic compression analysis method based on LSTM and multi-feature fusion

Abstract

The invention discloses a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion, and belongs to the technical field of electric digital data processing. Comprising the following steps: collecting a broadband random harmonic signal for time domain analysis, and extracting a period; judging whether each period signal can be compressed or retained or not by using LSTM (Long Short Term Memory); carrying out down-sampling and coding compression on the compressible periodic section signal; decoding and reconstructing the signal through an LSTM decoder; the decoded and reconstructed signals are preprocessed, multi-dimensional features of the signals are extracted, and evaluation indexes are normalized; calculating the compression ratio of the dynamic fusion index and the signal; and dynamically displaying the evaluation index and the dynamic fusion index. According to the method, self-adaptive decision of compression or not is realized, the compression ratio is improved, and meanwhile, the spectrum fidelity and phase consistency are also kept; and parts which do not need to be compressed are reserved as original samples, so that real-time performance and error control are both considered.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing technology, and in particular to a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion. Background Technology

[0002] With the large-scale integration of nonlinear loads such as new energy power generation equipment, power electronic devices, and electric vehicles into power systems, the types and distribution characteristics of harmonics in the power grid have become significantly more complex and broadband. Traditionally, harmonics are mainly concentrated within integer multiples of the fundamental power frequency, exhibiting relatively stable frequencies and amplitudes. However, in new power systems containing numerous power electronic converters and high-frequency switching devices, harmonic signals not only exhibit discrete harmonic characteristics in the low-frequency band but also display broadband random disturbances in the form of a continuous spectrum in the mid-to-high frequency range (from several kilohertz to tens of kilohertz). These disturbances exhibit strong non-stationarity and randomness in the time, frequency, and phase domains, posing potential threats to power quality, equipment safety, and system stability.

[0003] In signal compression and harmonic analysis, existing technologies mainly focus on fixed frequency bands, wavelet transform, and compressed sensing. Some researchers have combined compressed sensing with empirical wavelet transform to achieve frequency band decomposition and energy measurement of broadband interference signals, thereby improving the accuracy of compressed reconstruction. Amaya et al. further utilized compressed sensing technology to locate harmonic distortion sources, achieving high-precision frequency recovery under low sampling rate conditions. Although the above methods have improved the spectrum reconstruction capability to some extent, they all rely on the sparsity of the signal or predefined frequency band division. When faced with highly dynamic and highly random broadband harmonics, the reconstruction performance drops significantly. In addition, some existing technologies attempt to use time-frequency distribution functions, wavelet packet decomposition, or energy index construction methods for broadband signal analysis, but their computational complexity is high, cross-term interference is severe, and stability is insufficient under conditions of multi-source harmonics, noise interference, and rapid random disturbances. More importantly, existing methods generally use a single feature dimension as the evaluation basis, lacking cross-domain fusion and adaptive weighting mechanisms. This makes the evaluation results susceptible to local anomalies or feature redundancy, making it difficult to achieve a comprehensive and quantitative description of the time-varying characteristics of broadband random harmonics.

[0004] Existing methods for harmonic compression and analysis have significant limitations when dealing with broadband random harmonics:

[0005] (1) Compression rules rely on manual thresholds or fixed frequency band divisions, which cannot adapt to the dynamic random characteristics of disturbances. There is a lack of adaptive correlation between frequency domain feature extraction and sampling strategies, making it difficult to achieve “signal complexity driven” compression decisions.

[0006] (2) Linear interpolation or sparse reconstruction models are not responsive enough to high-order time-series features, resulting in a decrease in the phase consistency and spectral structure fidelity of the reconstructed signal;

[0007] (3) Lack of comprehensive characterization of randomness and non-stationarity. Existing indicators such as total harmonic distortion (THD), harmonic current ratio, and harmonic power factor are mostly based on steady-state or periodic signal assumptions, and cannot effectively assess the dynamic impact of broadband random harmonics under time-varying and non-periodic conditions.

[0008] (4) The indicators are too singular and lack feature integration. Most current harmonic indicators only focus on energy or amplitude characteristics, while ignoring key factors such as frequency shift, phase stability, and spectral sparsity. This results in insufficient accuracy in assessing broadband random disturbances and fails to reveal the intrinsic relationship between harmonic energy distribution and system dynamics.

[0009] Therefore, a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion is needed to achieve high-fidelity reconstruction, interpretable feature analysis, and dynamic quantitative characterization of broadband random harmonics in power systems. Summary of the Invention

[0010] The purpose of this invention is to propose a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion.

[0011] A broadband random harmonic compression analysis method based on LSTM and multi-feature fusion includes the following steps:

[0012] Acquire broadband random harmonic signals, perform time-domain analysis on the signals, and extract the signal period;

[0013] Use LSTM to determine whether the signal in each cycle segment is compressible or can be preserved.

[0014] Downsampling and encoding compression are performed on compressible periodic signals;

[0015] The downsampled and coded compressed signal is decoded and reconstructed using an LSTM decoder;

[0016] The decoded and reconstructed signal is preprocessed to extract multidimensional features of the signal, and the evaluation index of the multidimensional features is normalized.

[0017] The dynamic fusion index and signal compression ratio are calculated using a variance-mixed weighting mechanism;

[0018] The evaluation indicators of multidimensional features are dynamically displayed along with the dynamically fused indicators.

[0019] Furthermore, the downsampling and encoding compression of the compressible periodic signal is performed as follows: For the original periodic signal with N sampling points, M sampling points are retained after compression. The compressed time-domain sequence is then input into an LSTM-based encoder model to extract nonlinear dynamic features and perform deep compression.

[0020] Furthermore, the preprocessing of the decoded and reconstructed signal is as follows:

[0021] The decoded and reconstructed signal is segmented according to time windows. Within each time window, the Hilbert transform is used to obtain the analytical form of the signal, and the instantaneous amplitude and phase information are extracted.

[0022] Furthermore, the time window length is 40ms.

[0023] Furthermore, an overlap rate of 50% to 80% is set between adjacent time windows.

[0024] Furthermore, the evaluation indicators for multidimensional features include: amplitude standard deviation, spectral centroid, spectral flatness, phase coherence, energy kurtosis, and spectral sparsity.

[0025] Furthermore, the calculation formula for the dynamic fusion index is as follows:

[0026] ;

[0027] in, For time, The dynamic weight of the j-th indicator. Let j be the j-th evaluation index after normalization.

[0028] A broadband random harmonic compression analysis device based on LSTM and multi-feature fusion includes:

[0029] The period extraction module is used to acquire broadband random harmonic signals, perform time-domain analysis on the signals, and extract the period of the signals.

[0030] The signal determination module is used to determine whether the signal in each cycle segment can be compressed or retained using LSTM.

[0031] The encoding and compression module is used to downsample and encode compressible periodic signals;

[0032] The decoding and reconstruction module is used to decode and reconstruct the downsampled and encoded compressed signal using an LSTM decoder.

[0033] The feature processing module is used to preprocess the decoded and reconstructed signal, extract multi-dimensional features of the signal, and normalize the evaluation index of the multi-dimensional features.

[0034] The index calculation module is used to calculate the compression ratio of the dynamically fused index and signal using a variance-mixed weighting mechanism.

[0035] The dynamic display module is used to dynamically display the evaluation indicators of multi-dimensional features and the dynamic fusion indicators.

[0036] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps in a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion.

[0037] A storage medium storing a computer program that, when executed by a processor, implements the various steps of a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion.

[0038] The beneficial effects of this invention are as follows:

[0039] 1. This invention performs structured modeling of broadband signals through periodic segmentation and feature analysis. It uses an LSTM-based discriminator to replace the fixed frequency band threshold, achieving adaptive decision-making on whether to compress or not. For signal segments that need to be compressed, an LSTM encoder-decoder network is used for nonlinear compression and restoration, which improves the compression rate while maintaining spectral fidelity and phase consistency. For parts that do not need to be compressed, they are retained as is, taking into account both real-time performance and error control.

[0040] 2. This invention achieves accurate multimodal characterization of harmonics through multi-source features, adaptive weighting, variance fusion, and frequency domain clustering modeling, thereby enabling dynamic evaluation of reconstruction quality and providing reliable data support for power quality monitoring, harmonic pollution diagnosis, and power electronic control.

[0041] 3. This invention achieves high-fidelity recovery and quantifiable evaluation of broadband random harmonic signals under high compression rate and low redundancy transmission conditions through a three-in-one closed-loop mechanism of "compression decision-LSTM reconstruction-multi-feature evaluation". It does not require relying solely on fixed threshold compression, thus reducing interpolation restoration errors and enhancing the dynamics and adaptability of evaluation. Attached Figure Description

[0042] Figure 1 The flowchart shows a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion.

[0043] Figure 2 This is a comparison chart of the original signal and the compressed signal;

[0044] Figure 3 A comparison chart showing the amplitude and multi-band aggregation of the original and reconstructed signals;

[0045] Figure 4The results for the six evaluation indicators are shown in the graph;

[0046] Figure 5 This is a diagram showing the result of multidimensional feature fusion.

[0047] Figure 6 This is a schematic diagram of a broadband random harmonic compression analysis device based on LSTM and multi-feature fusion.

[0048] Figure 7 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0049] This invention proposes a broadband random harmonic compression analysis method based on LSTM and multi-feature fusion. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0050] Figure 1 The flowchart of the broadband random harmonic compression analysis method based on LSTM and multi-feature fusion is shown below, specifically including:

[0051] 1. Acquire broadband random harmonic signals, perform time-domain analysis on the signals, and extract the signal period.

[0052] The signal sampling rate is F s The fundamental frequency is f0, and the number of sampling points per cycle is N. c for:

[0053] ;

[0054] Then the time-domain discrete sampling sequence x of the nth sampling point in the cth period c [n] is defined as:

[0055] ;

[0056] Discrete sampling sequences in the time domain conduct The amplitude spectrum X at the k-th frequency point obtained by the discrete Fourier transform of the point c [k] is defined as:

[0057] ;

[0058] Corresponding frequency for:

[0059] ;

[0060] Selecting high frequency bands , This is the lower limit of frequency. This is the upper limit of the frequency.

[0061] The high-frequency energy E in cycle c cDefined as:

[0062] ;

[0063] 2. Use LSTM to determine whether the signal in each period segment can be compressed or retained.

[0064] For all periodic energy sets Calculate the mean and standard deviation :

[0065] ;

[0066] ;

[0067] Then adaptively determine the threshold for:

[0068] ;

[0069] It is a dynamic adjustment factor.

[0070] The high-frequency energy from the past K periods is used as the input sequence S of the LSTM. c :

[0071]

[0072] Label From high-frequency energy E c Determined by comparison with a threshold:

[0073]

[0074] 0 indicates that the original signal should be preserved, and 1 indicates that the signal needs to be compressed;

[0075] LSTM discriminator learns the mapping:

[0076] ;

[0077] ;

[0078] in The output of the LSTM layer, For class probability vectors, This is the output layer weight matrix. is the output layer bias vector, and Softmax is the activation function in the neural network.

[0079] Final judgment result for:

[0080] ;

[0081] 3. Downsampling and encoding compression are performed on compressible periodic signals.

[0082] For an original periodic signal with N sampling points, Y sampling points are retained after compression. The compressed time-domain sequence is then input into an LSTM-based encoder model to extract nonlinear dynamic features for deep compression.

[0083] If it is determined to be compression, then sub-sampling is performed according to the downsampling factor d, let:

[0084] ;

[0085] in Original signal sequence The new sequence obtained by downsampling according to the sampling factor d This represents the compressed signal sequence, where y represents the index range of the downsampled signal sequence, and N... c It is the length of the original signal sequence.

[0086] Fourth, the downsampled and encoded compressed signal is decoded and reconstructed using an LSTM decoder.

[0087] The training decoder z will compress the signal Reconstruct the original length :

[0088] ;

[0089] The training target is usually the mean squared error. :

[0090] ;

[0091] The full-length signal is obtained by sequentially piecing together the reconstruction results from each cycle. :

[0092] ;

[0093] in Describes the period c of the signal. The index positions of N samples c It is the number of samples within a period. It is the sample index within the current period.

[0094] Fifth, preprocess the decoded and reconstructed signal, extract multidimensional features of the signal, and normalize the evaluation index of the multidimensional features.

[0095] Construct the time vector t:

[0096] ;

[0097] The signal sampling frequency is F. S The total signal duration is T. total Define the sliding window length L W Number of sampling points within the window :

[0098] ;

[0099] Let the sampling overlap rate be φ, and calculate the frame shift H:

[0100] ;

[0101] The decoded and reconstructed signal is segmented according to time windows. Within each time window, the Hilbert transform is used to obtain the analytical form of the signal, and the instantaneous amplitude and phase information are extracted.

[0102] The total number of sampling points for reconstructing the discrete sequence signal x[n] is N, and the number of sampling points within the sliding window is N. w Total frames (M):

[0103] ;

[0104] Where H is the frame shift size.

[0105] The signal for frame 𝑖 is:

[0106] ;

[0107] in This represents the nth sampling point in frame n, where i is the frame index.

[0108] The time center point t of each frame i :

[0109] ;

[0110] Perform a Hilbert transform on each frame of the signal to obtain the analytic signal z. i [n]:

[0111] ;

[0112] in This represents the Hilbert transform.

[0113] Calculate instantaneous amplitude and phase :

[0114] ;

[0115] For each scale of the signal, the following six-dimensional characteristic index is calculated to characterize the multidimensional features of harmonic disturbances.

[0116] (1) Amplitude standard deviation (ASD) reflects the amplitude and transient intensity of the disturbance:

[0117] ;

[0118] ;

[0119] in, The mean amplitude;

[0120] (2) The spectral centroid (SC) describes the frequency shift trend:

[0121] Calculate the power spectral density:

[0122] ;

[0123] ;

[0124] ;

[0125] in, For the corresponding frequency, k represents the frequency point, and N... FFT It is the number of points in the Fast Fourier Transform. P is a very small constant. i [k] represents the calculated power spectral density of the i-th frame signal at frequency point k.

[0126] (3) Spectral flatness (SF) describes the flatness of the power spectrum of a harmonic signal:

[0127] ;

[0128] in , It is a very small constant;

[0129] (4) Phase coherence (PC) describes the consistency of the phase relationship between the frequency components of a harmonic signal:

[0130] ;

[0131] (5) Energy kurtosis (EK) describes the steepness of the energy distribution of a harmonic signal:

[0132] Calculate the energy sequence:

[0133] ;

[0134] ;

[0135] ;

[0136] in, E is the energy mean. i [n] represents the energy intensity of the nth sampling point of the i-th frame signal;

[0137] (6) Spectral sparsity (SS) describes the degree of concentration of spectral energy:

[0138] ;

[0139] The dimensions and numerical ranges of different features vary considerably. To ensure comparability, feature normalization is performed, and each feature vector is subjected to max-min normalization:

[0140] ;

[0141] in, This is the original calculated value of the j-th feature index in frame i; A very small constant is used to prevent the denominator from being zero.

[0142] VI. Calculate the dynamic fusion index and signal compression rate using a variance-mixed weighting mechanism.

[0143] By analyzing the sensitivity of features under different time windows, the fusion weights of each feature are dynamically adjusted. The historical window length L is set to balance the ability to capture instantaneous changes and long-term trends.

[0144] ;

[0145] Where M is the total number of frames in the signal time window;

[0146] For the current frame i, consider the history window [i-L+1, i]:

[0147] Define the historical feature matrix:

[0148] ;

[0149] Calculate the variance vector V(i) of each feature within the historical window:

[0150] ;

[0151] ;

[0152] ;

[0153] in Let the variance of feature j within the historical window be denoted as . Let be the mean value of feature j within the history window, and g be the starting value of the history window frame.

[0154] In this embodiment, based on the normalized six indicators, a variance mixed weighting mechanism is introduced to dynamically adjust the fusion contribution of each scale:

[0155] Adaptive weight formula:

[0156] ;

[0157] in, This is element-wise multiplication; This is the weighting adjustment coefficient.

[0158] Introducing memory factors :

[0159] ;

[0160] in, To control the memory strength of historical weights; To control the regression strength of the initial weights; w init To initialize the weight parameters.

[0161] Weight normalization:

[0162] ;

[0163] in, The adaptive dynamic weights for the i-th frame; Let be the weight of the j-th indicator in the i-th frame;

[0164] Finally, the dynamic fusion index is calculated:

[0165] ;

[0166] Among them, t i Let i be the time center point of the i-th frame. The adaptive dynamic weight of the j-th index in the i-th frame. This is the normalized result of the j-th index in the i-th frame.

[0167] Result evaluation and compression ratio calculation:

[0168] ;

[0169] in To preserve periodic sets, It is a compressed set of cycles, and dynamically displays TSHMI(t) and its sub-characteristics, which can be used for power quality anomaly early warning, harmonic disturbance location and trend analysis.

[0170] 7. Dynamically display the evaluation indicators of multi-dimensional features and the dynamic fusion indicators. Record the compression retention decision results, count the number of data points reduced in the compression cycle, and calculate the compression ratio in combination with the original signal points; dynamically display TSHMI(t) and sub-feature indicators, which can be used for power quality anomaly early warning, harmonic disturbance location, and trend analysis.

[0171] A test signal was generated by superimposing a standard sine wave (as the fundamental wave) with broadband random noise for computational verification. The original data was then subjected to a Fast Fourier Transform, and the signal was compressed and recovered based on a set high-frequency energy threshold. Figure 2 The comparison between the original and compressed signals is presented. It can be seen that, while maintaining the main harmonic structure and broadband random perturbation pattern essentially the same, the data size is significantly reduced, with an overall compression rate of 28%. Furthermore, to quantitatively evaluate the impact of the compression process on the spectral characteristics, aggregated statistical analysis is performed on the amplitude and error of the reconstructed signal in different frequency bands. Figure 3 The image shows a comparison of the amplitude and multi-band aggregation of the original and reconstructed signals. It can be observed that the amplitude and differences in each frequency band are generally small, indicating that the compression and recovery process did not significantly damage the main spectral characteristics of the broadband random harmonics. This verifies the effectiveness of the method in achieving effective compression while still maintaining the harmonic signal characteristics well. A Hilbert transform is performed on the compressed and recovered random harmonic signal, and multi-dimensional dynamic evaluation indices are calculated. Figure 4 The graph shows the evolution of harmonic signals in the same time period after the recovery of the six core feature indicators (amplitude standard deviation ASD, spectral centroid SC, spectral flatness SF, phase coherence PC, energy kurtosis EK, and spectral sparsity SS) on which the dynamic weight adjustment algorithm is based. Figure 5 The image shows the comprehensive output result after adaptive weighting of six-dimensional features, which is derived from... Figure 4 The six-dimensional basic feature index is dynamically fused through adaptive weighting under the influence of the memory factor α. This adaptive mechanism can intelligently adjust the contribution weight of each feature based on the historical characteristics of the harmonic signal, ultimately generating a comprehensive index.

[0172] Figure 6 This is a schematic diagram of a broadband random harmonic compression analysis device based on LSTM and multi-feature fusion, including:

[0173] The period extraction module is used to acquire broadband random harmonic signals, perform time-domain analysis on the signals, and extract the period of the signals.

[0174] The signal determination module is used to determine whether the signal in each cycle segment can be compressed or retained using LSTM.

[0175] The encoding and compression module is used to downsample and encode compressible periodic signals;

[0176] The decoding and reconstruction module is used to decode and reconstruct the downsampled and encoded compressed signal using an LSTM decoder.

[0177] The feature processing module is used to preprocess the decoded and reconstructed signal, extract multi-dimensional features of the signal, and normalize the evaluation index of the multi-dimensional features.

[0178] The index calculation module is used to calculate the compression ratio of the dynamically fused index and signal using a variance-mixed weighting mechanism.

[0179] The dynamic display module is used to dynamically display the evaluation indicators of multi-dimensional features and the dynamic fusion indicators.

[0180] In summary, this invention achieves adaptive decision-making regarding compression, improving the compression ratio while maintaining spectral fidelity and phase consistency. Uncompressed portions are retained as is, balancing real-time performance and error control. Furthermore, it enables accurate multi-mode characterization of harmonics, allowing for dynamic evaluation of reconstruction quality and providing reliable data support for power quality monitoring, harmonic pollution diagnosis, and power electronic control.

[0181] This embodiment also includes an electronic device and a storage medium. Figure 7 This is a schematic diagram of the electronic device of the present invention. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of the broadband random harmonic compression analysis method based on LSTM and multi-feature fusion. The computer program is stored on the storage medium, and when the computer program is executed by the processor, it implements the various steps of the broadband random harmonic compression analysis method based on LSTM and multi-feature fusion.

[0182] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0183] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0184] 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.

[0185] 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.

[0186] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0187] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A broadband random harmonic compression analysis method based on LSTM and multi-feature fusion, characterized in that, Includes the following steps: Acquire broadband random harmonic signals, perform time-domain analysis on the signals, and extract the signal period; Use LSTM to determine whether the signal in each cycle segment is compressible or can be preserved. Downsampling and encoding compression are performed on compressible periodic signals; The downsampled and coded compressed signal is decoded and reconstructed using an LSTM decoder; The decoded and reconstructed signal is preprocessed to extract multidimensional features of the signal, and the evaluation index of the multidimensional features is normalized. The dynamic fusion index and signal compression ratio are calculated using a variance-mixed weighting mechanism; The evaluation indicators of multidimensional features are dynamically displayed along with the dynamically fused indicators.

2. The broadband random harmonic compression analysis method based on LSTM and multi-feature fusion according to claim 1, characterized in that, The specific steps for downsampling and encoding compression of compressible periodic signals are as follows: For an original periodic signal with N sampling points, M sampling points are retained after compression. The compressed time-domain sequence is then input into an LSTM-based encoder model to extract nonlinear dynamic features for deep compression.

3. The broadband random harmonic compression analysis method based on LSTM and multi-feature fusion according to claim 1, characterized in that, The preprocessing of the decoded and reconstructed signal is as follows: The decoded and reconstructed signal is segmented according to time windows. Within each time window, the Hilbert transform is used to obtain the analytical form of the signal, and the instantaneous amplitude and phase information are extracted.

4. The broadband random harmonic compression analysis method based on LSTM and multi-feature fusion according to claim 3, characterized in that, The time window has a length of 40ms.

5. The broadband random harmonic compression analysis method based on LSTM and multi-feature fusion according to claim 4, characterized in that, Set an overlap rate of 50% to 80% between adjacent time windows.

6. The broadband random harmonic compression analysis method based on LSTM and multi-feature fusion according to claim 1, characterized in that, Evaluation metrics for multidimensional features include: amplitude standard deviation, spectral centroid, spectral flatness, phase coherence, energy kurtosis, and spectral sparsity.

7. The broadband random harmonic compression analysis method based on LSTM and multi-feature fusion according to claim 6, characterized in that, The calculation formula for the dynamic fusion index is as follows: ； in, For time, The dynamic weight of the j-th indicator. Let j be the j-th evaluation index after normalization.

8. A broadband random harmonic compression analysis device based on LSTM and multi-feature fusion, characterized in that, include: The period extraction module is used to acquire broadband random harmonic signals, perform time-domain analysis on the signals, and extract the period of the signals. The signal determination module is used to determine whether the signal in each cycle segment can be compressed or retained using LSTM. The encoding and compression module is used to downsample and encode compressible periodic signals; The decoding and reconstruction module is used to decode and reconstruct the downsampled and encoded compressed signal using an LSTM decoder. The feature processing module is used to preprocess the decoded and reconstructed signal, extract multi-dimensional features of the signal, and normalize the evaluation index of the multi-dimensional features. The index calculation module is used to calculate the compression ratio of the dynamically fused index and signal using a variance-mixed weighting mechanism. The dynamic display module is used to dynamically display the evaluation indicators of multi-dimensional features and the dynamic fusion indicators.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the broadband random harmonic compression analysis method based on LSTM and multi-feature fusion as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the broadband random harmonic compression analysis method based on LSTM and multi-feature fusion as described in any one of claims 1 to 7.

Images