CN-121978416-A - Lightning whistle sound wave identification method and system based on multi-scale time-frequency fusion

CN121978416ACN 121978416 ACN121978416 ACN 121978416ACN-121978416-A

Abstract

The invention provides a lightning whistle sound wave identification method and a system based on multi-scale time-frequency fusion, which are used for carrying out self-adaptive window length segmentation on a broadband very low frequency sequence, calculating short-time Fourier transform under various time-frequency resolutions, constructing a composite time-frequency image by time-scale bi-dimensional splicing, taking long-time evolution and short-time details into consideration, introducing a self-supervision gating attention-suppressing noise mechanism, learning noise statistical characteristics from historical observation, generating a spectrogram noise mask, carrying out gating fusion on a feature layer, realizing self-adaptive suppression of non-stationary noise, and highlighting whistle trace characteristics. And (3) performing target detection on the enhanced composite spectrogram by adopting an improved lightweight YOLO network, and outputting the time-frequency boundary and main characteristic parameters of the whistle event. The invention can realize the high-efficiency automatic identification and fine statistics of lightning whistle events in large-scale very low frequency observation data, and provides technical support for related applications such as space environment monitoring based on lightning whistle.

Inventors

NI BINBIN
SONG XUCHENG
GUO SIMING
GU XUDONG
XU WEI
WANG SHIWEI
ZHANG BOWEN
Chen Zhishe
LIN HONGWEI
LIU SHICHENG

Assignees

武汉大学

Dates

Publication Date: 20260505
Application Date: 20260108

Claims (10)

1. A lightning whistle wave identification method based on multi-scale time-frequency fusion is characterized by comprising the following steps: Constructing a very low frequency electromagnetic wave detection station, and collecting a very low frequency electromagnetic wave signal in real time; Preprocessing the time sequence signal of the very low frequency electromagnetic wave signal by utilizing the very low frequency electromagnetic wave observation data to obtain a preprocessed time sequence signal; based on whistle wave characteristics, performing multi-scale video segmentation and multi-scale spectrogram splicing on the preprocessed time sequence signals to obtain a composite time-frequency spectrogram; after self-supervision noise modeling and gating enhancement are adopted for the composite time-frequency spectrogram, recognition model training is carried out to output a plurality of recognition results; storing a plurality of identification results and calculating related physical parameters; and updating the pseudo tag of the related physical parameter manually on line, optimizing the parameter by adopting a preset model compression technology, outputting a lightning whistle wave identification result and carrying out visual analysis.
2. The method for identifying lightning whistle waves based on multi-scale time-frequency fusion according to claim 1, wherein preprocessing the time sequence signal of the very low frequency electromagnetic wave signal comprises DC removal, power frequency and pulse interference suppression, band-pass filtering and amplitude normalization.
3. The lightning whistle wave identification method based on multi-scale time-frequency fusion according to claim 1, wherein the steps of performing multi-scale video segmentation and multi-scale spectrogram splicing on the preprocessed time sequence signal based on whistle wave characteristics to obtain a composite time-frequency spectrogram comprise: performing self-adaptive segmentation on the preprocessed time sequence signals according to signal energy distribution to obtain a plurality of segmented time sequence signals; determining a preset short window for capturing rapid dispersion details, and a preset long window for extracting a span dispersion trend; respectively adopting a preset short window and a preset long window to carry out short-time Fourier transform on each segmented time sequence signal, and outputting a multi-scale video image; splicing the multi-scale video pictures according to the time sequence and the scale sequence to construct a composite time-frequency spectrogram; And extracting whistle signals based on whistle frequency domain features, and acquiring whistle feature parameters.
4. The lightning whistle wave identification method based on multi-scale time-frequency fusion according to claim 1, wherein after self-supervision noise modeling and gating enhancement are adopted for the composite time-frequency spectrogram, an identification model training is performed to output a plurality of identification results, and the method comprises the following steps: performing median filtering, fractional bit estimation and stability judgment on the background power distribution of each frequency segment to obtain a two-dimensional weight matrix, wherein the two-dimensional weight matrix is used for marking power frequency lines, polar stripe noise and broadband pulse interference, the two-dimensional weight matrix is used as a spectrogram noise mask, the spectrogram noise mask is used for reflecting the probability of whether each point is a non-whistle background, and the spectrogram noise mask is used as a priori prompt to be provided for a network in self-supervision training; In the feature extraction stage, automatically estimating and generating gating weight by a lightweight attention branch according to local time-frequency gradient, signal sparsity and adjacent scale consistency, wherein the gating weight is used for representing the importance of a current region on whistle recognition; The spectrogram noise mask and the gating weight are subjected to point-by-point fusion in a feature layer to form a denoised enhancement spectrogram; Inputting the denoised enhancement spectrogram into an improved lightweight YOLO11 network, and outputting a lightning whistle candidate frame and time-frequency parameters; Taking the denoised enhanced spectrogram and whistle characteristic parameters as weak labels, pre-training the recognition model until the recognition accuracy exceeds a preset accuracy threshold, and outputting an initial recognition result; if the initial recognition result is consistent with the preset whistle physical characteristics, the initial recognition result is saved as a final recognition result, otherwise, the initial recognition result is added into a negative event set to be trained again.
5. The multi-scale time-frequency fusion-based lightning whistle wave identification method according to claim 1, wherein storing a plurality of identification results and calculating relevant physical parameters comprises: the plurality of recognition results includes arrival time, first-order dispersion cutoff frequency, start-stop frequency, and duration; The relevant physical parameters include peak and average power, signal to noise ratio, dispersion parameters, and candidate box confidence.
6. The lightning whistle wave identification method based on multi-scale time-frequency fusion according to claim 1, wherein the artificial online updating of pseudo tags of related physical parameters, the parameter optimization by adopting a preset model compression technology, the output of the lightning whistle wave identification result and the visual analysis comprise: Automatically screening candidate events by using a high confidence threshold, pushing the screened events to a labeling interface, manually clicking a preset operation frame, writing back the passed or modified labels as pseudo labels, and synchronously generating a difficult case/negative event list for difficult sample mining of subsequent training; Performing online fine adjustment on the updated pseudo tag by adopting a preset index; The method comprises the steps of utilizing quantization and pruning to keep edge real-time change, including carrying out INT8 or mixed precision quantization on an identification network, carrying out structured pruning with the proportion larger than a preset proportion and sparsifying a detection head, and meeting real-time constraint that the inference time delay of a single composite spectrogram is smaller than a preset time delay threshold on edge equipment; And writing the final confirmation event record corresponding to the lightning whistle wave identification result into a database or object storage in a preset structural format, wherein the final confirmation event record comprises a plurality of preset fields.
7. Lightning whistle wave identification system based on multiscale time-frequency fusion, characterized by comprising: the acquisition module is used for constructing a very low frequency electromagnetic wave detection station and acquiring a very low frequency electromagnetic wave signal in real time; The preprocessing module is used for preprocessing the time sequence signal of the very low frequency electromagnetic wave signal by utilizing the very low frequency electromagnetic wave observation data to obtain a preprocessed time sequence signal; the splicing module is used for carrying out multi-scale video segmentation and multi-scale spectrogram splicing on the preprocessed time sequence signals based on whistle sound wave characteristics to obtain a composite time-frequency spectrogram; The training module is used for training the recognition model to output a plurality of recognition results after the self-supervision noise modeling and gating enhancement are adopted on the composite time-frequency spectrogram; the storage module is used for storing a plurality of identification results and calculating related physical parameters; and the updating module is used for updating the pseudo tag of the related physical parameter manually on line, optimizing the parameter by adopting a preset model compression technology, outputting a lightning whistle wave identification result and carrying out visual analysis.
8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements a lightning whistle wave identification method based on multi-scale time-frequency fusion according to any of claims 1 to 6 when executing the program.
9. A non-transitory computer readable storage medium, having stored thereon a computer program, which when executed by a processor implements a lightning whistle wave identification method based on multi-scale time-frequency fusion according to any of claims 1 to 6.
10. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements a lightning whistle wave identification method based on multi-scale time-frequency fusion according to any of claims 1 to 6.

Description

Lightning whistle sound wave identification method and system based on multi-scale time-frequency fusion Technical Field The invention relates to the technical field of very low frequency electromagnetic wave detection and application, in particular to a lightning whistle sound wave identification method and system based on multi-scale time-frequency fusion. Background Lightning whistle waves are typically dispersive electromagnetic fluctuations formed by ground lightning radiation of very low frequency (VLF, 3-30 kHz) energy coupled into the ionosphere-magnetic layer environment, propagating along the geomagnetic line and re-incident to the ground in the magnetically conjugate region. The waveform characteristics of the novel lightning-induced whistle are essentially different from whistle type fluctuation (such as chorus, hiss, EMIC wave band structure) naturally generated in the magnetic layer, wherein natural whistle is mostly generated by the action of magnetic layer electrons and fluctuating microwave particles, the frequency structure is complex, the dispersion characteristic is weak or multi-band, and the lightning-induced whistle presents an obvious single-band dispersion track with the frequency rapidly descending along with time. The existing lightning whistle recognition mostly adopts methods such as manual visual time-frequency diagram, fixed threshold or template matching. The fixed threshold/template method has low adaptability to low signal-to-noise ratio, non-stationary noise and various dispersion tracks in polar scenes, and is easy to cause missed detection and false detection. In recent years, convolution networks or universal target detectors based on single-scale spectrograms are used for whistle recognition, but three types of limitations generally exist, namely, firstly, single-scale time-frequency representation is difficult to simultaneously consider short-time detail and long-time dispersion trend, weak signals and background are difficult to distinguish, secondly, explicit supervision dependence is strong, model generalization capability and long-term stability are insufficient under the conditions that polar data labeling is scarce and distribution drifts along with seasons/local time/magnetic activity, thirdly, deployment constraint is obvious, model volume and calculation force consumption are high, and low-power consumption and near-real-time operation are difficult to realize on edge acquisition equipment. Meanwhile, lightning whistle wave recognition is limited by a plurality of practical factors, namely background noise of the lightning whistle wave recognition is obviously changed along with time, seasons and station environments, an offline training model is easy to attenuate in long-term operation, and the traditional recognition method lacks joint constraint on physical laws (such as dispersion slope, duration, frequency downlink track, dispersion parameters and the like) of the lightning whistle, so that the false recognition rate is difficult to reduce on the premise of not introducing a large amount of manual participation. In summary, there is a need for a solution that can (1) fuse and characterize weak whistle in multiple scales, (2) self-monitor noise suppression and gating enhancement under no or few labeling conditions, (3) support edge end real-time reasoning with lightweight structure, (4) perform secondary screening by combining with physical consistency rules, and (5) realize continuous self-adaptive automatic identification technology scheme through online pseudo-tags, so as to meet long-term, stable and efficient processing requirements of polar and high-noise VLF observation data. Disclosure of Invention The invention provides a lightning whistle sound wave identification method and a system based on multi-scale time-frequency fusion, which are used for solving the defects that in the prior art, the lightning whistle is difficult to automatically identify under the condition of a high-noise station, the stability of a long period is insufficient and the real-time performance of an edge end is limited, and realizing the processing of Very Low Frequency (VLF) electromagnetic observation data and the automatic identification of events under the high-noise environment. In a first aspect, the invention provides a lightning whistle wave identification method based on multi-scale time-frequency fusion, which comprises the following steps: Constructing a very low frequency electromagnetic wave detection station, and collecting a very low frequency electromagnetic wave signal in real time; Preprocessing the time sequence signal of the very low frequency electromagnetic wave signal by utilizing the very low frequency electromagnetic wave observation data to obtain a preprocessed time sequence signal; based on whistle wave characteristics, performing multi-scale video segmentation and multi-scale spectrogram splicing on the preprocessed time sequence signals to ob