Lightning whistle wave identification method and system based on multi-scale time-frequency fusion

Abstract

The invention provides a lightning whistle sound wave identification method and system based on multi-scale time-frequency fusion, and the method comprises the steps: carrying out the adaptive window length segmentation of a broadband very-low frequency sequence, calculating short-time Fourier transform under a plurality of time-frequency resolutions, carrying out the two-dimensional splicing of time and scale to construct a composite time-frequency image, giving consideration to long-time evolution and short-time details, and carrying out the recognition of lightning whistle sound waves. A self-supervised gating attention noise suppression mechanism is introduced, noise statistical characteristics are learned from historical observation, a spectrogram noise mask is generated, gating fusion is performed on a feature layer, self-adaptive suppression of non-stationary noise is realized, and whistle trace features are highlighted. And carrying out target detection on the enhanced composite spectrogram by adopting an improved lightweight YOLO network, and outputting a time-frequency boundary and main characteristic parameters of a whistle event. According to the method, efficient automatic identification and fine statistics of lightning whistle events can be realized in large-scale very-low-frequency observation data, and technical support is provided for related applications such as space environment monitoring based on lightning whistles.

Description

Technical Field

[0001] This invention relates to the field of very low frequency electromagnetic wave detection and application technology, and in particular to a lightning whistle sound wave identification method and system based on multi-scale time-frequency fusion. Background Technology

[0002] Lightning whistling waves are typical dispersive electromagnetic waves formed when very low frequency (VLF, 3–30 kHz) energy radiated by ground lightning couples into the ionosphere-magnetosphere environment, propagates along the geomagnetic field lines, and re-impacts the ground in the magnetic conjugate region. Their waveform characteristics are fundamentally different from naturally occurring whistling waves within the magnetosphere (such as chorus, hiss, and EMIC band structures): natural whistling waves are mostly generated by the interaction of magnetosphere electrons with wave microwave particles, exhibiting a complex frequency structure and weak or multi-band dispersion; while lightning-induced whistling waves show a clear single-band dispersion trajectory with a rapidly declining frequency over time.

[0003] Existing methods for lightning whistle recognition primarily employ manual inspection of time-frequency maps, fixed thresholds, or template matching. Manual inspection is inefficient and highly subjective, unsuitable for long-term and large-volume data. Fixed threshold / template methods are ill-suited to the low signal-to-noise ratio, non-stationary noise, and diverse dispersion trajectories found in polar environments, easily leading to missed and false detections. In recent years, attempts have been made to use convolutional networks or general target detectors based on single-scale spectrograms for whistle recognition, but these generally suffer from three limitations: First, single-scale time-frequency representations struggle to simultaneously capture short-term details and long-term dispersion trends, making it difficult to distinguish weak signals from the background. Second, they rely heavily on explicit supervision, resulting in insufficient model generalization and long-term stability in situations where polar data annotations are scarce and distribution shifts with seasons, local time, and magnetic activity. Third, deployment constraints are significant, with high model size and computational consumption, making it difficult to achieve low-power, near-real-time operation on edge acquisition devices.

[0004] Meanwhile, lightning whistle sound wave recognition is also limited by several practical factors: its background noise varies significantly with time, season and station environment, and offline training models are prone to performance degradation during long-term operation; and traditional recognition methods lack joint constraints on the physical laws of lightning whistles (such as dispersion slope, duration, frequency downlink trajectory, dispersion parameters, etc.), making it difficult to reduce the false recognition rate without introducing a large amount of manual intervention. In summary, there is an urgent need for an automatic recognition technology that can: (1) multi-scale fusion characterize weak whistles; (2) self-supervised noise suppression and gating enhancement under unlabeled or poorly labeled conditions; (3) support real-time inference at the edge with a lightweight structure; (4) perform secondary screening by combining physical consistency rules; and (5) achieve continuous adaptive automatic recognition through online pseudo-labels, so as to meet the long-term, stable and efficient processing needs of polar and high-noise VLF observation data. Summary of the Invention

[0005] This invention provides a lightning whistle sound wave identification method and system based on multi-scale time-frequency fusion, which solves the defects of existing technologies such as difficulty in automatic identification of lightning whistles under high-noise station conditions, insufficient long-period stability, and limited real-time performance at the edge end, and realizes very low frequency (VLF) electromagnetic observation data processing and automatic event identification in high-noise environments.

[0006] In a first aspect, the present invention provides a lightning whistle sound wave identification method based on multi-scale time-frequency fusion, comprising: Construct a very low frequency electromagnetic wave detection station to collect very low frequency electromagnetic wave signals in real time; Using very low frequency electromagnetic wave observation data, the time series signal of the very low frequency electromagnetic wave signal is preprocessed to obtain the preprocessed time series signal. Based on the characteristics of whistle waves, multi-scale video segmentation and multi-scale spectrogram splicing are performed on the preprocessed time-series signal to obtain a composite time-spectrum. After applying self-supervised noise modeling and gated enhancement to the composite time-spectrum graph, the recognition model is trained and multiple recognition results are output. Save multiple recognition results and calculate relevant physical parameters; The pseudo-labels of relevant physical parameters are updated manually online, and the parameters are optimized using a preset model compression technique. The results of lightning whistle sound wave recognition are then output and visualized.

[0007] The present invention provides a lightning whistle sound wave identification method based on multi-scale time-frequency fusion, which preprocesses the time sequence signal of very low frequency electromagnetic wave signal, including DC removal, power frequency and pulse interference suppression, bandpass filtering and amplitude normalization.

[0008] According to the present invention, a lightning whistle sound wave identification method based on multi-scale time-frequency fusion is provided. This method performs multi-scale video segmentation and multi-scale spectrogram stitching on a preprocessed time-series signal based on whistle sound wave features to obtain a composite time-spectrum image, including: The preprocessed time sequence signal is adaptively segmented according to the signal energy distribution to obtain multiple segmented time sequence signals. A preset short window is used to capture fast dispersion details, and a preset long window is used to extract cross-segment dispersion trends. Short-time Fourier transforms are performed on each segment of the time-series signal using preset short windows and preset long windows, respectively, to output multi-scale video images; Multi-scale video images are stitched together according to chronological and scale order to construct a composite time-spectrum image; Whistle signal is extracted based on whistle frequency domain features to obtain whistle feature parameters.

[0009] According to the present invention, a lightning whistle sound wave recognition method based on multi-scale time-frequency fusion is provided. After applying self-supervised noise modeling and gated enhancement to the composite time-frequency spectrum, the recognition model is trained and multiple recognition results are output, including: Median filtering, fractional estimation, and stability assessment are performed on the background power distribution of each frequency segment to obtain a two-dimensional weight matrix. The two-dimensional weight matrix is ​​used to label power line, polar stripe noise, and broadband pulse interference. The two-dimensional weight matrix is ​​used as a spectral noise mask, which reflects the probability of whether each point is a non-whistle background. The spectral noise mask is provided to the network as a priori cue during self-supervised training. During the feature extraction stage, a lightweight attention branch automatically estimates and generates gate weights based on local time-frequency gradients, signal sparsity, and adjacent scale consistency. These gate weights are used to represent the importance of the current region for whistle recognition. The spectral noise mask and the gating weights are fused point by point at the feature layer to form a denoised enhanced spectrum. The denoised enhanced spectrum is input into the improved lightweight YOLO11 network to output lightning whistle candidate boxes and time-frequency parameters; The enhanced spectrogram after denoising and the whistle feature parameters are used as weak labels to pre-train the recognition model until the recognition accuracy exceeds the preset accuracy threshold, and the initial recognition result is output. If the initial recognition result is determined to be consistent with the preset whistle physical characteristics, the initial recognition result is saved as the final recognition result; otherwise, the initial recognition result is added to the negative event set for retraining.

[0010] According to the present invention, a lightning whistle sound wave identification method based on multi-scale time-frequency fusion is provided, which saves multiple identification results and calculates relevant physical parameters, including: The multiple recognition results include arrival time, first-order dispersion cutoff frequency, start and end frequencies, and duration. The relevant physical parameters include peak and average power, signal-to-noise ratio, dispersion parameters, and candidate box confidence.

[0011] According to the present invention, a lightning whistle sound wave recognition method based on multi-scale time-frequency fusion is provided. The method involves manually updating pseudo-labels of relevant physical parameters online, optimizing parameters using a preset model compression technique, outputting lightning whistle sound wave recognition results, and performing visualization analysis. The method includes: Candidate events are automatically filtered using a high confidence threshold. The filtered events are then pushed to the annotation interface, where manual selection of preset operation boxes allows the approved or modified annotations to be written back as pseudo-labels. Simultaneously, a list of difficult / negative events is generated for mining difficult samples in subsequent training. The updated pseudo-labels are fine-tuned online using preset indicators; Real-time edge changes are maintained by using quantization and pruning, including INT8 or mixed precision quantization of the recognition network, structural pruning with a ratio greater than a preset ratio and sparsification of the detection head, and satisfying the real-time constraint that the inference delay of a single composite spectrum is less than a preset delay threshold on the edge device. The final confirmation event record corresponding to the lightning whistle sound wave recognition result is written to the database or object storage in a preset structured format, containing multiple preset fields.

[0012] Secondly, the present invention also provides a lightning whistle sound wave recognition system based on multi-scale time-frequency fusion, comprising: The acquisition module is used to build a very low frequency electromagnetic wave detection station and acquire very low frequency electromagnetic wave signals in real time. The preprocessing module is used to preprocess the time series signal of the very low frequency electromagnetic wave signal using very low frequency electromagnetic wave observation data to obtain the preprocessed time series signal. The splicing module is used to perform multi-scale video segmentation and multi-scale spectrogram splicing on the preprocessed time-series signal based on the characteristics of the whistle wave, so as to obtain a composite time-spectrum. The training module is used to train the recognition model and output multiple recognition results after applying self-supervised noise modeling and gated enhancement to the composite time-spectrum graph; The storage module is used to save multiple recognition results and calculate relevant physical parameters; The update module is used for manual online updating of pseudo-labels for relevant physical parameters. It uses preset model compression technology to optimize parameters, outputs lightning whistle sound wave recognition results, and performs visualization analysis.

[0013] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion as described above.

[0014] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion as described above.

[0015] This invention provides a method and system for identifying lightning whistle sound waves based on multi-scale time-frequency fusion. For VLF observation data in high-noise environments, it first performs multi-scale time-frequency stitching and self-supervised gating noise suppression. Then, a lightweight detection head enables rapid localization and parameter extraction of weak traces. A secondary screening is performed using physical consistency constraints such as dispersion slope, duration, and skeleton residuals. Through online pseudo-label updating and quantization / pruning synergy, low-power, long-term adaptive, and stable operation is achieved at the edge. This invention offers high recognition accuracy, low false alarm rate, flexible deployment, and scalability. The false detection rate of the trained model is less than 5%, while the false detection rate of traditional machine learning methods reaches 25.07%. The obtained high-confidence whistle events and dispersion features can directly serve the subsequent inversion of duct characteristics and plasma parameters, and support long-term automated monitoring and statistical analysis of magnetosphere-ionosphere coupling processes, possessing significant application value in space environment research. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is one of the flowcharts of the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion provided by the present invention; Figure 2 This is the second flowchart of the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion provided by the present invention; Figure 3 This is a flowchart of the recognition model provided by the present invention; Figure 4 This is a comparison diagram of the composite time-frequency diagram provided by this invention and the original time-frequency diagram; Figure 5 This is a result image of a whistle wave automatically identified in an embodiment provided by the present invention; Figure 6 This is a partial data result diagram obtained from the detection provided by the present invention; Figure 7 This is a time distribution diagram of the whistle wave detected by the present invention; Figure 8 This is a schematic diagram of the structure of the lightning whistle sound wave recognition system based on multi-scale time-frequency fusion provided by the present invention; Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] Figure 1 This is one of the flowcharts illustrating the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion provided in this embodiment of the invention, such as... Figure 1 As shown, it includes: Step 100: Construct a very low frequency electromagnetic wave detection station to collect very low frequency electromagnetic wave signals in real time; Step 200: Using very low frequency electromagnetic wave observation data, preprocess the time sequence signal of the very low frequency electromagnetic wave signal to obtain the preprocessed time sequence signal. Step 300: Based on the characteristics of the whistle wave, perform multi-scale video segmentation and multi-scale spectrogram stitching on the preprocessed time-series signal to obtain a composite time-spectrum map; Step 400: After applying self-supervised noise modeling and gated enhancement to the composite time-spectrum graph, train the recognition model and output multiple recognition results; Step 500: Save multiple recognition results and calculate relevant physical parameters; Step 600: The pseudo-labels of relevant physical parameters are updated manually online, the parameters are optimized using preset model compression technology, and the lightning whistle sound wave recognition results are output and visualized.

[0020] Specifically, this embodiment of the invention uses lightning whistles as the detection target. First, the continuous VLF observation sequence is subjected to multi-window adaptive segmentation and multi-scale STFT processing, and then stitched together in chronological and scale order to form a composite time-frequency image. Then, self-supervised gating attention is used to learn the statistical characteristics of background noise on unlabeled historical data, generating noise masks and gating weights, and fusing them point by point in the feature layer to suppress stripe, power frequency, and pulse interference, resulting in an enhanced spectrum. On this basis, an improved lightweight YOLO detection head is introduced to locate the whistle candidate box in the enhanced spectrum and output time-frequency parameters. Combined with physical consistency constraints (frequency downslope, duration, dispersion skeleton fitting residual, SNR threshold, etc.), a second screening is performed to eliminate confusion caused by false recognition and waveguide reflection components. Finally, the validated events are parameterized, stored in the database, and visualized. Pseudo-labels are generated using high-confidence results and sampled samples. Online fine-tuning is performed with a small learning rate and early stopping strategy, while INT8 / mixed precision quantization and structured pruning ensure low-power near real-time operation at the edge.

[0021] like Figure 2As shown, it specifically includes: Step 1: Construct a very low frequency electromagnetic wave detection station to collect very low frequency electromagnetic wave signals in real time; Step 2: Using very low frequency electromagnetic wave observation data, preprocess the broadband very low frequency (VLF) electromagnetic wave time sequence signal continuously collected by the observation station. The preprocessing includes DC removal, power frequency and pulse interference suppression, bandpass filtering and amplitude normalization. Step 31: Adaptively segment the continuous VLF time-series signal according to the signal energy distribution, and generate a multi-scale segmented sequence using both short and long windows. The 5-s short window is used to capture fast dispersion details, and the 60-s long window is used to extract cross-segment dispersion trends. A Short-Time Fourier Transform (STFT) is performed on each segment. The short window uses NFFT=1024 and NOVERLAP=512, while the long window uses NFFT=8192 and window length Nwin=4096, thus obtaining a multi-scale time-frequency map with both high temporal and frequency resolution. Step 32: Stitch the time-frequency graphs according to chronological and scale order to construct a composite time-frequency image; Step 33: Extract the whistle signal based on the whistle frequency domain features and obtain whistle feature parameters, including whistle arrival time. First-order dispersion cutoff frequency Start and end frequency and duration ; Step 41: Utilizing long-term accumulated unlabeled very low frequency (VLF) observation data, perform self-supervised noise modeling and gating enhancement on the composite time-frequency image. First, based on the time-frequency statistical characteristics of historical spectra, construct a "spectral noise mask." This involves performing median filtering, quantile estimation, and stability assessment on the background power distribution of each frequency segment to obtain a two-dimensional weight matrix capable of labeling power line noise, polar stripe noise, and broadband impulse interference. This noise mask reflects the probability of each point becoming part of the "non-whistle background" (positions with high background probability have weights approaching 0.8-1, while low-noise areas have weights of approximately 0.1-0.3), and is provided to the network as a priori cue during self-supervised training. Subsequently, "gating weights" are generated during the feature extraction stage. The value is automatically estimated by a lightweight attention branch based on local time-frequency gradients, signal sparsity, and neighboring scale consistency. It typically ranges from 0 to 1 and represents the importance of the current region for whistle recognition. This is achieved by point-by-point fusion of the noise mask M and the gating weights G at the feature layer. This method achieves explicit suppression of background noise and enhancement of weak whistle trajectories, resulting in a denoised enhanced spectrum. Finally, the enhanced spectrum is input into an improved lightweight YOLO11 network to achieve rapid detection of lightning whistle candidate boxes and extraction of time-frequency parameters. Step 42: Using the enhanced spectrogram and the labeled candidates from Step 33 as weak labels, perform pre-training or continuous training of the recognition model until the recognition accuracy reaches 95% or more. Step 43: Determine whether the recognition result is consistent with the physical characteristics of the whistle and the dispersion of the time-lapse image. For example, does the typical duration meet the requirements? Check if the SNR is higher than the adaptive threshold; check if the frequency-time skeleton fitting residual meets the constraints, etc. If not, add the sample to the negative event set, continue training, and return to step 41; Step 5: Save all identified results, including at least: arrival time. First-order dispersion cutoff frequency Start and end frequency and duration The calculation of relevant physical parameters includes at least: peak and average power, signal-to-noise ratio (SNR), dispersion parameter D, and candidate box confidence. Step 61, Manual online pseudo-tag updates include: (1) Automatically filter candidate events using a high confidence threshold (e.g., ≥0.95); (2) Push the filtered events to the annotation interface, where manual selection of the pass / reject / modify boxes is required; (3) Write back the approved or modified annotations as pseudo-labels; (4) Simultaneously generate a list of difficult examples / negative events for mining difficult samples in subsequent training.

[0022] Online fine-tuning employs: a small learning rate, an early stopping strategy, a sliding window replay buffer, and an exponential moving average (EMA) to mitigate catastrophic forgetting; Quantization / pruning preserves real-time edge changes, including: (1) Perform INT8 or mixed precision quantization (PTQ / QAT) on the recognition network; (2) Perform ≥30% structured pruning and thin out the detection heads; (3) The real-time constraint of single composite spectrum inference delay ≤ 50 ms is met on the edge device; Step 62, Result Output and Archiving includes: The final confirmed event record will be written to a database or object storage in a structured format such as CSV / JSON / TXT, with fields including at least: Event ID, UTC time, station ID, and arrival time. First-order dispersion cutoff frequency Start and end frequency Duration Peak and average power, signal-to-noise ratio (SNR), dispersion parameter D, and candidate box confidence.

[0023] Based on the above embodiments, this invention will be further described in detail using actual observation data. This embodiment uses observation data acquired by a high-precision very low frequency (VLF) electromagnetic wave detection system deployed at an observation station (62.2°S, 59.0°W). This system continuously recorded broadband VLF electromagnetic wave signals from 00:00 to 24:00 (UTC) on August 6, 2023, at a sampling rate of 50 kHz, with uninterrupted recording throughout the day. This data serves as the input data for this method. This embodiment is only for demonstration purposes; the method of this invention is also applicable to VLF lightning whistle identification in any region of the world.

[0024] Step 1: Construct a very low frequency electromagnetic wave detection station (approximately 62.2°S, 59.0°W) to acquire very low frequency electromagnetic wave signals in real time; Step 2: Preprocess the acquired VLF timing signal. Perform the following on the raw timing data of the whole day of August 6, 2023: DC removal; power frequency and harmonic suppression; pulse interference (strong pulses commonly seen in polar stations) suppression; 3–30 kHz bandpass filtering; amplitude normalization and dynamic range compression.

[0025] A sequence with stable noise characteristics, suitable for time-frequency analysis, is obtained; Step 31: Adaptive energy segmentation and multi-scale STFT calculation based on station noise conditions. Considering the significant time-varying nature of polar background noise, the slice length is adaptively determined based on instantaneous energy and local signal-to-noise ratio. For each slice, STFT is calculated using both short and long windows (5s and 60s). The 5s short window is used to capture fast dispersion details, and the 60s long window is used to extract cross-segment dispersion trends. Short-time Fourier transform (STFT) is performed on each segment. The short window uses NFFT=1024 and NOVERLAP=512, while the long window uses NFFT=8192 and window length Nwin=4096, thus obtaining multi-scale time-frequency maps with multiple time and frequency resolutions. Step 32: Concatenate the multi-scale time-frequency images in chronological and scale order to form a composite time-frequency image. Short-window spectrograms are used to highlight the detailed trajectory of the rapidly descending lightning whistle, while long-window spectrograms are used to present the complete trend. A single composite time-frequency image can be obtained by stitching along the time axis, covering the typical 1-15kHz lightning whistle frequency band. Step 33: Extract the whistle signal based on the whistle frequency domain features and obtain whistle feature parameters, including whistle arrival time. First-order dispersion cutoff frequency Start and end frequency and duration ; Step 41: Perform self-supervised gating enhancement and spectral denoising using historical Antarctic Great Wall Station spectral data. A self-supervised pre-trained model is established based on a large number of unlabeled Antarctic spectra to learn the statistical patterns of polar fringe noise and impulse interference. First, based on the time-frequency statistical characteristics of historical Antarctic Great Wall Station spectral data, a "spectral noise mask" is constructed. This involves performing median filtering, quantile estimation, and stability assessment on the background power distribution of each frequency segment to obtain a two-dimensional weight matrix that can label power frequency lines, polar fringe noise, and broadband impulse interference. This noise mask reflects the probability of each point becoming part of the "non-whistle background" (positions with high background probability have weights approaching 0.8-1, while low-noise areas have weights of approximately 0.1-0.3), and is provided to the network as a priori cue during self-supervised training. Subsequently, "gating weights" are generated during the feature extraction stage. The value is automatically estimated by a lightweight attention branch based on local time-frequency gradients, signal sparsity, and neighboring scale consistency. It typically ranges from 0 to 1 and represents the importance of the current region for whistle recognition. This is achieved by point-by-point fusion of the noise mask M and the gating weights G at the feature layer. This method achieves explicit suppression of background noise and enhancement of weak whistle trajectories, resulting in a denoised enhanced spectrum. Finally, the enhanced spectrum is input into an improved lightweight YOLO11 network to achieve rapid detection of lightning whistle candidate boxes and extraction of time-frequency parameters. A noise mask is generated by combining gated attention and fused point-by-point at the feature layer to suppress power frequency residue, striped background, and strong impulses, significantly enhancing the downward trajectory of the lightning whistle. Step 42: Use a lightweight YOLO11 network to perform lightning whistle event recognition on the enhanced spectrogram. Input the enhanced composite time-frequency map into the improved YOLO network, and output the whistle candidate boxes and their time-frequency parameters. Use the parameters from step 33 as weak labels for continuous training until the recognition accuracy stabilizes at ≥95%. Step 43: Perform cross-segment verification and spurious event removal based on the lightning whistle physical consistency rule. For each candidate event, calculate whether the typical duration meets the requirement. Check if the SNR is higher than the adaptive threshold; check if the frequency-time skeleton fitting residual meets the constraints. If not, add it to the negative event set and return to step 41 to continue training; if it meets the constraints, output it as a real lightning whistle event. Step 5: Save all identified results, including at least: arrival time. First-order dispersion cutoff frequency Start and end frequency and duration The calculation of relevant physical parameters includes at least: peak and average power, signal-to-noise ratio (SNR), dispersion parameter D, and candidate box confidence. Step 61: Perform online pseudo-label updates and adaptive model maintenance based on the observation station scenario, through: False labels are automatically filtered based on a confidence level ≥ 0.95; manual batch verification and write-back are performed; a set of difficult examples and negative events is constructed; small learning rate, early stopping, and EMA sliding update are implemented; the model stability and adaptability under long-term operation in Antarctica are achieved. Step 62: Output the whistle event results for the day from the observation station. Store the final confirmed event records in CSV format, with fields including: Event ID, UTC time, and arrival time. First-order dispersion cutoff frequency Start and end frequency Duration The dispersion parameter D and the confidence level of the candidate boxes.

[0026] The original sequence is segmented into multiple time-frequency segments and STFTs are calculated and then stitched together to form a composite time-frequency graph. Self-supervised gated attention is used for spectral noise reduction and enhancement. The enhanced spectrum is then input into an improved lightweight YOLO detector head to complete whistle candidate recognition and parameter extraction. A secondary screening is performed using physical consistency constraints such as dispersion slope, duration, skeleton fitting residual, and SNR. The module structure of the recognition model is shown below. Figure 3 As shown.

[0027] Within the above processing framework, Figure 4 The invention showcases the significantly enhanced whistle spectrogram, target detection bounding box, and time-frequency skeleton fusion results. For events that pass physical consistency verification, the system automatically extracts and outputs standardized parameters, including event arrival time, duration, start and end frequencies, dispersion slope, and local signal-to-noise ratio. Figure 6 This embodiment demonstrates the detection and parameterization results of a typical lightning whistle event. Figure 6 The red rectangle in the middle represents the time-frequency region of the whistle wave automatically located by the automatic identification method for lightning whistle waves. Through skeleton fitting, the typical frequency downtrend characteristics of this event can be obtained: the frequency drops by about 4.70 kHz, and the corresponding duration is about 0.372 seconds. Figure 6 The coordinates of the top left and bottom right corners of the candidate boxes, the start and end frequencies of the skeleton fitting, the arrival time, and the duration have been marked on the spectrum to illustrate the physical consistency constraint process of the automatic recognition algorithm. This event is also an example of a structured data table (CSV) output by the system, containing key physical quantities such as event timestamp, start and end frequencies, duration, dispersion parameter D, and local SNR.

[0028] Figure 7 The application results of the method of the present invention on the full-day VLF data of the observation station on August 6, 2023 are presented. Figure 7The red vertical line represents the lightning whistle events automatically identified by this invention. It can be seen that although the background noise fluctuates significantly throughout the day due to changes in the polar ionosphere, the method of this invention still maintains stable detection capability for lightning whistles under high noise interference. Between 05:00 and 10:00, the method of this invention can accurately identify a large number of weakly dispersive trajectories and provide the correct trigger time. Furthermore, the temporal distribution of the identified events exhibits typical diurnal variation characteristics, consistent with the physical laws of lightning whistle sound waves, further demonstrating the reliability of the identification results. The multi-scale time-frequency fusion and self-supervised gating enhancement design of this invention significantly improves the identification accuracy and robustness in complex noise environments.

[0029] In summary, this invention introduces multi-scale time-frequency fusion, gated self-supervised noise reduction, and a lightweight YOLO detection mechanism throughout the entire process of lightning whistle sound wave recognition, resulting in four significant improvements in recognition capability compared to traditional methods: 1. The detectability of weak whistles is significantly enhanced. Multi-scale STFT fusion can simultaneously preserve short-term details and long-term dispersion trend weak events, reducing the false negative rate by about 30-50%, and making weak trajectories that are difficult to detect by traditional methods stand out significantly in the enhanced spectrum.

[0030] 2. Significantly improved stability under strong noise interference environment. Gated noise masks can adaptively suppress stripe noise, impulse interference and device drift, enabling this method to maintain a high recognition rate in high-noise segments, while traditional thresholding methods are almost completely ineffective in such environments.

[0031] 3. The false positive rate and false negative rate have been significantly reduced. By combining dispersive skeleton consistency verification, this invention significantly reduces the false detection probability of non-whistle events, making the output results more reliable and more suitable as input for subsequent physical inversion.

[0032] 4. Significantly reduced recognition time, suitable for automatic data processing of large-scale continuous observations. It eliminates the need for manual segment-by-segment inspection and can achieve fully automated processing on 24-hour continuous observation data from a single station. The identification time is reduced by two orders of magnitude compared to manual inspection. It can complete single composite spectrum inference on edge devices with a latency of ≤50ms, making it particularly suitable for real-time processing and long-term operation of large-scale continuous observation data.

[0033] This invention employs a lightweight YOLO algorithm and online pseudo-tag updates, enabling the system to run in real-time on edge devices and process long-term continuous data streams without manual intervention—a requirement that traditional template methods cannot meet. The resulting event samples and dispersion features can be directly used for subsequent path and plasma parameter inversion, and support long-term automated monitoring and analysis of magnetosphere-ionosphere coupling processes, demonstrating significant application value in space environment research.

[0034] The lightning whistle sound wave recognition system based on multi-scale time-frequency fusion provided by the present invention will be described below. The lightning whistle sound wave recognition system based on multi-scale time-frequency fusion described below can be referred to in correspondence with the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion described above.

[0035] Figure 8 This is a schematic diagram of the structure of the lightning whistle sound wave recognition system based on multi-scale time-frequency fusion provided in an embodiment of the present invention, as shown below. Figure 8 As shown, it includes: acquisition module 81, preprocessing module 82, stitching module 83, training module 84, storage module 85, and update module 86, wherein: The acquisition module 81 is used to construct a very low frequency electromagnetic wave detection station and acquire very low frequency electromagnetic wave signals in real time; the preprocessing module 82 is used to preprocess the time-series signal of the very low frequency electromagnetic wave using the very low frequency electromagnetic wave observation data to obtain a preprocessed time-series signal; the stitching module 83 is used to perform multi-scale video segmentation and multi-scale spectrum stitching on the preprocessed time-series signal based on the whistle wave characteristics to obtain a composite time-spectrum map; the training module 84 is used to train the recognition model and output multiple recognition results after applying self-supervised noise modeling and gated enhancement to the composite time-spectrum map; the saving module 85 is used to save multiple recognition results and calculate relevant physical parameters; the updating module 86 is used for manual online updating of the pseudo-labels of relevant physical parameters, parameter optimization using preset model compression technology, outputting the lightning whistle wave recognition results and performing visualization analysis.

[0036] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9As shown, the electronic device may include: a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other through the communication bus 940. The processor 910 can call logic instructions in the memory 930 to execute a lightning whistle sound wave recognition method based on multi-scale time-frequency fusion. This method includes: constructing a very low frequency (VLF) electromagnetic wave detection station to acquire VLF electromagnetic wave signals in real time; preprocessing the time-series signal of the VLF electromagnetic wave using the VLF electromagnetic wave observation data to obtain a preprocessed time-series signal; performing multi-scale video segmentation and multi-scale spectrum stitching on the preprocessed time-series signal based on whistle sound wave characteristics to obtain a composite time-spectrum; training a recognition model on the composite time-spectrum using self-supervised noise modeling and gated enhancement to output multiple recognition results; saving multiple recognition results and calculating relevant physical parameters; manually updating the pseudo-labels of relevant physical parameters online, optimizing parameters using a preset model compression technique, outputting the lightning whistle sound wave recognition results, and performing visualization analysis.

[0037] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0038] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion provided by the above methods. The method includes: constructing a very low frequency electromagnetic wave detection station to collect very low frequency electromagnetic wave signals in real time; using the very low frequency electromagnetic wave observation data to preprocess the time-series signal of the very low frequency electromagnetic wave signal to obtain a preprocessed time-series signal; performing multi-scale video segmentation and multi-scale spectrum stitching on the preprocessed time-series signal based on whistle sound wave characteristics to obtain a composite time-spectrum; training a recognition model and outputting multiple recognition results after applying self-supervised noise modeling and gated enhancement to the composite time-spectrum; saving multiple recognition results and calculating relevant physical parameters; manually updating the pseudo-labels of relevant physical parameters online, optimizing parameters using a preset model compression technique, outputting the lightning whistle sound wave recognition results, and performing visualization analysis.

[0039] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program performs the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion provided by the above methods. The method includes: constructing a very low frequency electromagnetic wave detection station to acquire very low frequency electromagnetic wave signals in real time; using the very low frequency electromagnetic wave observation data, preprocessing the time-series signal of the very low frequency electromagnetic wave signal to obtain a preprocessed time-series signal; performing multi-scale video segmentation and multi-scale spectrum stitching on the preprocessed time-series signal based on whistle sound wave characteristics to obtain a composite time-spectrum; training a recognition model on the composite time-spectrum using self-supervised noise modeling and gated enhancement, and outputting multiple recognition results; saving multiple recognition results and calculating relevant physical parameters; manually updating the pseudo-labels of relevant physical parameters online, optimizing parameters using a preset model compression technique, outputting the lightning whistle sound wave recognition results, and performing visualization analysis.

[0040] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0041] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0042] 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying lightning whistle sound waves based on multi-scale time-frequency fusion, characterized in that, include: Construct a very low frequency electromagnetic wave detection station to collect very low frequency electromagnetic wave signals in real time; Using very low frequency electromagnetic wave observation data, the time series signal of the very low frequency electromagnetic wave signal is preprocessed to obtain the preprocessed time series signal. Based on the characteristics of whistle waves, multi-scale video segmentation and multi-scale spectrogram splicing are performed on the preprocessed time-series signal to obtain a composite time-spectrum. After applying self-supervised noise modeling and gated enhancement to the composite time-spectrum graph, the recognition model is trained and multiple recognition results are output. Save multiple recognition results and calculate relevant physical parameters; The pseudo-labels of relevant physical parameters are updated manually online, and the parameters are optimized using a preset model compression technique. The results of lightning whistle sound wave recognition are then output and visualized.

2. The lightning whistle sound wave recognition method based on multi-scale time-frequency fusion according to claim 1, characterized in that, Preprocessing of the timing signal of very low frequency electromagnetic wave signals includes DC removal, power frequency and pulse interference suppression, bandpass filtering, and amplitude normalization.

3. The lightning whistle sound wave recognition method based on multi-scale time-frequency fusion according to claim 1, characterized in that, Based on the characteristics of whistle waves, multi-scale video segmentation and multi-scale spectrogram stitching are performed on the preprocessed time-series signal to obtain a composite time-spectrum image, including: The preprocessed time sequence signal is adaptively segmented according to the signal energy distribution to obtain multiple segmented time sequence signals. A preset short window is used to capture fast dispersion details, and a preset long window is used to extract cross-segment dispersion trends. Short-time Fourier transforms are performed on each segment of the time-series signal using preset short windows and preset long windows, respectively, to output multi-scale video images; Multi-scale video images are stitched together according to chronological and scale order to construct a composite time-spectrum image; Whistle signal is extracted based on whistle frequency domain features to obtain whistle feature parameters.

4. The lightning whistle sound wave recognition method based on multi-scale time-frequency fusion according to claim 1, characterized in that, After applying self-supervised noise modeling and gated enhancement to the composite time-spectrum graph, the recognition model is trained and outputs multiple recognition results, including: Median filtering, fractional estimation, and stability assessment are performed on the background power distribution of each frequency segment to obtain a two-dimensional weight matrix. The two-dimensional weight matrix is ​​used to label power line, polar stripe noise, and broadband pulse interference. The two-dimensional weight matrix is ​​used as a spectral noise mask, which reflects the probability of whether each point is a non-whistle background. The spectral noise mask is provided to the network as a priori cue during self-supervised training. During the feature extraction stage, a lightweight attention branch automatically estimates and generates gate weights based on local time-frequency gradients, signal sparsity, and adjacent scale consistency. These gate weights are used to represent the importance of the current region for whistle recognition. The spectral noise mask and the gating weights are fused point by point at the feature layer to form a denoised enhanced spectrum. The denoised enhanced spectrum is input into the improved lightweight YOLO11 network to output lightning whistle candidate boxes and time-frequency parameters; The enhanced spectrogram after denoising and the whistle feature parameters are used as weak labels to pre-train the recognition model until the recognition accuracy exceeds the preset accuracy threshold, and the initial recognition result is output. If the initial recognition result is determined to be consistent with the preset whistle physical characteristics, the initial recognition result is saved as the final recognition result; otherwise, the initial recognition result is added to the negative event set for retraining.

5. The lightning whistle sound wave recognition method based on multi-scale time-frequency fusion according to claim 1, characterized in that, Save multiple recognition results and calculate relevant physical parameters, including: The multiple recognition results include arrival time, first-order dispersion cutoff frequency, start and end frequencies, 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 sound wave recognition method based on multi-scale time-frequency fusion according to claim 1, characterized in that, The pseudo-labels for relevant physical parameters are updated manually online. Parameter optimization is performed using a pre-defined model compression technique. The results of lightning whistle sound wave recognition are then output and visualized for analysis, including: Candidate events are automatically filtered using a high confidence threshold. The filtered events are then pushed to the annotation interface, where manual selection of preset operation boxes allows the approved or modified annotations to be written back as pseudo-labels. Simultaneously, a list of difficult / negative events is generated for mining difficult samples in subsequent training. The updated pseudo-labels are fine-tuned online using preset indicators; Real-time edge changes are maintained by using quantization and pruning, including INT8 or mixed precision quantization of the recognition network, structural pruning with a ratio greater than a preset ratio and sparsification of the detection head, and satisfying the real-time constraint that the inference delay of a single composite spectrum is less than a preset delay threshold on the edge device. The final confirmation event record corresponding to the lightning whistle sound wave recognition result is written to the database or object storage in a preset structured format, containing multiple preset fields.

7. A lightning whistle sound wave recognition system based on multi-scale time-frequency fusion, characterized in that, include: The acquisition module is used to build a very low frequency electromagnetic wave detection station and acquire very low frequency electromagnetic wave signals in real time. The preprocessing module is used to preprocess the time series signal of the very low frequency electromagnetic wave signal using very low frequency electromagnetic wave observation data to obtain the preprocessed time series signal. The splicing module is used to perform multi-scale video segmentation and multi-scale spectrogram splicing on the preprocessed time-series signal based on the characteristics of the whistle wave, so as to obtain a composite time-spectrum. The training module is used to train the recognition model and output multiple recognition results after applying self-supervised noise modeling and gated enhancement to the composite time-spectrum graph; The storage module is used to save multiple recognition results and calculate relevant physical parameters; The update module is used for manual online updating of pseudo-labels for relevant physical parameters. It uses preset model compression technology to optimize parameters, outputs lightning whistle sound wave recognition results, and performs visualization analysis.

8. 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 program, it implements the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the lightning whistle sound wave recognition method based on multi-scale time-frequency fusion as described in any one of claims 1 to 6.

Images