Lightning waveform identification model training method and lightning waveform identification method

Abstract

Embodiments of the present disclosure provide a lightning waveform identification model training method and a lightning waveform identification method, which are applied to the technical field of lightning detection and identification. The method comprises: synchronously collecting impulse data and double-channel waveform data of multiple sites through a wideband magnetic field sensor; performing time difference positioning and event screening based on the impulse data, matching the corresponding waveform data according to the screening result, binding the waveform data and the true value label corresponding to the event, and forming the labeled training sample as a labeled data set; collecting unlabeled double-channel waveform data as an unlabeled data set, and jointly constituting a mixed data set with the labeled data set and the unlabeled data set; and training a teacher-student model based on the mixed data set to obtain a lightning waveform identification model. In this way, the dependence on artificial labeling in the deep learning model training process can be effectively reduced, and the overall robustness and identification accuracy of the system can be significantly improved in the case of insufficient site quantity.

Description

Technical Field

[0001] This disclosure relates to the field of lightning detection and identification technology, and in particular to a training method for a lightning waveform identification model and a lightning waveform identification method. Background Technology

[0002] In the field of lightning detection and identification technology, accurately obtaining the spatial location and type of lightning (such as cloud-to-ground lightning, etc.) is of great significance for meteorological departments to carry out forest fire early warning, power system protection, flight scheduling, and urban emergency management. Currently, mainstream lightning location technologies mostly adopt the Time-of-Age (TOA) method and the Time-of-Age Direction-Finding (TOA+DOA) method, which achieves spatial location of lightning events through the collaborative work of multiple observation stations. Accurate calculation of two-dimensional geographic location usually requires effective data from at least four observation stations, while obtaining three-dimensional location information including altitude requires the joint participation of data from five or more stations. Furthermore, the determination of lightning type often depends on altitude information.

[0003] However, existing lightning location systems face numerous challenges in practical applications, severely limiting the accuracy, completeness, and practicality of their positioning results. On one hand, the deployment of observation stations is limited, with large spatial distances between stations and insufficient overall deployment density, making it difficult to meet the requirements for a large number of stations needed for high-precision positioning. On the other hand, complex field environments present strong electromagnetic interference, and complex terrain causes significant signal attenuation, further affecting data validity. Simultaneously, network limitations prevent stations from uploading complete raw waveform data in real time, creating a data transmission bottleneck. These factors combined result in the number of observation stations capable of simultaneously capturing and effectively processing the same lightning event often falling below the minimum threshold required for 3D positioning in actual operation scenarios. This directly leads to a lack of lightning height information, making height-based lightning type identification unreliable and severely weakening the overall practicality and intelligence of the positioning system, making it difficult to adapt to complex and sparse observation scenarios.

[0004] Furthermore, in deep learning-based lightning identification technology applications, traditional methods heavily rely on large-scale, high-quality labeled samples. Manual labeling is not only costly and inefficient, but also suffers from strong subjectivity and poor consistency, hindering model training efficiency and generalization ability. Simultaneously, existing deep learning models are often structurally complex, making efficient deployment on resource-constrained embedded edge devices difficult. This leads to data processing reliance on central servers, increasing data transmission pressure and processing latency, failing to meet real-time identification requirements. Therefore, there is an urgent need for a technical solution that can reduce reliance on manual labeling, adapt to edge computing environments, and still achieve efficient lightning type identification and location capability compensation even with insufficient site coverage, thereby improving the practicality and intelligence of lightning location systems in complex environments. Summary of the Invention

[0005] This disclosure provides a training method for a lightning waveform identification model and a lightning waveform identification method, which solves the technical problems of existing lightning positioning systems, such as the large number of positioning sites, lack of height information, high cost of manual annotation for model training, and insufficient operating efficiency and recognition accuracy due to model incompatibility with edge computing devices.

[0006] According to a first aspect of this disclosure, a training method for a lightning waveform discrimination model is provided. The method includes: simultaneously acquiring pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at various observation stations;

[0007] Based on the pulse data, time difference method is used for localization and event filtering, and the corresponding waveform data is matched according to the filtering results. The waveform data is then bound to the truth label corresponding to the event to form an annotated training sample.

[0008] The labeled training samples are used as the labeled dataset, and additional unlabeled dual-channel waveform data are collected as the unlabeled dataset. The labeled dataset and the unlabeled dataset are combined to form a hybrid dataset for model training.

[0009] The teacher-student model was trained based on the hybrid dataset to obtain a lightning waveform identification model.

[0010] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the step of performing time-difference localization and event filtering based on the pulse data, matching the corresponding waveform data according to the filtering results, and binding the waveform data with the truth label corresponding to the event to form labeled training samples includes:

[0011] Based on the pulse data from each observation station, a time difference positioning algorithm is used for joint positioning to obtain the spatial location and type of the lightning event;

[0012] Lightning events with a positioning error less than a preset threshold and a number of observation stations participating in the positioning not less than a preset number are selected, and the event type of the lightning event is used as the truth label of the event.

[0013] Based on the timestamp information of the selected lightning events, the original pulse data and dual-channel waveform data are back-matched.

[0014] The matched dual-channel waveform data is processed by channel selection and downsampling to generate single-channel waveform data;

[0015] The single-channel waveform data is normalized to obtain standardized single-channel waveform data;

[0016] The standardized single-channel waveform data is bound to the truth label corresponding to the event to form an annotated training sample.

[0017] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the process of channel selection and downsampling of the matched dual-channel waveform data to generate single-channel waveform data includes:

[0018] Calculate the peak-to-peak value of each channel in the dual-channel waveform data, and take the channel with the larger peak-to-peak value as the main channel input;

[0019] The waveform data of the selected main channel is downsampled using max pooling, and the data dimension is compressed to the preset target sampling rate to generate single-channel waveform data.

[0020] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the training of the teacher-student model based on the hybrid dataset to obtain the lightning waveform discrimination model includes:

[0021] The hybrid dataset is divided into a training set, a validation set, and a test set;

[0022] Construct an average teacher framework that includes student and teacher models;

[0023] The teacher and student models are trained using a training set. During the training process, data augmentation is applied to the input waveform data, and a dynamic learning rate adjustment mechanism is used to optimize the training process.

[0024] After each training round, the performance of the current student model and / or teacher model is evaluated using the validation set, and the parameters of the model with the best performance on the validation set are saved.

[0025] Once the training reaches the preset number of rounds or the early stopping condition is met, the best-performing student model is loaded and the accuracy of the lightning type identification is verified through the test set. When the accuracy reaches the preset index, the best-performing student model is used as the lightning waveform identification model.

[0026] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the teacher-student model comprises, in sequence, an input layer, a one-dimensional convolutional layer, a gated recurrent unit layer, and an output layer; wherein,

[0027] The one-dimensional convolutional layer contains several convolutional structures, followed by batch normalization layers, ReLU activation functions, and Dropout layers. It also introduces a one-dimensional frequency dynamic convolution mechanism to extract local waveform features and dynamically adjust the weight distribution according to different frequency components in the temporal dimension.

[0028] The gated recurrent unit layer is used to receive the feature sequence output by the one-dimensional convolutional layer, model the time dependency in the waveform data, and output global temporal features;

[0029] The output layer, constructed based on several fully connected layers, is used to reduce the dimensionality of the global temporal features and transform the dimensionality-reduced global temporal features into prediction probability vectors corresponding to different lightning types through the Softmax activation function.

[0030] According to a second aspect of this disclosure, a lightning waveform identification method is provided. The method includes: simultaneously acquiring pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at various observation stations;

[0031] The raw waveform data collected in real time by each observation station is input into the lightning waveform identification model as described above to generate a prediction probability vector corresponding to different lightning types. The lightning type with the highest probability value in the prediction probability vector is selected as the local identification result of the observation station for this lightning event.

[0032] The local identification result is bound and encapsulated with the timestamp corresponding to the lightning event and the metadata of this observation station to form an upload data packet;

[0033] The uploaded data packets are sent to the central processing system, which receives and integrates the uploaded data packets from all observation stations for lightning events within the same time window, and generates a spatial distribution map of lightning activity that reflects the types and spatial distribution of lightning activity in the region.

[0034] According to a third aspect of this disclosure, a training apparatus for a lightning waveform discrimination model is provided. The apparatus includes:

[0035] The data acquisition module is used to simultaneously acquire pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at each observation station;

[0036] The data labeling module is used to perform time difference method localization and event filtering based on the pulse data, and to match the corresponding waveform data according to the filtering results, and bind the waveform data with the truth label corresponding to the event to form labeled training samples.

[0037] The dataset construction module is used to take the labeled training samples as a labeled dataset and collect additional unlabeled dual-channel waveform data as an unlabeled dataset, and to combine the labeled dataset and the unlabeled dataset to form a hybrid dataset for model training.

[0038] The model training module is used to train the teacher-student model based on the hybrid dataset to obtain the lightning waveform identification model.

[0039] According to a fourth aspect of this disclosure, a lightning waveform identification device is provided. The device includes:

[0040] The data acquisition module is used to simultaneously acquire pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at each observation station;

[0041] The data identification module is used to input the raw waveform data collected in real time by each observation station into the lightning waveform identification model as described above, generate prediction probability vectors corresponding to different lightning types, and select the lightning type with the highest probability value in the prediction probability vector as the local identification result of the observation station for this lightning event.

[0042] The data packet generation module is used to bind and encapsulate the local identification result with the timestamp corresponding to the lightning event and the metadata of this observation station to form an upload data packet;

[0043] The data packet upload module is used to send the uploaded data packets to the central processing system, so that the central processing system can receive and integrate the uploaded data packets from all observation stations for lightning events within the same time window, and generate a lightning activity spatial distribution map reflecting the lightning activity type and spatial distribution in the region.

[0044] According to a fifth aspect of this disclosure, an electronic device is provided. The electronic device includes a memory and a processor, wherein a computer program is stored on the memory, and the processor executes the program to implement the method described above.

[0045] According to a sixth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method described above.

[0046] In this disclosure, firstly, pulse data and dual-channel waveform data are simultaneously acquired based on a broadband magnetic field sensor, which can completely capture the electromagnetic characteristics and directional information of lightning events (such as east-west and north-south magnetic field components), providing a rich and complete raw data foundation for subsequent feature extraction. This avoids the feature loss problem caused by incomplete single-channel data information and ensures that the model can learn more comprehensive lightning waveform features.

[0047] Secondly, by using time-difference positioning and event filtering, only lightning events with positioning errors meeting the threshold and a sufficient number of participating sites are retained. Their corresponding waveform data is then bound to ground truth labels, automatically generating high-quality labeled samples. This method of inferring labels based on high-precision positioning results eliminates the need for manual annotation, reducing data preparation time and manpower costs while avoiding the subjectivity and inconsistency issues of manual annotation. This effectively improves the reliability and generation efficiency of training samples.

[0048] Furthermore, combining labeled and unlabeled datasets to form a hybrid dataset allows for the full utilization of a large amount of unlabeled waveform data to expand the training data scale. The diverse operating conditions contained in the unlabeled data (such as lightning waveforms under different weather and terrain conditions) enrich the model's learning sample distribution, enabling the model to be exposed to a wider range of scene features during training, thereby improving its adaptability to complex environments and generalization performance, and solving the problem of insufficient model generalization ability caused by traditional reliance on limited labeled data.

[0049] Finally, the teacher-student model training method based on mixed datasets can fully leverage the advantages of semi-supervised learning. The teacher model generates stable soft labels by averaging the parameters of the student model, guiding the student model to learn on unlabeled data. This approach ensures classification accuracy using a small amount of high-quality labeled data while simultaneously uncovering latent feature patterns from a large amount of unlabeled data, enabling the model to achieve high recognition accuracy even with limited labeled data. Furthermore, this training method enhances the model's robustness to waveform variations (such as waveform distortion caused by noise interference and signal attenuation), laying the foundation for high-precision real-time identification on edge devices.

[0050] In summary, this disclosure effectively reduces reliance on manual annotation by optimizing data collection, sample generation, dataset construction, and training mechanisms, thereby improving the model's adaptability and recognition accuracy in complex scenarios. It provides reliable model support for efficient identification of lightning types, and alleviates the difficulty in type identification caused by insufficient number of sites and limited data quality in existing technologies.

[0051] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0052] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0053] Figure 1 A schematic diagram of an exemplary operating environment of an embodiment of this disclosure is shown;

[0054] Figure 2 A flowchart illustrating waveform processing identification for edge computing devices provided in embodiments of this disclosure is shown;

[0055] Figure 3 A flowchart illustrating the training method for the lightning waveform discrimination model provided in the embodiments of this disclosure is shown;

[0056] Figure 4 A flowchart of a lightning waveform identification method provided by an embodiment of this disclosure is shown;

[0057] Figure 5 A structural diagram of a training apparatus for a lightning waveform discrimination model provided in an embodiment of this disclosure is shown;

[0058] Figure 6 A structural diagram of a lightning waveform identification device provided in an embodiment of this disclosure is shown;

[0059] Figure 7 A structural diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0061] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0062] In this disclosure, firstly, pulse data and dual-channel waveform data are simultaneously acquired based on a broadband magnetic field sensor, which can completely capture the electromagnetic characteristics and directional information of lightning events (such as east-west and north-south magnetic field components), providing a rich and complete raw data foundation for subsequent feature extraction. This avoids the feature loss problem caused by incomplete single-channel data information and ensures that the model can learn more comprehensive lightning waveform features.

[0063] Secondly, by using time-difference positioning and event filtering, only lightning events with positioning errors meeting the threshold and a sufficient number of participating sites are retained. Their corresponding waveform data is then bound to ground truth labels, automatically generating high-quality labeled samples. This method of inferring labels based on high-precision positioning results eliminates the need for manual annotation, reducing data preparation time and manpower costs while avoiding the subjectivity and inconsistency issues of manual annotation. This effectively improves the reliability and generation efficiency of training samples.

[0064] Furthermore, combining labeled and unlabeled datasets to form a hybrid dataset allows for the full utilization of a large amount of unlabeled waveform data to expand the training data scale. The diverse operating conditions contained in the unlabeled data (such as lightning waveforms under different weather and terrain conditions) enrich the model's learning sample distribution, enabling the model to be exposed to a wider range of scene features during training, thereby improving its adaptability to complex environments and generalization performance, and solving the problem of insufficient model generalization ability caused by traditional reliance on limited labeled data.

[0065] Finally, the teacher-student model training method based on mixed datasets can fully leverage the advantages of semi-supervised learning. The teacher model generates stable soft labels by averaging the parameters of the student model, guiding the student model to learn on unlabeled data. This approach ensures classification accuracy using a small amount of high-quality labeled data while simultaneously uncovering latent feature patterns from a large amount of unlabeled data, enabling the model to achieve high recognition accuracy even with limited labeled data. Furthermore, this training method enhances the model's robustness to waveform variations (such as waveform distortion caused by noise interference and signal attenuation), laying the foundation for high-precision real-time identification on edge devices.

[0066] In summary, this disclosure effectively reduces reliance on manual annotation by optimizing data collection, sample generation, dataset construction, and training mechanisms, thereby improving the model's adaptability and recognition accuracy in complex scenarios. It provides reliable model support for efficient identification of lightning types, and alleviates the difficulty in type identification caused by insufficient number of sites and limited data quality in existing technologies.

[0067] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of the present disclosure can be implemented is shown. The operating environment includes multiple observation stations, each equipped with a waveform acquisition device for continuously acquiring raw waveform data of the earth's electric or electromagnetic fields. The acquired waveform data is input in real time to a local edge computing device, where a lightning waveform discrimination model deployed therein rapidly discriminates the data and generates preliminary identification results.

[0068] After identification, each edge computing device sends the identification results (such as whether a lightning event occurred, category label, confidence level, event timestamp, etc.) back to the central processing system in structured data format. The central processing system is responsible for further data fusion, spatial positioning, and multi-station linkage analysis to generate a more complete lightning activity map with higher spatial resolution.

[0069] Figure 2 It shows Figure 1 The flowchart shown is for waveform processing identification of edge computing devices.

[0070] The edge computing device first performs local preprocessing on the raw waveform data collected by the observation station. This step includes, but is not limited to:

[0071] Waveform channel selection: Select valid magnetic field channel data and remove redundant or interfering channels;

[0072] Signal denoising: Algorithms such as bandpass filtering, wavelet denoising, or adaptive filters are used to suppress power frequency interference and background noise;

[0073] Downsampling: To reduce model inference latency and resource consumption, high-sampling-rate waveforms are downsampled appropriately while retaining key features;

[0074] This step ensures that the data input to the model has clear lightning characteristics in both the time and frequency domains, thus providing a high-quality data foundation for subsequent recognition and inference.

[0075] The preprocessed waveform segments are input into a lightning waveform identification model deployed in a local edge computing device for inference.

[0076] The model inference phase outputs multi-class probability distributions or confidence values, reflecting the probability that the waveform segment belongs to different types of lightning events (such as cloud-to-ground lightning, cloud-to-cloud lightning, positive polarity, negative polarity, etc.).

[0077] Based on the model's inference results, edge computing devices use a set threshold strategy or the maximum probability method to determine the final identification category, and combine this with metadata such as timestamps, confidence levels, and site numbers to generate a structured identification result package. This result is then cached or directly uploaded to the central processing system for subsequent processing such as spatial positioning and multi-site joint analysis.

[0078] Figure 3 A flowchart illustrating the training method for the lightning waveform discrimination model provided in embodiments of this disclosure is shown, such as... Figure 3 As shown, the training method 300 for the lightning waveform discrimination model may include the following steps:

[0079] The S310, based on broadband magnetic field sensors at various observation stations, simultaneously acquires pulse data and dual-channel waveform data generated by lightning events.

[0080] Specifically, the broadband magnetic field sensors deployed at each observation station must cover the typical frequency range of lightning pulse signals to ensure complete capture of the instantaneous electromagnetic field changes generated by different types of lightning (such as cloud lightning and ground lightning). The sensors employ a dual-channel design, separately acquiring the east-west (EW) and north-south (NS) magnetic field components. Differential input and shielded enclosure designs enhance electromagnetic interference resistance, reducing the impact of environmental interference such as industrial electromagnetic noise and terrain reflections on the signal. To achieve synchronous acquisition across multiple stations, each sensor must be connected to a high-precision time synchronization module (such as GPS or BeiDou timing unit) to ensure that the time synchronization error between different stations is controlled at the microsecond level, guaranteeing precise alignment of waveform data for the same lightning event on the time axis at different stations.

[0081] Specifically, the acquisition triggering mechanism adopts an adaptive threshold strategy. When the detected signal amplitude exceeds a preset multiple of the background noise (e.g., 5-10 times), continuous acquisition is automatically initiated. The acquisition duration must cover the typical lightning pulse duration, and key pulse parameters such as pulse start time, peak time, and peak-to-peak value are recorded simultaneously. During the acquisition process, in addition to dual-channel waveform data, metadata must be associated in real time, including the unique site number, sensor latitude, longitude and altitude, sampling rate, equipment operating status (e.g., whether it is operating normally, whether there are fault alarms), acquisition timestamps, and other information. Furthermore, the metadata and waveform data are bound together by a unique event ID to facilitate subsequent data retrieval and traceability.

[0082] Specifically, the sensor has a built-in temporary cache module that can temporarily store the collected data to avoid data loss due to sudden transmission interruptions. Data can be transmitted when the edge computing device is idle. It also supports marking abnormal data (such as distorted signals when the sensor fails) to provide a basis for subsequent data cleaning.

[0083] S320 performs time difference localization and event filtering based on pulse data, and matches the corresponding waveform data according to the filtering results. The waveform data is then bound to the truth label corresponding to the event to form an labeled training sample.

[0084] In some embodiments, time-difference localization and event filtering are performed based on pulse data, and corresponding waveform data is matched according to the filtering results. The waveform data is then bound to the ground truth label corresponding to the event to form labeled training samples, including:

[0085] Based on the pulse data from each observation station, a time difference positioning algorithm is used for joint positioning to obtain the spatial location and type of lightning events;

[0086] Lightning events with a positioning error less than a preset threshold and a number of observation stations participating in the positioning not less than a preset number are selected, and the event type of the lightning event is used as the truth label of the event.

[0087] Based on the timestamp information of the selected lightning events, the original pulse data and dual-channel waveform data are backtracked and matched.

[0088] The matched dual-channel waveform data is processed by channel selection and downsampling to generate single-channel waveform data;

[0089] The single-channel waveform data is normalized to obtain standardized single-channel waveform data;

[0090] Standardized single-channel waveform data is bound to the truth label corresponding to the event to form labeled training samples.

[0091] In some embodiments, channel selection and downsampling processing are performed on the matched dual-channel waveform data to generate single-channel waveform data, including:

[0092] Calculate the peak-to-peak value of each channel in the dual-channel waveform data, and use the channel with the larger peak-to-peak value as the main channel input;

[0093] The waveform data of the selected main channel is downsampled using max pooling, and the data dimension is compressed to the preset target sampling rate to generate single-channel waveform data.

[0094] Specifically, based on the pulse data collected by each observation station, the central station uses the Time Difference of Arrival (TOA) algorithm to perform joint calculations across multiple stations. By calculating the time difference of arrival of the same lightning event at different stations and combining this with the known latitude and longitude coordinates of the stations, the three-dimensional spatial location (latitude, longitude, and altitude) and preliminary type (such as cloud-to-ground lightning IC, ground-to-ground lightning CG) of the lightning event are calculated. To ensure the reliability of the positioning results, strict event screening conditions can be set (e.g., the number of observation stations participating in the positioning should not be less than 5 to meet the basic data requirements for three-dimensional positioning; the positioning results should contain valid altitude information to provide a physical basis for type judgment; at the same time, the positioning chi-square value should be less than 1000 to quantitatively assess that the positioning error is within an acceptable range and ensure the physical reliability of the results).

[0095] Specifically, after screening, the lightning event types that meet the criteria are used as the truth labels for the events. Then, based on the precise timestamps corresponding to the screened lightning events, the dual-channel (EW, NS) waveform data corresponding to the original pulse data of each participating site are back-matched. Time axis alignment ensures that the waveform segments strictly correspond to the target events, avoiding cross-event data overlap. For the matched dual-channel waveform data, the peak-to-peak value (i.e., the difference between the maximum and minimum values ​​in the waveform) of each channel within the effective pulse period is first calculated. The channel with the larger peak-to-peak value is selected as the main channel to retain waveform information with stronger signal energy and more significant characteristics. Then, the selected main channel waveform is downsampled using max pooling. The pooling window size is set according to the target sampling length, compressing the data dimension while retaining waveform abrupt changes (such as pulse leading edges and peak points), generating single-channel waveform data.

[0096] Specifically, amplitude normalization is performed on the single-channel waveform data, scaling the waveform amplitude to a fixed range (e.g., [-1, 1]). This eliminates amplitude scale differences between different events, improving the numerical stability of the model during training. Finally, the standardized single-channel waveform data is bound to the ground truth labels corresponding to the event (e.g., "ground flash" or "cloud flash"), forming uniformly structured and accurately labeled training samples, providing a high-quality data foundation for subsequent deep learning model training.

[0097] S330 uses the labeled training samples as the labeled dataset and collects additional unlabeled dual-channel waveform data as the unlabeled dataset, combining the labeled dataset and the unlabeled dataset to form a hybrid dataset for model training.

[0098] Specifically, the labeled training samples generated through the aforementioned steps are integrated into a labeled dataset. This dataset contains rigorously screened standardized single-channel waveform data and corresponding ground truth labels for lightning types (such as cloud-to-ground lightning). Each sample is associated with metadata such as the original acquisition site information and timestamps to ensure data traceability and consistency. Simultaneously, additional unlabeled dual-channel waveform data is collected as an unlabeled dataset. This unlabeled data originates from a large number of lightning event waveforms collected daily by various observation stations, covering signals under different weather conditions, terrain environments, and electromagnetic interference scenarios. This includes waveforms that could not generate reliable ground truth labels due to insufficient participating stations (less than 5), waveforms with slight noise interference but still identifiable as lightning events, and some marginal signals (such as pulses that are suspected to be lightning but difficult to definitively determine), to enrich data diversity.

[0099] Specifically, when constructing a hybrid dataset, the unlabeled data needs to undergo the same preprocessing operations as the labeled dataset. That is, it is first converted into a single-channel waveform by using a channel selection strategy (selecting the channel with the larger peak-to-peak value), then downsampled to a fixed length (such as 1000 sampling points) by max pooling, and then normalized by amplitude to ensure that the unlabeled data and the labeled data are consistent in format and scale.

[0100] In addition, the samples in the mixed dataset need to be deduplicated to remove the waveforms of the same lightning event that were collected repeatedly. At the same time, the invalid data with severe distortion (such as noise caused by sensor failure) or without effective pulse characteristics are filtered out through the data quality assessment mechanism. Finally, a mixed dataset is formed, which consists of the labeled dataset and the preprocessed unlabeled dataset. This provides a training data foundation with rich scene information and uniform format for subsequent semi-supervised learning training, so as to improve the model's adaptability and generalization performance to complex real-world environments.

[0101] S340, a lightning waveform identification model is obtained by training a teacher-student model based on a mixed dataset.

[0102] In some embodiments, the lightning waveform identification model is obtained by training the teacher-student model based on a hybrid dataset, including:

[0103] The hybrid dataset is divided into a training set, a validation set, and a test set;

[0104] Construct an average teacher framework that includes student and teacher models;

[0105] The teacher and student models are trained using a training set. During the training process, data augmentation is applied to the input waveform data, and a dynamic learning rate adjustment mechanism is used to optimize the training process.

[0106] After each training round, the performance of the current student model and / or teacher model is evaluated using the validation set, and the parameters of the model with the best performance on the validation set are saved.

[0107] Once the training reaches the preset number of rounds or the early stopping condition is met, the best-performing student model is loaded and the accuracy of the lightning type identification is verified through the test set. When the accuracy reaches the preset index, the best-performing student model is used as the lightning waveform identification model.

[0108] Specifically, the mixed dataset is divided into training, validation, and test sets according to a preset ratio. During the division process, the sample ratio of different lightning types (such as cloud lightning and ground lightning) within each set must be kept consistent to avoid data distribution bias affecting the model's generalization. Subsequently, an average teacher framework is constructed, in which the student model and the teacher model use the same lightweight network structure (integrating one-dimensional frequency dynamic convolution and GRU) with identical initial parameters. The student model updates its parameters through the backpropagation algorithm, while the teacher model updates its parameters by performing an exponential moving average on the student model's parameters.

[0109] During training, diverse data augmentation processes are applied to the waveform data in the training set: Gaussian white noise and low-frequency interference signals are added to both labeled and unlabeled samples, and the unlabeled samples undergo additional time-axis perturbation to simulate signal variations in real-world scenarios. The learning rate is dynamically adjusted based on the number of training epochs or the performance on the validation set, which can accelerate model convergence in the early stages, stabilize model performance in the later stages, and avoid overfitting.

[0110] Specifically, for labeled samples in the training set, their standardized single-channel waveform data is input into the student model, and the difference between the lightning type probability distribution predicted by the student model and the true label of the sample is calculated as the supervision loss. For unlabeled samples in the training set, data perturbation is applied to their standardized single-channel waveform data to generate two slightly different versions, which are then input into the student model and the teacher model respectively. The difference between the predicted probability distributions output by the two models is calculated as the consistency loss. The supervision loss and the consistency loss are weighted and summed according to preset weight coefficients to obtain the total loss of the model, and the parameters of the student model are updated using the backpropagation algorithm with this total loss.

[0111] Specifically, after each training round, the performance of the student model is evaluated using a validation set. Metrics include type recognition accuracy and F1 score (with a focus on recall for ground flash recognition to meet actual early warning needs). The model parameters with the highest F1 score on the validation set are saved in real time. An early stopping mechanism is implemented during training; if the performance on the validation set shows no improvement for 10 consecutive rounds, training is terminated to avoid overfitting.

[0112] Once the training reaches the preset number of rounds or the early stopping condition is triggered, the saved optimal parameters are loaded, and the performance of the student model is verified through the test set. Not only is the overall recognition accuracy calculated, but its recognition performance in high noise, low signal-to-noise ratio and edge signal scenarios is also evaluated to ensure adaptability to complex environments. When the accuracy of the test set reaches the preset indicators (such as accuracy ≥ 92% and F1 score ≥ 0.91), the student model with the best performance is used as the lightning waveform identification model.

[0113] In some embodiments, the teacher-student model includes, in sequence, an input layer, a one-dimensional convolutional layer, a gated recurrent unit layer, and an output layer; wherein...

[0114] A one-dimensional convolutional layer contains several convolutional structures, followed by batch normalization layers, ReLU activation functions, and Dropout layers. It also introduces a one-dimensional frequency dynamic convolution mechanism to extract local waveform features and dynamically adjust the weight distribution according to different frequency components in the temporal dimension.

[0115] The gated recurrent unit layer is used to receive the feature sequence output by the one-dimensional convolutional layer, model the temporal dependencies in the waveform data, and output global temporal features;

[0116] The output layer, built upon several fully connected layers, is used to reduce the dimensionality of global temporal features and transforms the dimensionality-reduced global temporal features into prediction probability vectors corresponding to different lightning types through the Softmax activation function.

[0117] Specifically, the teacher-student model integrates one-dimensional frequency dynamic convolution with gated recurrent unit (GRU) structure, possessing excellent ability to extract abrupt waveform features and time series modeling capabilities, and is adapted to resource-constrained edge computing platforms.

[0118] This model takes one-dimensional waveform data as input, which is normalized and standardized to ensure the uniformity of numerical ranges for each feature dimension during training. The first part of the model consists of several one-dimensional convolutional layers used to extract local features from the original waveform. In particular, this disclosure introduces a one-dimensional frequency dynamic convolution mechanism, enabling the convolution kernel to dynamically adjust the weight distribution according to different frequency components in the temporal dimension, thereby improving the network's ability to perceive pulse-like waveforms. By introducing a dynamic receptive field, the model can enhance its adaptability to different waveform variation patterns without significantly increasing the number of parameters.

[0119] Specifically, after convolutional feature extraction, the feature sequence is fed into a gated recurrent unit (GRU) network. The GRU layer is used to model the temporal dependencies in the waveform data. Its structure is more compact than LSTM, effectively reducing computational and storage overhead while maintaining excellent temporal modeling capabilities. The output of this GRU layer serves as global temporal features, which are further connected to a fully connected layer for dimensionality reduction and classification.

[0120] To enhance the model's generalization ability and suppress overfitting, batch normalization and activation functions (such as ReLU) are applied after the convolutional layers, and dropout layers are inserted at appropriate positions. Furthermore, a softmax activation function is used at the end of the model to output classification probabilities, used to distinguish between different types of lightning events (cloud lightning, ground lightning).

[0121] Specifically, the model output is a vector of length 2, where each vector corresponds to a predicted probability of the lightning type. Furthermore, to improve confidence assessment capabilities, the model output may include a prediction confidence threshold for post-processing. The specific structure of the model is shown in Table 1.

[0122]

[0123]

[0124] Figure 4 A flowchart of a lightning waveform identification method provided by an embodiment of this disclosure is shown, such as... Figure 4 As shown, the lightning waveform identification method 400 may include the following steps:

[0125] The S410, based on broadband magnetic field sensors at various observation stations, simultaneously acquires pulse data and dual-channel waveform data generated by lightning events.

[0126] S420: Input the raw waveform data collected in real time by each observation station into the lightning waveform identification model as described above to generate prediction probability vectors corresponding to different lightning types. Select the lightning type with the highest probability value in the prediction probability vector as the local identification result of the observation station for this lightning event.

[0127] S430 binds and encapsulates the local identification result with the timestamp corresponding to the lightning event and the metadata of this observation station to form an upload data packet.

[0128] S440 sends the uploaded data packets to the central processing system, so that the central processing system can receive and integrate the uploaded data packets from all observation stations for lightning events within the same time window, and generate a spatial distribution map of lightning activity that reflects the types and spatial distribution of lightning activity in the region.

[0129] In summary, this disclosure achieves the following technical effects:

[0130] (1) The training samples are automatically selected and generated based on high-quality localization results, which reduces the dependence on manually labeled data and reduces the time and manpower costs in the training process of deep learning models.

[0131] (2) Based on the lightning waveform identification model for edge computing devices, the computational burden of edge computing devices is reduced and the real-time performance and response speed of model inference are improved by optimizing the model structure and processing flow.

[0132] (3) By optimizing the lightweight deployment and inference mechanism of the lightning waveform identification model at the site end, the model's operating efficiency and recognition accuracy on resource-constrained devices were improved, enabling real-time processing and intelligent recognition of lightning waveform data at the acquisition end, thus alleviating the pressure on the central server in terms of data transmission and computation. It also improved the compensation capabilities for auxiliary identification and spatial positioning of lightning types when the number of participating sites was insufficient.

[0133] The above is an introduction to the method embodiments. The following describes the present disclosure further through device embodiments.

[0134] Figure 5 A structural diagram of a training apparatus for a lightning waveform discrimination model provided in an embodiment of this disclosure is shown, as follows: Figure 5 As shown, the training device 500 for the lightning waveform discrimination model may include:

[0135] The data acquisition module 510 is used to simultaneously acquire pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at each observation station.

[0136] The data annotation module 520 is used for time difference localization and event filtering based on pulse data, and matches the corresponding waveform data according to the filtering results, and binds the waveform data with the truth label corresponding to the event to form an annotated training sample.

[0137] The dataset construction module 530 is used to take the labeled training samples as the labeled dataset and collect additional unlabeled dual-channel waveform data as the unlabeled dataset, and combine the labeled dataset and the unlabeled dataset to form a hybrid dataset for model training.

[0138] Model training module 540 is used to train the teacher-student model based on a mixed dataset to obtain a lightning waveform identification model.

[0139] Figure 6 A structural diagram of a lightning waveform identification device provided in an embodiment of this disclosure is shown, as follows: Figure 6 As shown, the lightning waveform identification device 600 may include:

[0140] The data acquisition module 610 is used to simultaneously acquire pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at each observation station.

[0141] The data identification module 620 is used to input the raw waveform data collected in real time by each observation station into the lightning waveform identification model, generate prediction probability vectors corresponding to different lightning types, and select the lightning type with the highest probability value in the prediction probability vector as the local identification result of the observation station for this lightning event.

[0142] The data packet generation module 630 is used to bind and encapsulate the local identification result with the timestamp corresponding to the lightning event and the metadata of this observation station to form an upload data packet.

[0143] The data packet upload module 640 is used to send the uploaded data packets to the central processing system, so that the central processing system can receive and integrate the uploaded data packets from all observation stations for lightning events within the same time window, and generate a lightning activity spatial distribution map reflecting the lightning activity type and spatial distribution in the region.

[0144] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps may be performed in other orders or simultaneously.

[0145] Secondly, those skilled in the art should also know that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by this disclosure.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0147] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0148] Figure 7 A structural diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Electronic device 700 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 700 may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0149] like Figure 7As shown, the electronic device 700 may include a computing unit 701, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 702 or a computer program loaded from a storage unit 708 into a random access memory (RAM) 703. The RAM 703 may also store various programs and data required for the operation of the electronic device 700. The computing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0150] Multiple components in electronic device 700 are connected to I / O interface 705, including: input unit 706, such as keyboard, mouse, etc.; output unit 707, such as various types of displays, speakers, etc.; storage unit 708, such as disk, optical disk, etc.; and communication unit 709, such as network card, modem, wireless transceiver, etc. Communication unit 709 allows electronic device 700 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0151] The computing unit 701 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 701 performs the various methods and processes described above, such as method 300 or method 400. For example, in some embodiments, method 300 or method 400 may be implemented as a computer program product, including a computer program tangibly contained in a computer-readable medium, such as storage unit 708. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 700 via ROM 702 and / or communication unit 709. When the computer program is loaded into RAM 703 and executed by the computing unit 701, one or more steps of method 300 or method 400 described above may be performed. Alternatively, in other embodiments, the computing unit 701 may be configured to perform method 300 or method 400 by any other suitable means (e.g., by means of firmware).

[0152] The various embodiments described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), payload programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0153] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0154] In the context of this disclosure, a computer-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0155] It should be noted that this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute method 300 or method 400 and achieve the corresponding technical effects achieved by the embodiments of this disclosure in executing their methods. For the sake of brevity, these will not be elaborated here.

[0156] In addition, this disclosure also provides a computer program product including a computer program that, when executed by a processor, implements method 300 or method 400.

[0157] To provide interaction with a user, the embodiments described above can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0158] The embodiments described above can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0159] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0160] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0161] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for a lightning waveform discrimination model, characterized in that, include: Based on the broadband magnetic field sensors at each observation station, pulse data and dual-channel waveform data generated by lightning events are collected simultaneously. Based on the pulse data from each observation station, a time difference positioning algorithm is used for joint positioning to obtain the spatial location and type of the lightning event; Lightning events with a positioning error less than a preset threshold and a number of observation stations participating in the positioning not less than a preset number are selected, and the event type of the lightning event is used as the truth label of the event. Based on the timestamp information of the selected lightning events, the original pulse data and dual-channel waveform data are back-matched. The matched dual-channel waveform data is processed by channel selection and downsampling to generate single-channel waveform data; The single-channel waveform data is normalized to obtain standardized single-channel waveform data; The standardized single-channel waveform data is bound to the truth label corresponding to the event to form labeled training samples; The labeled training samples are used as the labeled dataset, and additional unlabeled dual-channel waveform data are collected as the unlabeled dataset. The labeled dataset and the unlabeled dataset are combined to form a hybrid dataset for model training. The teacher-student model was trained based on the hybrid dataset to obtain a lightning waveform identification model.

2. The method according to claim 1, characterized in that, The step of performing channel selection and downsampling processing on the matched dual-channel waveform data to generate single-channel waveform data includes: Calculate the peak-to-peak value of each channel in the dual-channel waveform data, and take the channel with the larger peak-to-peak value as the main channel input; The waveform data of the selected main channel is downsampled using max pooling, and the data dimension is compressed to the preset target sampling rate to generate single-channel waveform data.

3. The method according to claim 2, characterized in that, The lightning waveform identification model obtained by training the teacher-student model based on the hybrid dataset includes: The hybrid dataset is divided into a training set, a validation set, and a test set; Construct an average teacher framework that includes student and teacher models; The teacher and student models are trained using a training set. During the training process, data augmentation is applied to the input waveform data, and a dynamic learning rate adjustment mechanism is used to optimize the training process. After each training round, the performance of the current student model and / or teacher model is evaluated using the validation set, and the parameters of the model with the best performance on the validation set are saved. Once the training reaches the preset number of rounds or meets the early stopping condition, the best-performing student model is loaded and the lightning type identification accuracy of the best-performing student model is verified through the test set. When the accuracy reaches the preset index, the best-performing student model is used as the lightning waveform identification model.

4. The method according to claim 3, characterized in that, The teacher-student model comprises, in sequence, an input layer, a one-dimensional convolutional layer, a gated recurrent unit layer, and an output layer; wherein... The one-dimensional convolutional layer contains several convolutional structures, followed by batch normalization layers, ReLU activation functions, and Dropout layers. It also introduces a one-dimensional frequency dynamic convolution mechanism to extract local waveform features and dynamically adjust the weight distribution according to different frequency components in the temporal dimension. The gated recurrent unit layer is used to receive the feature sequence output by the one-dimensional convolutional layer, model the time dependency in the waveform data, and output global temporal features; The output layer, constructed based on several fully connected layers, is used to reduce the dimensionality of the global temporal features and transform the dimensionality-reduced global temporal features into prediction probability vectors corresponding to different lightning types through the Softmax activation function.

5. A method for identifying lightning waveforms, characterized in that, include: Based on the broadband magnetic field sensors at each observation station, pulse data and dual-channel waveform data generated by lightning events are collected simultaneously. The raw waveform data collected in real time by each observation station is input into the lightning waveform identification model trained by the training method of the lightning waveform identification model as described in any one of claims 1 to 4, generating a prediction probability vector corresponding to different lightning types, and selecting the lightning type with the highest probability value in the prediction probability vector as the local identification result of the observation station for this lightning event. The local identification result is bound and encapsulated with the timestamp corresponding to the lightning event and the metadata of this observation station to form an upload data packet; The uploaded data packets are sent to the central processing system, which receives and integrates the uploaded data packets from all observation stations for lightning events within the same time window, and generates a spatial distribution map of lightning activity that reflects the types and spatial distribution of lightning activity in the region.

6. A training device for a lightning waveform discrimination model, characterized in that, include: The data acquisition module is used to simultaneously acquire pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at each observation station; The data annotation module is used to perform joint positioning based on the pulse data of each observation station using a time difference positioning algorithm to obtain the spatial location and type of lightning events; Lightning events with a positioning error less than a preset threshold and a number of observation stations participating in the positioning not less than a preset number are selected, and the event type of the lightning event is used as the truth label of the event. Based on the timestamp information of the selected lightning events, the original pulse data and dual-channel waveform data are back-matched. The matched dual-channel waveform data is processed by channel selection and downsampling to generate single-channel waveform data; The single-channel waveform data is normalized to obtain standardized single-channel waveform data; The standardized single-channel waveform data is bound to the truth label corresponding to the event to form labeled training samples; The dataset construction module is used to take the labeled training samples as a labeled dataset and collect additional unlabeled dual-channel waveform data as an unlabeled dataset, and to combine the labeled dataset and the unlabeled dataset to form a hybrid dataset for model training. The model training module is used to train the teacher-student model based on the hybrid dataset to obtain the lightning waveform identification model.

7. A lightning waveform identification device, characterized in that, include: The data acquisition module is used to simultaneously acquire pulse data and dual-channel waveform data generated by lightning events based on broadband magnetic field sensors at each observation station; The data identification module is used to input the raw waveform data collected in real time by each observation station into the lightning waveform identification model trained by the training method of the lightning waveform identification model as described in any one of claims 1 to 4, generate prediction probability vectors corresponding to different lightning types, and select the lightning type with the highest probability value in the prediction probability vector as the local identification result of the observation station for this lightning event. The data packet generation module is used to bind and encapsulate the local identification result with the timestamp corresponding to the lightning event and the metadata of this observation station to form an upload data packet; The data packet upload module is used to send the uploaded data packets to the central processing system, so that the central processing system can receive and integrate the uploaded data packets from all observation stations for lightning events within the same time window, and generate a lightning activity spatial distribution map reflecting the lightning activity type and spatial distribution in the region.

8. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1-5.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 5.

Images