A method and system for detecting signals from unfamiliar unmanned aerial vehicles (UAVs) based on environment learning

Abstract

This invention provides a method and system for detecting signals from unfamiliar unmanned aerial vehicles (UAVs) based on environmental learning. The method includes: S1, converting spatial electromagnetic signals into time-frequency maps and detecting valid signals to obtain signal layer features; S2, extracting spatial features of each valid signal, including azimuth information and distance change trend information; S3, during the environmental learning period, clustering the signal layer features and spatial features, and storing the clusters as environmental signals in an environmental database; S4, during normal operation, comparing the current multidimensional features with the environmental database, and determining an anomaly and issuing a warning if no match is found; S5, continuously tracking the abnormal target, judging its movement state based on the spatial feature change trend, and adjusting the threat level. This invention establishes a background signal benchmark through environmental learning, reversing the detection logic from "matching known UAV models" to "excluding known environments," achieving open set detection, dynamic environmental adaptation, and hierarchical warning for any unfamiliar UAV signal.

Description

Technical Field

[0001] This invention relates to the fields of low-altitude safety and radio direction finding, and in particular to a method and system for detecting signals from unfamiliar unmanned aerial vehicles (UAVs) based on environmental learning. Background Technology

[0002] Currently, UAV detection technology mainly falls into three categories: radar detection, photoelectric detection, and radio signal detection. Radar detection is limited by low-altitude clutter interference and the physical bottleneck of the target's small radar cross-section; photoelectric detection is significantly affected by weather, visibility, and line-of-sight obstruction, making continuous all-weather monitoring difficult. In contrast, radio signal detection achieves target detection by passively listening to the UAV's downlink data transmission and image transmission signals. It has significant advantages such as silent detection, no weather or visual obstruction limitations, and all-weather, all-time operation, and has become the mainstream technology in the current UAV detection field. In the field of radio signal detection, existing technical solutions can be roughly divided into the following three categories based on their core detection logic: One type is the matching-based recognition technology based on a pre-built feature library. This technology pre-collects communication signal samples of known drone models, trains a deep learning model to build a feature library for the model, and matches the signal to be tested against templates in the library during actual detection to output the model recognition result. This type of solution is currently the mainstream in commercial applications, but its recognition capability is completely limited by the coverage of the training data. It cannot recognize signals from new models, modified drones with custom protocols, or unfamiliar drones with non-standard protocols, and has inherent detection blind spots. Furthermore, the feature library requires continuous manual updates, resulting in high maintenance costs and significant delays.

[0003] Another type is blind detection technology without a prior library. This type of solution attempts to achieve UAV link detection through the behavioral or statistical characteristics of signals, without relying on a specific UAV model library. However, it is usually designed only for specific communication systems, has a narrow scope of application, and is currently mostly in the laboratory research stage, only able to achieve limited detection capabilities for specific scenarios and specific protocol types.

[0004] In addition, the aforementioned existing technologies generally suffer from the following common defects: Poor environmental adaptability: All of them adopt static recognition logic and do not learn autonomously and dynamically model the electromagnetic environment of the deployment scenario. Under the electromagnetic environment fluctuations in different scenarios and at different times, they are prone to misjudging legitimate background signals as drone targets, making it difficult to balance the false alarm rate and the false detection rate. Lack of spatial feature fusion: Most existing solutions only output signal layer features (frequency, protocol type, etc.) without deeply integrating signal features with spatial features such as target orientation and distance. This makes it impossible to continuously track a single target, and also impossible to distinguish between harmless transit targets and intrusion targets based on motion status, thus failing to meet the practical needs of graded early warning.

[0005] Therefore, how to effectively detect unfamiliar drone signals without over-relying on prior drone model databases, while adapting to complex dynamic electromagnetic environments and integrating multi-dimensional features to complete target tracking and threat assessment, has become a pressing technical challenge in the field of low-altitude security. Summary of the Invention

[0006] This invention addresses the technical problems existing in the prior art by providing a method and system for detecting unfamiliar drone signals based on environment learning. By reversing the detection logic from "matching known drone models" to "excluding known environments" and combining multidimensional feature clustering and continuous spatial feature tracking, it achieves open set detection of any unfamiliar drone signal, dynamic environment adaptation, and threat level classification early warning.

[0007] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for detecting signals from unfamiliar unmanned aerial vehicles (UAVs) based on environment learning includes the following steps: Step S1: Convert the received spatial electromagnetic signal into a time-frequency diagram, and detect the effective signals in the time-frequency diagram to obtain the signal layer characteristics of each effective signal; Step S2: Extract the spatial features corresponding to each valid signal, wherein the spatial features include the target's orientation information and information reflecting the trend of target distance change over time; Step S3: During the preset environmental learning period, the signal layer features and spatial features of each effective signal are clustered as multidimensional features, and the clusters formed by the clustering are stored in the environmental database as environmental signal features. Step S4: During normal operation, compare the multidimensional features formed by the signal layer features and spatial features of the currently detected valid signal with the environmental database. If no match is found, the signal is determined to be an unfamiliar abnormal target and an early warning is generated. Step S5: Continuously track the unfamiliar and abnormal target, determine its motion state based on the trend of its spatial characteristics changing over time, and adjust the threat level based on the motion state.

[0008] As a further improvement to the method of the present invention: step S1 includes: The received spatial electromagnetic signals are converted into high-precision time-frequency diagrams and coarse time-frequency diagrams, respectively. The first deep learning model performs detection based on the coarse time-frequency map and outputs the position coordinates of each valid signal, as well as the signal category label, center frequency, bandwidth, and pulse width. The second deep learning model extracts the corresponding signal region from the high-precision time-frequency map based on the position coordinates, and outputs the protocol type or modulation method of each effective signal. Signal layer features are constructed based on the signal category label, protocol type or modulation method, center frequency, bandwidth, and pulse width.

[0009] As a further improvement to the method of the present invention: the signal category label includes OFDM-UAV, single-carrier digital modulation, and single-carrier analog modulation; among the protocol types or modulation methods, OFDM-UAV includes enhanced WIFI and image transmission protocols, single-carrier digital modulation includes spread spectrum modulation and frequency shift keying modulation, and single-carrier analog modulation includes frequency modulation, phase modulation, and amplitude modulation.

[0010] As a further improvement to the method of the present invention: in step S3, the multidimensional features include signal category label, protocol type or modulation method, frequency point, bandwidth, reporting time, azimuth and signal strength or estimated distance; the clustering adopts a density-based clustering algorithm, sets a neighborhood radius for each feature dimension, aggregates samples that meet the neighborhood conditions into clusters, and removes discrete noise samples with a sample number lower than a preset threshold.

[0011] As a further improvement to the method of the present invention: in the clustering, the neighborhood radius of the signal category label, protocol type or modulation method is zero, and the neighborhood radius of the reported time, azimuth and signal strength or estimated distance is smaller than the neighborhood radius of the frequency point and bandwidth.

[0012] As a further improvement to the method of the present invention: in step S5, the motion state includes stationary, approaching, moving away, and detouring; when it is determined that the target is in the approaching state, the threat level is increased; when it is determined that the target is in the moving away or detouring state, the threat level is maintained or decreased.

[0013] As a further improvement to the method of the present invention, the method further includes a dynamic update step of the environmental database: during normal operation, if a signal that is determined to be an abnormal target is confirmed as an environmental signal after continuous tracking, the multidimensional features of the signal are added to the environmental database.

[0014] This invention also provides an unfamiliar drone signal detection system based on environment learning, comprising: Signal receiving module, used to receive space electromagnetic signals; A time-frequency diagram generation module is used to convert the spatial electromagnetic signal into a time-frequency diagram; The signal detection module is used to detect the effective signals in the time-frequency diagram and obtain the signal layer features of each effective signal; The spatial feature extraction module is used to extract the spatial features corresponding to each valid signal. The spatial features include the target's orientation information and information reflecting the trend of target distance changing over time. The environment learning module is used to cluster the signal layer features and spatial features of each effective signal as multidimensional features within a preset environment learning period, and store each cluster formed by the clustering as an environmental signal feature in the environment database. An anomaly detection module is used to compare the multidimensional features formed by the signal layer features and spatial features of the currently detected valid signal with the environmental database during normal operation. If no match is found, the signal is determined to be an unfamiliar abnormal target and an early warning is generated. The threat assessment module is used to continuously track the unfamiliar and abnormal target, determine its movement status based on the trend of its spatial characteristics changing over time, and adjust the threat level based on the movement status.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that is programmed or configured to execute the environment-learning-based unfamiliar UAV signal detection method via a processor.

[0016] The present invention also provides a computer program product, including a computer program programmed or configured to execute the environment-learning-based unfamiliar UAV signal detection method via a processor.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention organically integrates five steps—signal layer feature extraction, spatial feature extraction, environmental learning and database construction, anomaly detection, and threat assessment—into a complete detection closed loop, achieving full-process coverage from signal input to tiered early warning output. At the detection logic level, this invention reverses the closed-set recognition mode of existing technologies that "match known drone models" to an open-set detection mode that "excludes known environments." By constructing a background signal benchmark for the deployment scenario through environmental learning, the system can completely break free from dependence on drone model feature databases. It can effectively detect signals from any unfamiliar drone without pre-collecting any drone samples, fundamentally solving the problem of missed detection of unknown targets. At the multi-dimensional feature fusion level, this invention treats signal layer features and spatial features as a unified whole throughout the entire process for joint clustering, comparison, and tracking assessment, rather than processing them separately. This achieves stable and continuous tracking of targets and discrimination of their motion states, effectively distinguishing between fixed background signals in the environment, temporarily passing harmless targets, and continuously approaching intrusive targets, and dynamically adjusting the threat level accordingly, avoiding the limitation of existing technologies that can only provide indiscriminate early warnings. Through the coordinated operation of the above five steps, this invention achieves closed-loop management of the entire process of detecting, tracking, assessing threats, and issuing graded early warnings for unfamiliar and abnormal targets, thus meeting the graded early warning requirements for differentiated responses to targets of different threat levels in low-altitude security combat scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart of the unfamiliar drone signal detection method based on environment learning in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the principle of the unfamiliar drone signal detection system based on environment learning in an embodiment of the present invention.

[0020] Figure 3 This is a flowchart illustrating the generation process based on the time-frequency diagram of the original signal in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the angular feature results of the environmental training clustering target in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the estimated distance feature results of the environmental training clustering target in an embodiment of the present invention.

[0023] Figure 6 This is an example diagram of parameters in the environmental signal database configuration file in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, this embodiment provides a method for detecting signals from unfamiliar drones based on environmental learning, including the following steps: Step S1: Convert the received spatial electromagnetic signal into a time-frequency diagram, and detect the effective signals in the time-frequency diagram to obtain the signal layer characteristics of each effective signal; Step S2: Extract the spatial features corresponding to each valid signal, where the spatial features include the target's orientation information and information reflecting the trend of target distance change over time; Step S3: During the preset environmental learning period, the signal layer features and spatial features of each effective signal are clustered as multi-dimensional features, and the clusters formed by the clusters are stored in the environmental database as environmental signal features. Step S4: During normal operation, compare the multidimensional features formed by the signal layer features and spatial features of the currently detected valid signal with the environmental database. If no match is found, the signal is determined to be an unfamiliar abnormal target and an early warning is generated. Step S5: Continuously track unfamiliar and abnormal targets, determine their motion status based on the trend of their spatial characteristics over time, and adjust the threat level based on the motion status.

[0026] The core technical concept of this embodiment differs fundamentally from existing technologies: Existing technologies are based on "whitelist" matching logic, triggering alarms only when signal features match a pre-built drone feature database; unmatched signals are considered safe or ignored, leading to the missed detection of a large number of unfamiliar drone signals. In contrast, this invention is based on "blacklist" matching logic, storing normal background signals confirmed through environmental learning in the deployment scenario into an environmental database. During normal operation, an alert is triggered only when signal features do not match the environmental database, thus shifting the detection focus from "known threats" to "unknown anomalies," achieving open-set detection of any unfamiliar drone signal.

[0027] To implement the above method, the overall system principle of this embodiment is as follows: Figure 2 As shown, the system is divided into eight core functional modules. The complete link from signal input to early warning output is as follows: signal receiving module, time-frequency map generation module, first-level deep learning recognition module, second-level deep learning recognition module, spatial feature extraction module, environment learning module, environment database module, and target early warning module. Among them, the second-level deep learning recognition module and the spatial feature extraction module are parallel processing modules, and more dimensional features can be added to improve system robustness. The specific implementation of the present invention will be described in detail below with reference to the above modules. It should be noted that the above system architecture is only a preferred embodiment of the method of the present invention, and the method of the present invention is not limited to this specific system architecture. Those skilled in the art can add, remove, or adjust the modules according to actual deployment needs.

[0028] Step S1 in this embodiment includes: Step 101: Convert the received spatial electromagnetic signals into high-precision time-frequency diagrams and coarse time-frequency diagrams respectively.

[0029] Specifically, after acquiring the zero intermediate frequency (IF) signal, the time-frequency graph generation module constructs a time-frequency graph for deep learning, including STFT calculation, high-precision time-frequency graph generation, and dimensionality-reduced time-frequency graph generation. Its main purpose is to convert a one-dimensional time-domain complex signal into a two-dimensional time-frequency graph, while generating both a high-precision time-frequency graph and a dimensionality-reduced coarse time-frequency graph to meet the different accuracy requirements of two-level subdivision recognition and one-level generalization recognition, respectively.

[0030] The STFT calculation unit uses a Hamming window as the sliding window function. Its core parameters are: window length of 2048 points, overlap of adjacent windows of 1024 points, and FFT points of 2048 points. For a 32ms time-domain signal with a sampling rate of 122.88MHz (total sampling points of 3,932,160), after STFT sliding calculation, the total number of points in the time domain is 3840, and the number of points in the frequency domain is 2048, ultimately generating a high-precision time-frequency graph of 2048×3840 dimensions, as shown below. Figure 3 As shown. The resulting high-precision time-frequency map has a frequency domain resolution of 122.88MHz / 2048=0.06MHz and a time domain resolution of 1024 / 122.88MHz≈8.3μs. This resolution can fully and clearly present the time-frequency characteristics of various signals such as LTE, WiFi, Bluetooth, shortwave broadcasting, drone remote control, data transmission, digital image transmission, and analog image transmission.

[0031] The dimensionality-reduced time-frequency map generation unit performs lossless dimensionality reduction extraction on the high-precision time-frequency map, extracting the frequency domain dimension at a 2x interval and the time domain dimension at a 4x interval, ultimately generating a coarse time-frequency map with 1024×960 dimensions. The frequency domain resolution of the dimensionality-reduced coarse time-frequency map is 0.12MHz, and the time domain resolution is 33.3μs, fully meeting the accuracy requirements for coarse identification of signal categories. Dimensionality reduction effectively reduces redundant information in the time-frequency map, improves the generalization learning effect of the first-level model, and significantly reduces the model size and subsequent inference time, thus improving system real-time performance. It should be noted that the parameters for constructing the time-frequency map can be set according to the actual equipment, as long as the time domain and frequency domain resolutions meet the above requirements.

[0032] Step 102: Detect based on a coarse time-frequency map using the first deep learning model, and output the position coordinates of each valid signal, as well as the signal category label, center frequency, bandwidth, and pulse width; The second deep learning model extracts the corresponding signal region from the high-precision time-frequency map based on the position coordinates, and outputs the protocol type or modulation method of each effective signal.

[0033] Step 103: Construct signal layer features based on signal category labels, protocol type or modulation method, center frequency, bandwidth, and pulse width.

[0034] In this embodiment, a two-level deep learning model architecture is used to detect valid signals in the time-frequency graph, thereby obtaining the signal layer features of each valid signal. The core function of the first-level deep learning recognition module is to employ generalization learning to minimize the detection of drone signals and completely remove restrictions on false alarm signals. Considering that while drone signals can be categorized into four main types—remote control, data transmission, digital image transmission, and analog image transmission—and that the communication modes and frame structures of various drones differ, their signal systems remain largely within the existing framework. Based on communication systems, they can be further classified into three categories: OFDM-UAV, single-carrier analog modulation, and single-carrier digital modulation. Therefore, when training drone data, the first-level model specifically divides drone signals into these three categories to enable the model to perform generalization learning of their signal systems.

[0035] To achieve the above functions, the first-level model needs to possess the following three capabilities: finding various signals in a time-frequency graph with complex background noise; performing simple broad category distinctions among the various signals; and determining the accurate location of each signal in the time-frequency graph. In this embodiment, the first-level signal category generalization recognition model adopts an improved Yolov5 network architecture. During the model training phase, the dimensionality-reduced time-frequency graph is labeled according to six categories: OFDM-UAV, single-carrier analog modulation, single-carrier digital modulation, WiFi, LTE, and other signal categories. Among them, known classic drone signals, enhanced WiFi signals with independent channels, and environmental signals similar to classic drone signals or enhanced WiFi are all classified into the OFDM-UAV category; the WiFi category only labels WiFi signals on public channels; and other signals in the environment whose types cannot be identified are classified into the other signal category. The labeled samples used for training cover all classic environmental signals within the required detection frequency band, various environmental interference signals with characteristics similar to drone signals, and real drone signals from various mainstream brands and models. During the training process, there is no need for targeted training for specific drone models; only generalization learning of signal systems needs to be completed. After the model enters the inference phase, it outputs the position coordinates of all valid signals in the time-frequency graph, including the start and end points in the time domain and the start and end points in the frequency domain. At the same time, it outputs the signal category label corresponding to each valid signal, thereby completing the generalization classification of the entire signal.

[0036] The core function of the Level 2 deep learning recognition module is to provide simplified recognition results for protocol types or modulation methods. The Level 2 signal subdivision recognition model group includes three major signal recognition sub-models: OFDM-UAV, single-carrier digital modulation, and single-carrier analog modulation. For WiFi, LTE, and other signal types, no Level 2 model is required; the recognition results from the Level 1 model can be directly reused.

[0037] In this embodiment, all secondary model groups adopt an improved ResNet18 convolutional neural network architecture. Their input is a high-precision time-frequency map cropped from the signal location coordinates output by the primary model. The specific implementation logic is as follows: based on the effective signal time-domain and frequency-domain location coordinates provided by the primary model, the time-frequency region corresponding to the target signal is accurately cropped from the 2048×3840 high-precision time-frequency map. This region is then placed into a fixed-size blank pixel map and fed into the corresponding secondary sub-model for recognition. The fixed size is set to 256×960, corresponding to a frequency bandwidth of 15.36MHz and a time-domain length of 8ms, which is sufficient to support fine-grained recognition for each category. If the target size is too large, a fixed-size time-frequency region is cropped from its center. In practical applications, the above dimensions can be adjusted according to specific project requirements.

[0038] Regarding the specific outputs of each secondary model, the OFDM-UAV signal recognition sub-model outputs simplified protocol type identification results, including Enhanced WiFi, DJI Lightbridge, O2, O3, O4, O4+, Autel Skylink 2.0, 3.0, and unknown protocols, totaling 9 categories. The single-carrier digital modulation signal recognition sub-model outputs modulation mode identification results, including LORA, GFSK, FSK, PSK, and unknown modulation, totaling 5 categories. The single-carrier analog modulation signal recognition sub-model outputs modulation mode identification results, including FM, PM, AM, SSB, and unknown modulation, totaling 5 categories.

[0039] At this point, the secondary models have completed the identification of protocol types or modulation schemes for different signal categories, and have summarized the identification results at the output of the secondary model group. The secondary model group finally outputs a simplified protocol type or modulation scheme of the target signal, which, combined with the signal category, center frequency, bandwidth, pulse width, and other core signal layer features output by the primary model, provides signal layer feature data for subsequent multi-dimensional feature library construction.

[0040] After the signal layer features are extracted in step S1, step S2 further extracts the spatial features corresponding to each valid signal. In step S2 of this embodiment, the spatial feature extraction module provides reference spatial layer feature data for subsequent multi-dimensional feature library construction, specifically including two steps: high-precision phase ratio direction finding calculation and signal strength and distance calculation.

[0041] The high-precision phase comparison direction finding process employs a phase comparison direction finding mode. The basic principle of phase comparison direction finding is to utilize multiple antenna array elements to simultaneously receive the same target signal, measure the phase difference between the signal arrival at different array elements, and inversely deduce the target azimuth angle based on the geometric relationship between the phase difference and the signal's direction of arrival. In this embodiment, the phase comparison direction finding module receives the zero-IF complex signal from the antenna array, extracts the phase difference between each channel, and calculates and outputs the target's azimuth angle information relative to the receiving array, serving as one of the spatial feature dimensions for subsequent multi-dimensional feature library construction. The phase comparison direction finding mode has the advantages of high accuracy and strong stability, and can accurately describe the motion characteristics of the target signal.

[0042] The signal strength and distance calculation stage estimates the trend of target distance changes by using signal strength. Since the target's equivalent radiated power is unknown, the distance estimated in this stage does not need to be consistent with the actual distance; it only needs to reflect whether the target is approaching or moving away within a certain period of time.

[0043] The aforementioned azimuth information and distance change trend information together constitute the spatial features of this invention. These spatial features, together with the signal layer features, will serve as input data for subsequent multi-dimensional feature library construction.

[0044] This embodiment, after obtaining signal layer features and spatial features separately, further fuses these features. This fusion is not a simple feature concatenation, but rather a joint modeling of signal layer features and spatial features at the feature level through multidimensional feature clustering. This aggregates discrete single-observation data into environmental signal targets with complete multidimensional attribute descriptions. Environmental learning is the key step in realizing this fusion process. Its main function is to package various existing environmental signals in the electromagnetic environment into storable multidimensional feature value combinations, specifically including two stages: environmental signal acquisition and multidimensional feature data clustering. The overall logic of the fusion process is as follows: first, an initial environmental database is established through centralized learning; then, the database is dynamically supplemented and updated through continuous learning; finally, discrete observation data are aggregated into complete environmental signal targets through multidimensional feature density clustering, providing a judgment benchmark for anomaly detection in step S4.

[0045] In the environmental signal acquisition phase, a combination of centralized learning and gap filling is employed. For newly deployed scenarios, an initial one-hour environmental learning process is conducted to record data on the vast majority of signals in the environment, constructing an initial environmental signal database. During subsequent operation, if new anomalous signal targets appear, their attributes can be determined by observing their signal and spatial characteristics over a certain period. For example, if an FM-type signal target is reported, and long-term tracking reveals that its angle and distance remain largely unchanged, this target is likely a newly emerging environmental signal. However, if the reported target is an FSK-modulated single-carrier signal, and its angle and distance show significant changes over time, it is highly probable that a drone intrusion has occurred. Anomalous targets confirmed as environmental signals are added to the environmental database, thereby completing continuous learning of the electromagnetic environment.

[0046] In the multidimensional feature data clustering stage, its core function is to perform multidimensional feature density clustering on each reported data during the training period, using signal layer features, spatial layer features, and reporting time, thus aggregating discrete single-report data into complete environmental signal targets. The clustering algorithm, as a crucial step in this method, directly determines the accuracy of signal tracking. In this embodiment, multidimensional features specifically include signal category labels, protocol type or modulation method, frequency point, bandwidth, reporting time, azimuth, and signal strength or estimated distance. It is understandable that the richer the dimensionality of the feature values, the higher the overall robustness of the system.

[0047] In terms of clustering strategy, this embodiment adopts a density-based clustering algorithm. The specific clustering strategy is as follows: signal category labels, protocol types, or modulation methods are used as signal-level features, with the constraint that they must be consistent, meaning the neighborhood radius of these features is zero; reported time, azimuth, and signal strength or estimated distance are used as spatial-level features, with stricter constraints, meaning their neighborhood radius is smaller, and the specific value can be set according to the device refresh rate; frequency and bandwidth are used as auxiliary features, with constraints that can be appropriately relaxed, meaning their neighborhood radius is larger than the above spatial-level features. In other words, the neighborhood radius of signal category labels, protocol types, or modulation methods is zero, and the neighborhood radius of reported time, azimuth, and signal strength or estimated distance is smaller than the neighborhood radius of frequency and bandwidth. During clustering, the above-mentioned neighborhood radius is set for each feature parameter to ensure that the clustering algorithm outputs results accurately according to the design requirements. Then, the minimum number of samples for each cluster is further set to eliminate discrete noise samples.

[0048] This method uses multi-dimensional feature values ​​for clustering, which can significantly improve clustering accuracy. Furthermore, it eliminates the need to pre-define the number of clusters and can adaptively match requirements based on the neighborhood radius of specified feature parameters. After clustering is complete, each resulting cluster is stored as an environmental signal in the environmental database for use in subsequent anomaly detection and comparison in step S4.

[0049] At this point, the environmental database has been completed. In step S4 of this embodiment, the environmental database is used to store the complete multi-dimensional feature parameters of the effective clusters output by clustering. The environmental signal database can be saved using a database or configuration file, storing the feature vectors of all environmental signals as the core benchmark for judging abnormal targets during subsequent normal operation. It also supports manual one-click addition, deletion, and editing of environmental signal features, enabling dynamic updates to the database. For newly appearing abnormal signals, if they are determined to be environmental signals, they can be directly added to the environmental signal database.

[0050] In this embodiment, the target warning module serves as the final output unit of the system, used for continuous tracking and threat assessment of detected abnormal targets. Once step S4 determines that a signal is an unfamiliar abnormal target and generates a warning, the target warning module takes over the continuous tracking of that target.

[0051] During tracking, the system determines the movement status of abnormal targets by analyzing the spatial characteristics of these targets over time. In this embodiment, movement status includes stationary, approaching, moving away, and detouring. When a target is determined to be approaching, it indicates that the target is moving closer to the protected area, and the threat level should be increased. When a target is determined to be moving away or detouring, it indicates that the intrusion risk is low or that the target is leaving, and the threat level should be maintained or decreased. The specific classification of threat levels and the format of the warning output can be flexibly designed by the actual user according to the specific security scenario, and this invention does not impose any restrictions on this.

[0052] Through the above mechanism, the system realizes closed-loop management of the entire process of discovering, tracking, threat assessment and graded early warning of unfamiliar and abnormal targets, effectively solving the problem of the disconnect between detection and early warning capabilities in existing technologies.

[0053] To further verify the effectiveness of the aforementioned closed-loop management process, especially the actual effect of multidimensional feature density clustering in environment learning, this embodiment presents a spatial feature clustering result diagram during the environment learning training process, such as... Figure 4 and Figure 5 As shown, during the environmental learning process, multiple reports of the same environmental signal target exhibited good clustering characteristics in both the angle and estimated distance dimensions. Among them, Figure 4 The mid-angle features remain relatively stable over time. Figure 5The estimated distance characteristics also showed relatively small fluctuations, indicating that the environmental signals were effectively aggregated into stable clusters after multidimensional feature density clustering. It should be noted that the above results include three types of UAV data, but the UAV data class, as a potential threat target, is not included in the environmental database; the remaining data classes are all included in the environmental database to construct an environmental signal feature benchmark. The above clustering results verify that the multidimensional feature density clustering method proposed in this invention can effectively distinguish between environmental background signals and UAV signals, providing a reliable basis for subsequent anomaly detection.

[0054] Based on the above clustering results, the environmental database persistently stores each cluster of environmental signals formed by clustering in the form of a configuration file. For example... Figure 6 As shown, the multi-dimensional feature records corresponding to three environmental signal clusters already included in the database are displayed. From left to right, they are OFDM-UAV targets, single-carrier digital modulation targets, and single-carrier analog modulation targets. The specific parameters included in each record are: target_idx represents the cluster's ID in the database; sig_type represents the signal category label; uav_type represents the protocol type or modulation method; freq_begin and freq_end represent the frequency range of the signal; band_min and band_max represent the minimum and maximum bandwidth values; sig_ang represents the angular feature center value of the cluster, and allow_ang_err represents the allowable error range for the angle; sig_dis represents the estimated distance center value of the cluster, and allow_dis_err represents the allowable error range for the distance. These parameters cover both signal layer features and spatial layer features, collectively forming the complete multi-dimensional feature vector of a single environmental signal cluster.

[0055] It is understood that the above parameters are only illustrative examples. In practical applications, the multidimensional feature parameters can be further expanded according to specific deployment scenarios and detection requirements, such as adding time-domain pulse width information and frequency hopping range, to further improve the robustness and environmental adaptability of the system.

[0056] In summary, this embodiment has the following beneficial effects: First, this embodiment employs a method combining deep learning generalized training, multi-dimensional feature value target tracking, and environment learning for abnormal signal detection. The fundamental difference from other detection methods lies in this: existing technologies use a "whitelist" mechanism, only triggering alarms for signals registered in the aircraft model feature database, ignoring signals outside the database regardless of whether they are drones; this invention uses a "blacklist" mechanism, only ignoring known environmental signals confirmed through environment learning, while triggering alarms for all other newly appearing signals. This fundamental reversal of detection logic allows this invention to completely eliminate its dependence on the aircraft model feature database, enabling open-set detection of any unfamiliar drone signal without prior collection of any drone samples, fundamentally solving the inherent defect of existing technologies in detecting unknown targets.

[0057] Second, this embodiment achieves dynamic tracking and anomaly detection of environmental signals through an environmental learning mechanism, enabling the system to adapt to the electromagnetic environment of different deployment scenarios. Specifically, in a newly deployed scenario, the system first establishes an initial environmental database through centralized learning, and then continuously supplements and updates the environmental signal characteristics through continuous learning during long-term operation. This allows the equipment to adapt to the dynamic changes in the electromagnetic environment, achieving long-term stable operation that becomes more accurate with use, and significantly reducing the false alarm rate and missed detection rate in complex electromagnetic environments.

[0058] Third, this embodiment, by fusing the spatial and signal layer features of the target signal, provides a curve showing the change of the target's spatial features over time. This allows for accurate identification of whether the target is a fixed environmental signal, a harmless target temporarily passing through, or a continuously approaching, unfamiliar drone target with intrusion risk, and dynamically adjusts the threat level accordingly. This mechanism achieves closed-loop management of the entire process of detecting, tracking, assessing threats, and issuing graded warnings for unfamiliar and abnormal targets, meeting the graded warning requirements for differentiated responses to targets of different threat levels in actual security scenarios.

[0059] This embodiment also provides an unfamiliar drone signal detection system based on environment learning, including: Signal receiving module, used to receive space electromagnetic signals; The time-frequency diagram generation module is used to convert spatial electromagnetic signals into time-frequency diagrams; The signal detection module is used to detect the effective signals in the time-frequency graph and obtain the signal layer characteristics of each effective signal; The spatial feature extraction module is used to extract the spatial features corresponding to each valid signal. The spatial features include the target's orientation information and information reflecting the trend of target distance change over time. The environment learning module is used to cluster the signal layer features and spatial features of each effective signal as multi-dimensional features within a preset environment learning period, and store the clusters formed by the clusters as environmental signal features in the environment database. The anomaly detection module is used to compare the multidimensional features formed by the signal layer features and spatial features of the currently detected valid signals with the environmental database during normal operation. If no match is found, the signal is determined to be an unfamiliar abnormal target and an early warning is generated. The threat assessment module is used to continuously track unfamiliar and abnormal targets, determine their movement status based on the trend of their spatial characteristics over time, and adjust the threat level based on the movement status.

[0060] This embodiment also provides a computer-readable storage medium storing a computer program / instructions that are programmed or configured to execute an environment-learning-based unfamiliar drone signal detection method via a processor.

[0061] This embodiment also provides a computer program product, including a computer program / instructions that are programmed or configured to execute an environment-learning-based unfamiliar drone signal detection method via a processor.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for detecting signals from unfamiliar unmanned aerial vehicles (UAVs) based on environmental learning, characterized in that, Includes the following steps: Step S1: Convert the received spatial electromagnetic signal into a time-frequency diagram, and detect the effective signals in the time-frequency diagram to obtain the signal layer characteristics of each effective signal; Step S2: Extract the spatial features corresponding to each valid signal, wherein the spatial features include the target's orientation information and information reflecting the trend of target distance change over time; Step S3: During the preset environmental learning period, the signal layer features and spatial features of each effective signal are clustered as multidimensional features, and the clusters formed by the clustering are stored in the environmental database as environmental signal features. Step S4: During normal operation, compare the multidimensional features formed by the signal layer features and spatial features of the currently detected valid signal with the environmental database. If no match is found, the signal is determined to be an unfamiliar abnormal target and an early warning is generated. Step S5: Continuously track the unfamiliar and abnormal target, determine its motion state based on the trend of its spatial characteristics changing over time, and adjust the threat level based on the motion state.

2. The method for detecting unfamiliar UAV signals based on environment learning according to claim 1, characterized in that, Step S1 includes: The received spatial electromagnetic signals are converted into high-precision time-frequency diagrams and coarse time-frequency diagrams, respectively. The first deep learning model performs detection based on the coarse time-frequency map and outputs the position coordinates of each valid signal, as well as the signal category label, center frequency, bandwidth, and pulse width. The second deep learning model extracts the corresponding signal region from the high-precision time-frequency map based on the position coordinates, and outputs the protocol type or modulation method of each effective signal. Signal layer features are constructed based on the signal category label, protocol type or modulation method, center frequency, bandwidth, and pulse width.

3. The method for detecting unfamiliar UAV signals based on environment learning according to claim 2, characterized in that, The signal category labels include OFDM-UAV, single-carrier digital modulation, and single-carrier analog modulation; among the protocol types or modulation methods, OFDM-UAV includes enhanced WIFI and image transmission protocols, single-carrier digital modulation includes spread spectrum modulation and frequency shift keying modulation, and single-carrier analog modulation includes frequency modulation, phase modulation, and amplitude modulation.

4. The method for detecting unfamiliar UAV signals based on environment learning according to claim 1, characterized in that, In step S3, the multidimensional features include signal category labels, protocol type or modulation method, frequency point, bandwidth, reporting time, azimuth and signal strength or estimated distance; the clustering adopts a density-based clustering algorithm, sets a neighborhood radius for each feature dimension, aggregates samples that meet the neighborhood conditions into clusters, and removes discrete noise samples with a sample number lower than a preset threshold.

5. The method for detecting unfamiliar UAV signals based on environment learning according to claim 4, characterized in that, In the clustering, the neighborhood radius of the signal category label, protocol type, or modulation method is zero, and the neighborhood radius of the reported time, location, signal strength, or estimated distance is smaller than the neighborhood radius of the frequency point and bandwidth.

6. The method for detecting unfamiliar UAV signals based on environment learning according to claim 1, characterized in that, In step S5, the movement state includes stationary, approaching, moving away, and detouring; when the target is determined to be in an approaching state, the threat level is increased; when the target is determined to be in a moving away or detouring state, the threat level is maintained or decreased.

7. The method for detecting unfamiliar UAV signals based on environmental learning according to any one of claims 1 to 6, characterized in that, The method also includes a dynamic update step for the environmental database: during normal operation, if a signal identified as an abnormal target is confirmed as an environmental signal after continuous tracking, the multidimensional features of the signal are added to the environmental database.

8. A signal detection system for unfamiliar unmanned aerial vehicles (UAVs) based on environmental learning, characterized in that, include: Signal receiving module, used to receive space electromagnetic signals; A time-frequency diagram generation module is used to convert the spatial electromagnetic signal into a time-frequency diagram; The signal detection module is used to detect the effective signals in the time-frequency diagram and obtain the signal layer features of each effective signal; The spatial feature extraction module is used to extract the spatial features corresponding to each valid signal. The spatial features include the target's orientation information and information reflecting the trend of target distance changing over time. The environment learning module is used to cluster the signal layer features and spatial features of each effective signal as multidimensional features within a preset environment learning period, and store each cluster formed by the clustering as an environmental signal feature in the environment database. An anomaly detection module is used to compare the multidimensional features formed by the signal layer features and spatial features of the currently detected valid signal with the environmental database during normal operation. If no match is found, the signal is determined to be an unfamiliar abnormal target and an early warning is generated. The threat assessment module is used to continuously track the unfamiliar and abnormal target, determine its movement status based on the trend of its spatial characteristics changing over time, and adjust the threat level based on the movement status.

9. A computer-readable storage medium storing a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute, via a processor, the unfamiliar UAV signal detection method based on environmental learning as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute, via a processor, the unfamiliar UAV signal detection method based on environmental learning as described in any one of claims 1 to 7.

Images