Multi-mode electromagnetic environment detection system and method

Abstract

The invention relates to the technical field of electromagnetic environment detection, and discloses a multi-mode electromagnetic environment detection system and method. The system acquires multi-modal data such as electromagnetic field intensity and spectrum distribution through a multi-modal sensor array to generate a data set, and realizes interference source identification and dynamic adjustment of detection parameters through nonlinear signal processing, depth feature extraction based on a convolutional neural network, topology analysis of a graph neural network and adaptive optimization of reinforcement learning. And an electromagnetic field simulation module is also arranged to simulate electromagnetic wave propagation and early warn abnormal radiation, and an abnormal mode library is constructed to identify potential interference types. The system can comprehensively and accurately detect an electromagnetic environment, efficiently position an interference source, adaptively adjust detection parameters, improve detection accuracy, real-time performance and system stability, and effectively meet complex electromagnetic environment monitoring requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic environment detection, and in particular to a multi-modal electromagnetic environment detection system and method. Background Art

[0002] In modern society, the electromagnetic environment is becoming increasingly complex, and the widespread use of various electronic devices has made electromagnetic interference (EMI) a more prominent problem. From everyday communications equipment and household appliances to large-scale industrial equipment and automated production lines, and even various electronic equipment in the military, all are constantly transmitting and receiving electromagnetic signals. These signals interweave, creating a complex and ever-changing electromagnetic environment.

[0003] In the communications sector, electromagnetic interference can lead to reduced signal transmission quality, signal interruptions, increased bit error rates, and other issues, seriously impacting the reliability and stability of communications. For example, in densely populated urban areas, where numerous base stations, mobile phones, and Wi-Fi devices operate simultaneously, electromagnetic signals interfere with each other, compromising communication quality and resulting in a poor user experience. In aerospace, electromagnetic interference can affect the proper operation of aircraft navigation systems, communication systems, and various electronic equipment, posing a serious threat to flight safety. If an aircraft's navigation system is affected by electromagnetic interference, it can cause navigation errors, deviate from its route, and lead to serious accidents.

[0004] In industrial production, electromagnetic interference can also affect production equipment, leading to equipment failure and reduced production efficiency. For example, in automated factories, a large number of industrial robots and control systems rely on electronic signals for precise control. Once exposed to electromagnetic interference, these devices may experience control errors, produce substandard products, or even cause equipment damage.

[0005] Traditional electromagnetic environment monitoring technologies have numerous limitations. Firstly, their detection methods are relatively limited, often measuring only single parameters such as electromagnetic field strength or spectrum distribution, failing to fully capture information about the electromagnetic environment. This results in inaccurate and incomplete assessments of the electromagnetic environment, making it difficult to detect potential electromagnetic interference issues. Secondly, traditional monitoring technologies lack the ability to effectively identify and locate interference sources in complex electromagnetic environments. When multiple interference sources are present simultaneously, it is difficult to accurately determine their location and nature, making it impossible to implement targeted interference suppression measures.

[0006] Furthermore, traditional detection systems often lack adaptability due to their inability to automatically adjust detection parameters based on changes in the electromagnetic environment. Under varying environmental conditions, such as significant fluctuations in temperature and humidity, the detection system's performance may be affected, leading to inaccurate results. Furthermore, as the electromagnetic environment becomes increasingly complex, the requirements for real-time performance and accuracy of detection systems are becoming increasingly stringent, making it difficult for traditional detection technologies to meet these requirements. Summary of the Invention

[0007] The object of the present invention is to provide a multi-modal electromagnetic environment detection system and method to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multimodal electromagnetic environment detection system, the system comprising:

[0009] Multi-source data acquisition module, used to synchronously collect electromagnetic field intensity, spectrum distribution, polarization direction and ambient temperature and humidity data through a multi-modal sensor array to generate a multi-modal spatiotemporal dataset;

[0010] a signal preprocessing module, configured to perform nonlinear signal processing on the multimodal spatiotemporal data set to generate a time-frequency signal sequence;

[0011] A feature extraction module is used to build a deep feature extraction model based on a convolutional neural network, perform multi-scale spatiotemporal feature fusion on the time-frequency signal sequence, extract the time-frequency sparsity and spatial correlation characteristics of electromagnetic interference, and generate a multimodal feature coding matrix;

[0012] An interference source identification module is used to design a topological analysis model based on a graph neural network, construct a node relationship diagram of the electromagnetic radiation propagation path according to the multimodal feature coding matrix, and generate an interference source location probability distribution map;

[0013] The dynamic parameter adjustment module is used to establish an adaptive optimization strategy based on reinforcement learning, dynamically adjust the detection sensitivity and anti-interference filter parameters according to the interference source positioning probability distribution map, and generate environmental adaptive detection instructions.

[0014] Preferably, the nonlinear signal processing uses a wavelet packet decomposition and adaptive filtering combined algorithm to separate broadband noise components and steady-state electromagnetic characteristics, including:

[0015] The electromagnetic field intensity signal is decomposed by wavelet packets and a sub-band energy distribution sequence is generated through multi-resolution analysis;

[0016] Construct an adaptive filter coefficient update equation, use the high-frequency sub-band component as the reference noise input, and iteratively optimize the filter transfer function;

[0017] The kurtosis detection algorithm is used to identify abnormal frequency points of steady-state spectrum components and segment normal electromagnetic environment signal segments.

[0018] Preferably, the construction of the deep feature extraction model includes:

[0019] In the feature encoding layer, dilated convolution kernels are introduced to capture long-range temporal dependencies and output multi-scale feature map tensors;

[0020] In the feature fusion layer, the channel attention mechanism is used to weightedly aggregate spatiotemporal features of different scales, and residual connections are superimposed to prevent gradient disappearance;

[0021] The parameters of the feature encoding layer and the fusion layer are optimized through a contrastive learning strategy to minimize the ratio of the intra-class feature distance to the inter-class feature distance.

[0022] Preferably, the design of the topology analysis model includes:

[0023] Define the electromagnetic radiation source as a graph node, the propagation path attenuation coefficient as the edge weight, and construct a directed weighted graph structure;

[0024] A graph embedding algorithm is used to map nodes into a low-dimensional vector space and calculate the similarity matrix between nodes;

[0025] The spectral clustering algorithm is used to divide the high-probability interference source areas and generate electromagnetic radiation heat maps.

[0026] Preferably, the establishment of the adaptive optimization strategy includes:

[0027] The state space is defined as a combination vector of signal-to-noise ratio, spectrum flatness, and ambient temperature and humidity change rate;

[0028] Construct a proximal policy optimization algorithm framework, using advantage function estimation and importance sampling to optimize the stability of the policy network;

[0029] The reward function is designed as the weighted difference between interference suppression gain and detection energy consumption, driving the policy network to generate optimal parameter adjustment instructions.

[0030] Preferably, the system further comprises:

[0031] The electromagnetic field simulation module is used to establish an electromagnetic wave propagation model based on Maxwell's equations and simulate the superposition effect of multipath effect and obstacle diffraction in real time;

[0032] The measured data is input into the simulation model to generate a virtual field strength distribution map, and the abnormal radiation warning signal is triggered through difference analysis.

[0033] Preferably, the generation of the multimodal spatiotemporal dataset includes:

[0034] Synchronously collect antenna array received signals, spectrum analyzer scan data, and temperature and humidity sensor readings;

[0035] The non-negative matrix factorization algorithm is used to reduce the feature dimension of heterogeneous data and construct a low-rank approximate data cube;

[0036] The sensor quantization error is eliminated by wavelet threshold denoising method to enhance the detectability of weak signals.

[0037] Preferably, the generation of the environment adaptive detection instruction includes:

[0038] Establish a multi-constraint optimization model with detection accuracy, real-time performance, and equipment life as boundary conditions;

[0039] The interior point method is used to solve the convex optimization problem with nonlinear constraints and generate feasible solution sets of filter cutoff frequency and gain.

[0040] The multi-objective priority of the solution set is evaluated by the entropy weight method, and the optimal parameter configuration scheme is output.

[0041] Preferably, the system further comprises:

[0042] An abnormal pattern library based on an autoencoder is constructed to perform unsupervised feature learning on historical electromagnetic interference data; a dynamic mode decomposition algorithm is used to extract transient patterns of real-time signals and generate potential interference type identification results.

[0043] Preferably, the present invention further includes a multimodal electromagnetic environment detection method, the method comprising the following steps:

[0044] Step 1: Use a multimodal sensor array to synchronously collect electromagnetic field intensity, spectrum distribution, polarization direction, and ambient temperature and humidity data to generate a multimodal spatiotemporal dataset;

[0045] Step 2: performing nonlinear signal processing on the multimodal spatiotemporal dataset to generate a time-frequency signal sequence;

[0046] Step 3: Construct a deep feature extraction model based on a convolutional neural network, perform multi-scale spatiotemporal feature fusion on the time-frequency signal sequence, extract the time-frequency sparsity and spatial correlation characteristics of the electromagnetic interference, and generate a multimodal feature coding matrix;

[0047] Step 4: Based on the topological analysis model of the graph neural network, a node relationship diagram of the electromagnetic radiation propagation path is constructed according to the multimodal feature coding matrix, and a probability distribution diagram of the interference source location is generated;

[0048] Step 5: Using the adaptive optimization strategy of reinforcement learning, dynamically adjust the detection sensitivity and anti-interference filter parameters according to the interference source positioning probability distribution map to generate an environment adaptive detection instruction.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The multimodal electromagnetic environment detection system and method of the present invention have significant beneficial effects in many aspects. In terms of comprehensive detection, the electromagnetic field intensity, spectrum distribution, polarization direction and ambient temperature and humidity data are synchronously collected through a multimodal sensor array to generate a multimodal spatiotemporal data set, thereby achieving the acquisition of all-round information on the electromagnetic environment. Compared with traditional technologies that can only detect a single parameter, this method can provide richer and more comprehensive data, allowing users to have a deeper and more accurate understanding of the electromagnetic environment, and provide a solid foundation for subsequent analysis and decision-making. For example, in a complex urban electromagnetic environment, not only can the electromagnetic field intensity of different frequency bands be grasped, but also the polarization direction of the signal and the impact of ambient temperature and humidity on the electromagnetic signal can be understood, thereby more accurately evaluating the quality of the electromagnetic environment.

[0051] In terms of signal processing and feature extraction, nonlinear signal processing, such as wavelet packet decomposition and adaptive filtering combined algorithms, is used to effectively separate broadband noise components from steady-state electromagnetic characteristics, generating high-quality time-frequency signal sequences. A deep feature extraction model based on a convolutional neural network is constructed to perform multi-scale spatiotemporal feature fusion, extracting the time-frequency sparsity and spatial correlation characteristics of electromagnetic interference and generating a multimodal feature encoding matrix. This deep feature extraction method can extract key information from complex electromagnetic signals, greatly improving the accuracy of electromagnetic interference feature recognition. Even in the presence of weak signals or strong interference backgrounds, valuable features can be accurately extracted, providing strong support for subsequent interference source identification.

[0052] The interference source identification module uses a graph neural network-based topology analysis model to construct a node relationship diagram of the electromagnetic radiation propagation path based on a multimodal feature encoding matrix, generating a probability distribution map for the interference source location. This innovative method can intuitively display the electromagnetic radiation propagation path and the areas where the interference source may be located, and is more accurate and efficient than traditional positioning methods. In practical applications, whether in complex indoor electronic equipment environments or in vast outdoor communication base station areas, it can quickly and accurately locate interference sources, significantly reducing the time and labor costs of troubleshooting interference sources.

[0053] The dynamic parameter adjustment module establishes an adaptive optimization strategy based on reinforcement learning. It dynamically adjusts detection sensitivity and anti-interference filter parameters based on the interference source location probability distribution map, generating environmentally adaptive detection instructions. This module enables the system to automatically optimize detection parameters based on real-time changes in the electromagnetic environment, improving detection accuracy and efficiency. For example, if electromagnetic interference intensity suddenly increases, the system can automatically increase detection sensitivity and adjust anti-interference filter parameters to effectively suppress interference and ensure the reliability of detection results. Furthermore, by establishing a multi-constraint optimization model that comprehensively considers factors such as detection accuracy, real-time performance, and equipment lifespan, it uses the interior point method to solve and the entropy weight method to evaluate, outputting the optimal parameter configuration solution, further improving system performance and stability.

[0054] Furthermore, the system's electromagnetic field simulation module establishes an electromagnetic wave propagation model based on Maxwell's equations, simulating the combined effects of multipath and obstacle diffraction in real time. Measured data is input into the simulation model to generate a virtual field strength distribution map, which triggers abnormal radiation warning signals through differential analysis. This function not only predicts changes in the electromagnetic environment in advance but also promptly detects abnormal radiation, providing an important means for ensuring electromagnetic environmental safety. An autoencoder-based abnormal pattern library, combined with a dynamic mode decomposition algorithm, performs unsupervised feature learning on historical electromagnetic interference data, extracts transient patterns from real-time signals, and generates potential interference type identification results. This further enhances the system's ability to analyze and identify electromagnetic interference, improving the intelligence and reliability of the entire detection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a working principle diagram of the multi-modal electromagnetic environment detection system of the present invention;

[0056] Figure 2 Flowchart built for deep feature extraction model;

[0057] Figure 3 This is the working principle diagram of the electromagnetic field simulation module;

[0058] Figure 4 A graph of steps generated for the environment adaptive detection instructions. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] See also Figure 1-Figure 4 The present invention aims to provide a multimodal electromagnetic environment detection system and method to achieve comprehensive and accurate detection of the electromagnetic environment and interference source location, and to adaptively adjust detection parameters based on environmental changes. The specific implementation methods of the present invention are described in detail below.

[0061] The multimodal electromagnetic environment detection system mainly includes a multi-source data acquisition module, a signal preprocessing module, a feature extraction module, an interference source identification module and a dynamic parameter adjustment module.

[0062] During system operation, the multi-source data acquisition module comes into play first. This module uses a multimodal sensor array to synchronously collect electromagnetic field intensity, spectrum distribution, polarization direction, and ambient temperature and humidity data, thereby generating a multimodal spatiotemporal dataset. For example, an antenna array is used to receive electromagnetic field signals to obtain electromagnetic field intensity data, a spectrum analyzer is used to scan the spectrum distribution information, a polarization sensor is used to measure polarization direction, and a temperature and humidity sensor is used to collect ambient temperature and humidity data.

[0063] The signal preprocessing module performs nonlinear signal processing on the multimodal spatiotemporal dataset generated by the multi-source data acquisition module, ultimately generating a time-frequency signal sequence. This process removes noise and interference from the data, providing more reliable data for subsequent analysis.

[0064] The feature extraction module builds a deep feature extraction model based on convolutional neural networks. This model fuses multi-scale spatiotemporal features of the time-frequency signal sequence to extract the time-frequency sparsity and spatial correlation characteristics of electromagnetic interference and generate a multimodal feature encoding matrix. The powerful feature extraction capabilities of convolutional neural networks can extract key features from complex time-frequency signals, laying the foundation for interference source identification.

[0065] The interference source identification module is designed based on a topological analysis model using a graph neural network. Based on a multimodal feature encoding matrix, it constructs a node relationship diagram of the electromagnetic radiation propagation path, which in turn generates a probability distribution map for interference source location. This approach provides an intuitive understanding of the areas where interference sources are likely to be located, providing strong support for subsequent interference source location.

[0066] The dynamic parameter adjustment module establishes an adaptive optimization strategy based on reinforcement learning. It dynamically adjusts detection sensitivity and anti-interference filter parameters based on the interference source location probability distribution map, generating environmentally adaptive detection instructions. This enables the system to automatically adjust detection parameters based on changes in the actual electromagnetic environment, improving detection accuracy and efficiency.

[0067] The present invention will be further described below in conjunction with Examples 1 to 6:

[0068] Example 1:

[0069] In the process of the multi-source data acquisition module generating a multimodal spatiotemporal data set and the signal preprocessing module performing nonlinear signal processing, the specific implementation method is as follows:

[0070] Generating a multimodal spatiotemporal dataset first requires the simultaneous acquisition of antenna array received signals, spectrum analyzer scan data, and temperature and humidity sensor readings. The antenna array typically consists of multiple antennas of different types arranged in a specific layout, such as uniform distribution or an optimized geometric structure, to ensure comprehensive reception of electromagnetic field signals from different directions and characteristics. When scanning with the spectrum analyzer, appropriate parameters such as the scanning frequency range and scanning interval must be set to ensure comprehensive and accurate spectrum distribution information is obtained. The temperature and humidity sensor is installed in a suitable location near the testing equipment to avoid interference from heat sources, cold sources, or other factors, ensuring accurate measurement of ambient temperature and humidity.

[0071] After collecting this heterogeneous data, we use a non-negative matrix factorization algorithm to reduce its dimensionality. The principle of this algorithm is to decompose a non-negative matrix into the product of two non-negative matrices, namely, V = WH, where V is the original heterogeneous data matrix, W is the basis matrix, and H is the coefficient matrix. This decomposition can reduce the data's dimensionality while preserving its key features, thus reducing the computational complexity of subsequent processing. During the decomposition process, appropriate parameters such as decomposition accuracy and number of iterations should be selected based on the data characteristics and actual needs.

[0072] Wavelet threshold denoising eliminates sensor quantization errors. Based on the principles of wavelet transform, wavelet threshold denoising decomposes the signal and then processes the wavelet coefficients according to a threshold. During this process, it is necessary to select appropriate wavelet basis functions and thresholds. Different wavelet basis functions decompose the signal differently, so the appropriate threshold should be selected based on the signal's characteristics. Determining the threshold is also crucial. A threshold that is too high may over-remove the signal, resulting in distortion; a threshold that is too low may not effectively remove noise. Wavelet threshold denoising enhances the detectability of weak signals and improves data quality.

[0073] When performing nonlinear signal processing in the signal preprocessing module, a combined wavelet packet decomposition and adaptive filtering algorithm is used to separate broadband noise components from steady-state electromagnetic signatures. The electromagnetic field intensity signal is first subjected to wavelet packet decomposition, and a subband energy distribution sequence is generated through multi-resolution analysis. Wavelet packet decomposition further decomposes the high-frequency portion of the signal, building upon wavelet decomposition, enabling a more detailed analysis of the signal's frequency components. During the decomposition process, appropriate wavelet packet basis functions and the number of decomposition levels must be selected. Too many decomposition levels may result in excessive computational effort, while too few decomposition levels may prevent adequate signal analysis.

[0074] Then, an adaptive filter coefficient update equation is constructed, using the high-frequency subband components as reference noise inputs to iteratively optimize the filter transfer function. The coefficient update equation for an adaptive filter is generally based on the least mean square error (LMS) algorithm or the recursive least squares (RLS) algorithm. Taking the LMS algorithm as an example, its coefficient update formula is w(n+1)=w(n)+μe(n)x(n), where w(n) is the filter coefficient vector at the nth iteration, μ is the step size factor, which controls the speed of the filter coefficient update. A large μ may cause filter instability, while a small μ will slow convergence. e(n) is the error signal at the nth iteration, that is, the difference between the expected output and the actual output. x(n) is the input signal at the nth iteration. Through continuous iteration, the filter transfer function is optimized to better suppress noise.

[0075] Finally, a kurtosis detection algorithm is used to identify abnormal frequencies in the steady-state spectrum components and segment the normal electromagnetic environment signal. Kurtosis is a statistic used to measure the distribution characteristics of a signal. For normal steady-state signals, the kurtosis value has a certain range. When abnormal frequencies are present in the signal, the kurtosis value changes significantly. By setting an appropriate kurtosis threshold, abnormal frequencies can be identified and normal electromagnetic environment signal segments can be segmented, providing accurate data for subsequent analysis.

[0076] Example 2:

[0077] When the feature extraction module constructs a deep feature extraction model, the specific implementation is as follows:

[0078] The deep feature extraction model is built on a convolutional neural network and primarily consists of a feature encoding layer and a feature fusion layer. A dilated convolution kernel is introduced in the feature encoding layer to capture long-range temporal dependencies and output a multi-scale feature map tensor. The dilated convolution kernel adds a dilation rate parameter to the regular convolution kernel. Assuming a regular convolution kernel size of k×k and a dilation rate of d, the spacing between kernel elements during the convolution operation is d. By setting different dilation rates, the convolution kernel can sense signals at different scales, thereby capturing long-range temporal dependencies. For example, when the dilation rate d = 1, the dilated convolution kernel is equivalent to a regular convolution kernel; when d = 2, the spacing between kernel elements is 1 pixel, enabling it to sense signals over a wider range. In practical applications, the appropriate dilation rate combination should be selected based on the signal's time series length and feature distribution.

[0079] In the feature fusion layer, a channel attention mechanism is used to weightedly aggregate spatiotemporal features at different scales, and residual connections are superimposed to prevent vanishing gradients. The channel attention mechanism weights features across different channels to highlight important channel features. Its implementation generally includes steps such as global average pooling, a fully connected layer, an activation function, and weight calculation. First, global average pooling is performed on the feature map to compress the features of each channel into a single value, resulting in a global feature representation of the channel. The global features are then processed through a fully connected layer and an activation function to obtain a weight for each channel. Finally, the weights are multiplied by the original feature map to achieve weighted aggregation of features across different channels. Residual connections directly add the input to the output: y = x + F(x), where x is the input, F(x) is the output of a layer of the convolutional neural network, and y is the output after the residual connection. This approach facilitates gradient propagation during backpropagation, preventing the vanishing gradient problem and improving model training performance.

[0080] In order to optimize the parameters of the feature encoding layer and the fusion layer, a contrastive learning strategy is adopted. The core idea of the contrastive learning strategy is to minimize the ratio of the intra-class feature distance to the inter-class feature distance. Suppose there are two samples x i and x j If they belong to the same class, we hope that the distance d(x i ,x j ) is as small as possible; if they belong to different classes, the distance is expected to be as large as possible. By defining the contrast loss function Where N is the number of samples, M is the number of comparison samples, and z i is the sample x i The feature vector after model processing, z i + is x i The eigenvectors of positive samples belonging to the same class, sim, represent the similarity between two eigenvectors, typically calculated using cosine similarity. τ is a temperature parameter used to adjust the smoothness of the contrast loss. By minimizing this loss function, the model parameters are continuously optimized, enabling the model to better extract the time-frequency sparsity and spatial correlation characteristics of electromagnetic interference and generate a multimodal feature encoding matrix.

[0081] Example 3:

[0082] When designing the topology analysis model for the interference source identification module, the specific implementation is as follows:

[0083] Electromagnetic radiation sources are defined as graph nodes, and the propagation path attenuation coefficients are defined as edge weights to construct a directed weighted graph structure. In practical applications, electromagnetic radiation sources may be distributed in different locations and have different radiation intensities and frequency characteristics. Each electromagnetic radiation source is considered a node in the graph, and node attributes may include the source's location, radiation intensity, and frequency. The propagation path attenuation coefficient is determined based on the attenuation characteristics of electromagnetic waves during propagation and is dependent on factors such as propagation distance and the propagation medium. Through these definitions and calculations, a directed weighted graph is constructed that reflects the propagation relationships of electromagnetic radiation.

[0084] A graph embedding algorithm is used to map nodes to a low-dimensional vector space and calculate the similarity matrix between nodes. There are many graph embedding algorithms, such as DeepWalk, LINE, etc. Taking the DeepWalk algorithm as an example, it generates a series of node sequences by performing random walks on the graph, and then regards these node sequences as sentences in the text, and uses word vector training methods in natural language processing, such as the Skip-gram model, to map the nodes to a low-dimensional vector space. During the mapping process, parameters such as the length of the random walk, the number of steps, and the dimension of the embedding vector need to be set. After obtaining the low-dimensional vector representation of the node, the similarity matrix between the nodes can be calculated. Similarity is generally measured using cosine similarity or Euclidean distance. For example, the cosine similarity formula is where v i and v j are the low-dimensional vector representations of the two nodes, · represents the dot product of the vectors, and ‖·‖ represents the norm of the vectors.

[0085] A spectral clustering algorithm is used to delineate high-probability interference source areas and generate an electromagnetic radiation heat map. The spectral clustering algorithm performs clustering based on the eigenvalues and eigenvectors of the graph's Laplacian matrix. During the clustering process, the number of cluster categories must be determined. This can be determined based on specific evaluation metrics such as the silhouette coefficient and the Calinski-Harabasz index. Based on the clustering results, high-probability interference source areas are identified and displayed as a heat map on a map or a plan view of the detection area, visually demonstrating the distribution of interference sources.

[0086] Embodiment 4:

[0087] When the dynamic parameter adjustment module establishes an adaptive optimization strategy and generates an environment adaptive detection instruction, the specific implementation method is as follows:

[0088] When establishing an adaptive optimization strategy, the state space is first defined as a combination vector of signal-to-noise ratio, spectrum flatness, and ambient temperature and humidity change rate. The signal-to-noise ratio (SNR) is an indicator that measures the ratio of signal strength to noise strength. Its calculation formula is: where s iis the sampling value of the signal, n i is the noise sampling value. Spectral flatness measures the uniformity of the spectrum and is typically expressed by calculating spectral entropy or other related statistics. The rate of change of ambient temperature and humidity is determined by continuously measuring the ambient temperature and humidity and calculating the ratio of the temperature and humidity difference between adjacent measurements to the time interval. These three parameters are combined into a vector to represent the state space of the system.

[0089] A proximal policy optimization algorithm framework is constructed, using advantage function estimation and importance sampling to optimize the stability of the policy network. The proximal policy optimization algorithm (PPO) is a reinforcement learning algorithm based on policy gradients. In the PPO algorithm, advantage function estimation is used to measure the gap between the current policy and the optimal policy. The advantage function is generally obtained by calculating the difference between the value function under the current policy and the reference value function, that is, A(s,a) = V(s) - V ref (s), where A(s,a) is the advantage function, V(s) is the value function of state s under the current strategy, and V ref (s) is the reference value function. Importance sampling is used to leverage previously sampled data when updating the policy, improving data utilization. Through these methods, the stability of the policy network is optimized, enabling it to more effectively learn the optimal policy.

[0090] Then, the reward function is designed as the weighted difference between the interference suppression gain and the detection energy consumption, driving the policy network to generate the optimal parameter adjustment instructions. The interference suppression gain can be measured by calculating the improvement in the signal-to-noise ratio before and after interference suppression, that is, Detection energy consumption is related to factors such as the operating time and power of the detection device. For example, the energy consumption formula can be expressed as E = P × t, where E is energy consumption, P is device power, and t is operating time. The reward function R = w1 × Gain - w2 × E, where w1 and w2 are weighting coefficients. Their values are adjusted based on actual needs to balance interference suppression and energy consumption. The policy network continuously adjusts detection sensitivity and anti-interference filter parameters based on feedback from the reward function.

[0091] When generating environment-adaptive detection instructions, a multi-constraint optimization model is established with detection accuracy, real-time performance, and equipment life as boundary conditions. Detection accuracy can be measured by indicators such as false alarm rate and missed alarm rate. For example, the false alarm rate formula is The formula for the false negative rate is: Real-time performance is related to the time interval between data collection and processing, as well as the delay in outputting test results. Equipment lifespan is related to factors such as age, operating temperature, and voltage, and can be assessed through the development of a lifespan prediction model. Using these indicators as boundary conditions, a multi-constraint optimization model is constructed.

[0092] The interior point method is used to solve a convex optimization problem with nonlinear constraints, generating a set of feasible solutions for the filter cutoff frequency and gain. The interior point method is an effective approach for solving optimization problems. It gradually approaches the optimal solution by finding a point within the feasible region. During the solution process, the optimization problem must be transformed into a standard convex optimization form, and an appropriate interior point algorithm, such as the primal-dual interior point method, is selected based on the characteristics of the problem. Through iterative calculations using the interior point method, a set of feasible solutions for the filter cutoff frequency and gain is obtained. Finally, the entropy weighting method is used to evaluate the multi-objective priority of the solution set and output the optimal parameter configuration. The entropy weighting method is an objective weighting method that determines weights based on the degree of data dispersion. Each solution in the feasible solution set is comprehensively evaluated based on the entropy weights. The solution with the highest overall score is selected as the optimal parameter configuration, and the environment adaptive detection instructions are output.

[0093] Example 5:

[0094] This embodiment primarily relates to the specific implementation of the electromagnetic field simulation module in the system. The electromagnetic field simulation module is used to establish an electromagnetic wave propagation model based on Maxwell's equations, simulating the combined effects of multipath and obstacle diffraction in real time. Maxwell's equations are a set of equations that describe the fundamental laws of the electromagnetic field, including Gauss's law for the electric field, Gauss's law for the magnetic field, Faraday's law of electromagnetic induction, and Ampere's circuit law. In the actual modeling process, numerical calculation methods are used based on these equations and combined with specific boundary and initial conditions to solve the propagation characteristics of electromagnetic waves. Common numerical calculation methods include the finite element method and the finite-difference time-domain method. For example, the finite-difference time-domain method discretizes space and time, converting Maxwell's equations into difference equations for solution. When discretizing space, the calculation area is divided into many small grid cells, and when discretizing time, an appropriate time step is set. Through continuous iterative calculations, the propagation of electromagnetic waves in space is simulated.

[0095] When simulating the multipath effect, we consider the reflection and refraction of electromagnetic waves in different media during propagation. For example, in an urban environment, electromagnetic waves will reflect on the surface of buildings, forming multiple propagation paths. To simulate this situation, we need to calculate the propagation delay and attenuation of each reflection path based on the location, shape, material, and other information of the building. Assuming there is a reflection path with a propagation distance of d1 and a propagation speed of v, the propagation delay is The attenuation is related to factors such as the propagation distance, the electrical conductivity and magnetic permeability of the propagation medium. When encountering reflection from an ideal conductor after propagation in free space, the reflection coefficient can be calculated based on the Fresnel reflection formula to obtain the attenuation of the reflected wave.

[0096] The simulation of obstacle diffraction is based on the Geometric Theory of Diffraction (GTD) or the Uniform Geometric Theory of Diffraction (UTD). Taking GTD as an example, when electromagnetic waves encounter obstacles such as edges and corners, diffraction occurs. The intensity and propagation direction of the diffracted waves are determined by calculating the diffraction coefficient. The diffraction coefficient is related to parameters such as the shape of the obstacle, the frequency of the electromagnetic wave, and the angle of incidence. For example, for an ideally conductive wedge-shaped obstacle, its diffraction coefficient can be calculated using a specific formula. During the calculation process, it is necessary to accurately obtain information such as the angle of the wedge and the polarization mode of the electromagnetic wave.

[0097] Measured data is input into the simulation model to generate a virtual field strength distribution map, and differential analysis triggers an abnormal radiation warning signal. In practice, the measured data, such as electromagnetic field strength and spectrum distribution, collected by the multi-source data acquisition module are preprocessed according to the input requirements of the simulation model. For example, the measured electromagnetic field strength data is interpolated according to the spatial grid division of the model to ensure that it matches the simulation model's computational grid.

[0098] The pre-processed measured data is input into the established electromagnetic wave propagation model, and the model is run to obtain a virtual field strength distribution map. This distribution map reflects the theoretical electromagnetic field strength distribution under the current environmental conditions and radiation source. Then, the virtual field strength distribution map is compared and analyzed with the measured field strength data. The virtual field strength value E at the corresponding position can be calculated by point-by-point comparison. v (x,y,z) and the measured field strength value E m The difference between (x, y, z) ΔE(x, y, z) = |E v (x,y,z)-E m (x,y,z)|.

[0099] A reasonable threshold T is set. When the number of locations where ΔE(x, y, z) exceeds the threshold T exceeds a certain percentage, or when ΔE(x, y, z) remains above the threshold T within a critical area, an abnormal radiation warning signal is triggered. The setting of threshold T requires comprehensive consideration of factors such as the system's detection accuracy requirements, ambient noise levels, and historical data. For example, through statistical analysis of a large amount of historical data, the distribution range of the difference between virtual and measured field strengths under normal electromagnetic environments can be determined. Based on this, a reasonable threshold T can be set to ensure timely detection of actual abnormal radiation conditions while avoiding excessive false alarms.

[0100] In this way, the electromagnetic field simulation module can assist the system in more accurately determining whether there are any abnormalities in the electromagnetic environment, providing strong support for ensuring the safety and stability of the electromagnetic environment.

[0101] Example 6:

[0102] This example focuses on the specific implementation of the system's functions based on autoencoders and the dynamic mode decomposition algorithm. First, an autoencoder-based abnormal pattern library is constructed to perform unsupervised feature learning on historical electromagnetic interference data. An autoencoder is a neural network structure consisting of an encoder and a decoder. The encoder maps input data into a low-dimensional feature space, while the decoder reconstructs the low-dimensional features into an approximation of the original data.

[0103] Historical electromagnetic interference data is formatted and fed into an autoencoder. Assuming the input data is x and the encoder's mapping function is f, the encoded low-dimensional feature z = f(x). When training the autoencoder, the goal is to minimize the reconstruction error, i.e., minimize ‖xg(z)‖, where g is the decoder's mapping function. By continuously adjusting the encoder and decoder network parameters, the autoencoder is able to learn the key features of the historical electromagnetic interference data.

[0104] During the training process, choosing the right neural network architecture and training algorithm is crucial. For example, a multilayer perceptron (MLP) can be used as the underlying structure of an autoencoder, with parameters updated using stochastic gradient descent (SGD) or its variants such as Adagrad and Adadelta. Furthermore, the number of hidden layers and nodes should be appropriately set based on the data size and feature complexity. Generally speaking, increasing the number of hidden layers and nodes improves the model's expressiveness, but can also lead to overfitting, necessitating fine-tuning through methods such as cross-validation.

[0105] After sufficient training, the low-dimensional features learned by the autoencoder form the basis of the anomaly pattern library. These low-dimensional features represent typical patterns in historical electromagnetic interference data and can be used to compare and analyze real-time signals to determine whether anomalies exist.

[0106] The dynamic mode decomposition algorithm is used to extract transient modes from real-time signals and generate potential interference type identification results. The dynamic mode decomposition (DMD) algorithm is a data-driven analysis method that can extract dynamic modes from time series data. Real-time electromagnetic signals are collected and arranged in chronological order into a matrix form, X.

[0107] The DMD algorithm first performs singular value decomposition (SVD) on the matrix X to obtain X = UΣV * , where U and V are unitary matrices, and Σ is a diagonal matrix with singular values. Then, through a series of calculations, the DMD modes and DMD eigenvalues are obtained. The DMD modes reflect the different dynamic modes of the signal, while the DMD eigenvalues are related to the time evolution characteristics of these modes.

[0108] Based on DMD modes and eigenvalues, the transient behavior of real-time signals can be analyzed. For example, by observing the energy distribution and time evolution trends of different modes, abnormal dynamic patterns can be determined. If the energy of certain modes changes abnormally within a short period of time, or if dynamic characteristics similar to abnormal patterns in historical data are observed, potential interference can be identified.

[0109] Compare the results of dynamic modal decomposition with the abnormal pattern library based on the autoencoder. If the transient pattern of the real-time signal matches some patterns in the abnormal pattern library, the potential interference type can be preliminarily determined. The matching degree can be calculated using methods such as Euclidean distance and cosine similarity. For example, calculate the low-dimensional feature vector z of the real-time signal real and the low-dimensional feature vector z of a pattern in the abnormal pattern library abnormal Cosine similarity of When the similarity exceeds a certain threshold, the two are considered to match.

[0110] In this way, the system can use historical data and real-time signal analysis to more effectively identify potential interference types, provide richer information for electromagnetic environment detection and interference source location, and improve the detection capability and reliability of the entire system.

[0111] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0112] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multimodal electromagnetic environment detection system, characterized in that: The system comprises: Multi-source data acquisition module, used to synchronously collect electromagnetic field intensity, spectrum distribution, polarization direction and ambient temperature and humidity data through a multi-modal sensor array to generate a multi-modal spatiotemporal dataset; a signal preprocessing module, configured to perform nonlinear signal processing on the multimodal spatiotemporal data set to generate a time-frequency signal sequence; A feature extraction module is used to build a deep feature extraction model based on a convolutional neural network, perform multi-scale spatiotemporal feature fusion on the time-frequency signal sequence, extract the time-frequency sparsity and spatial correlation characteristics of electromagnetic interference, and generate a multimodal feature coding matrix; An interference source identification module is used to design a topological analysis model based on a graph neural network, construct a node relationship diagram of the electromagnetic radiation propagation path according to the multimodal feature coding matrix, and generate an interference source location probability distribution map; The dynamic parameter adjustment module is used to establish an adaptive optimization strategy based on reinforcement learning, dynamically adjust the detection sensitivity and anti-interference filter parameters according to the interference source positioning probability distribution map, and generate environmental adaptive detection instructions.

2. The multimodal electromagnetic environment detection system according to claim 1, wherein: The nonlinear signal processing uses a wavelet packet decomposition and adaptive filtering combined algorithm to separate broadband noise components and steady-state electromagnetic characteristics, including: The electromagnetic field intensity signal is decomposed by wavelet packets and a sub-band energy distribution sequence is generated through multi-resolution analysis; Construct an adaptive filter coefficient update equation, use the high-frequency sub-band component as the reference noise input, and iteratively optimize the filter transfer function; The kurtosis detection algorithm is used to identify abnormal frequency points of steady-state spectrum components and segment normal electromagnetic environment signal segments.

3. The multimodal electromagnetic environment detection system according to claim 1, wherein: The construction of the deep feature extraction model includes: In the feature encoding layer, dilated convolution kernels are introduced to capture long-range temporal dependencies and output multi-scale feature map tensors; In the feature fusion layer, the channel attention mechanism is used to weightedly aggregate spatiotemporal features of different scales, and residual connections are superimposed to prevent gradient disappearance; The parameters of the feature encoding layer and the fusion layer are optimized through a contrastive learning strategy to minimize the ratio of the intra-class feature distance to the inter-class feature distance.

4. The multimodal electromagnetic environment detection system according to claim 1, wherein: The design of the topology analysis model includes: Define the electromagnetic radiation source as a graph node, the propagation path attenuation coefficient as the edge weight, and construct a directed weighted graph structure; A graph embedding algorithm is used to map nodes into a low-dimensional vector space and calculate the similarity matrix between nodes; The spectral clustering algorithm is used to divide the high-probability interference source areas and generate electromagnetic radiation heat maps.

5. The multimodal electromagnetic environment detection system according to claim 1, wherein: The establishment of the adaptive optimization strategy includes: The state space is defined as a combination vector of signal-to-noise ratio, spectrum flatness, and ambient temperature and humidity change rate; Construct a proximal policy optimization algorithm framework, using advantage function estimation and importance sampling to optimize the stability of the policy network; The reward function is designed as the weighted difference between interference suppression gain and detection energy consumption, driving the policy network to generate optimal parameter adjustment instructions.

6. The multimodal electromagnetic environment detection system according to claim 1, wherein: The system further comprises: The electromagnetic field simulation module is used to establish an electromagnetic wave propagation model based on Maxwell's equations and simulate the superposition effect of multipath effect and obstacle diffraction in real time; The measured data is input into the simulation model to generate a virtual field strength distribution map, and the abnormal radiation warning signal is triggered through difference analysis.

7. The multimodal electromagnetic environment detection system according to claim 1, wherein: The generation of the multimodal spatiotemporal dataset includes: Synchronously collect antenna array received signals, spectrum analyzer scan data, and temperature and humidity sensor readings; The non-negative matrix factorization algorithm is used to reduce the feature dimension of heterogeneous data and construct a low-rank approximate data cube; The sensor quantization error is eliminated by wavelet threshold denoising method to enhance the detectability of weak signals.

8. The multimodal electromagnetic environment detection system according to claim 1, wherein: The generation of the environment adaptive detection instruction includes: Establish a multi-constraint optimization model with detection accuracy, real-time performance, and equipment life as boundary conditions; The interior point method is used to solve the convex optimization problem with nonlinear constraints and generate feasible solutions of filter cutoff frequency and gain. The multi-objective priority of the solution set is evaluated by the entropy weight method, and the optimal parameter configuration scheme is output.

9. The multimodal electromagnetic environment detection system according to claim 1, wherein: The system further comprises: An abnormal pattern library based on an autoencoder is constructed to perform unsupervised feature learning on historical electromagnetic interference data; a dynamic mode decomposition algorithm is used to extract transient patterns of real-time signals and generate potential interference type identification results.

10. A multimodal electromagnetic environment detection method, characterized in that: The method comprises the following steps: Step 1: Use a multimodal sensor array to synchronously collect electromagnetic field intensity, spectrum distribution, polarization direction, and ambient temperature and humidity data to generate a multimodal spatiotemporal dataset; Step 2: performing nonlinear signal processing on the multimodal spatiotemporal data set to generate a time-frequency signal sequence; Step 3: Construct a deep feature extraction model based on a convolutional neural network, perform multi-scale spatiotemporal feature fusion on the time-frequency signal sequence, extract the time-frequency sparsity and spatial correlation characteristics of the electromagnetic interference, and generate a multimodal feature coding matrix; Step 4: Based on the topological analysis model of the graph neural network, a node relationship diagram of the electromagnetic radiation propagation path is constructed according to the multimodal feature coding matrix, and a probability distribution diagram of the interference source location is generated; Step 5: Using the adaptive optimization strategy of reinforcement learning, dynamically adjust the detection sensitivity and anti-interference filter parameters according to the interference source positioning probability distribution map to generate an environment adaptive detection instruction.

Images