High-dynamic self-adaptive environment interference compensation and calibration system

Abstract

The invention discloses a high-dynamic adaptive environmental interference compensation and calibration system, which relates to the technical field of environmental interference compensation, and comprises a high-dynamic adaptive deep learning model subsystem and a preprocessing module for generating a fusion feature vector in original radar data; a front-end feature extraction module performs feature extraction on the fusion feature vector through a multi-layer convolutional network and a multi-head attention Transform module; the self-adaptive physical constraint module corrects the extracted features based on physical residual errors through a full-connection network to generate corrected features; the adaptive optimization subsystem is embedded into the high-dynamic adaptive deep learning model subsystem and comprises an IRNN adaptive compensation branch, a multi-view data modeling module, a fusion module and a separation module; and the online environmental parameter calibration subsystem is used for dynamically calibrating system parameters based on the environmental parameters. According to the invention, environment interference can be captured and compensated in real time in a complex environment, and weak life signals of life bodies can be accurately detected.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental interference compensation, and more specifically, to a high-dynamic adaptive environmental interference compensation and calibration system. Background Art

[0002] Currently, millimeter-wave radar technology has been widely used in the field of micro-motion signal detection. In particular, through fixed filtering means such as Kalman filtering and wavelet transform, and statistical modeling methods such as Gaussian mixture models, the basic detection ability of weak signals of living bodies has been achieved; models such as CNN and LSTM based on deep learning have also made preliminary progress in noise feature extraction. However, in the actual detection environment, the reflection of static obstacles, dynamic interference fluctuations, multipath effects and electromagnetic noise are highly coupled, resulting in significant limitations for existing methods: fixed filters are difficult to match the time-varying characteristics of interference signals due to the fixed time-domain response, statistical modeling relies on a large amount of prior data and cannot cope with environmental mutations, and deep learning models are prone to generating false features due to the lack of physical constraints; at the same time, traditional algorithms lack effective mechanisms in aspects such as multipath signal separation and environmental parameter drift compensation, resulting in a decrease in the signal-to-noise ratio of the target signal and making it difficult to effectively distinguish real target signals from interference. Therefore, how to improve the target detection accuracy of millimeter-wave radar signals in complex scenarios is an urgent problem for those skilled in the art to solve. Summary of the Invention

[0003] In view of this, the present invention provides a high-dynamic adaptive environmental interference compensation and calibration system, which overcomes the above defects.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A high-dynamic adaptive environmental interference compensation and calibration system, comprising:

[0006] A high-dynamic adaptive deep learning model subsystem, including a preprocessing module, a front-end feature extraction module, and an adaptive physical constraint module; the preprocessing module generates a fused feature vector in the original radar data through amplitude extraction, frequency-domain conversion, attenuation coefficient calculation, and static reflection feature extraction; the front-end feature extraction module extracts features from the fused feature vector through a multi-layer convolutional network and a multi-head attention Transformer module; the adaptive physical constraint module predicts the scattering factor, propagation coefficient, and phase shift based on the extracted features through a fully connected network, calculates the physical residual with the corresponding theoretical values, and corrects the extracted features based on the physical residual to generate corrected features;

[0007] The adaptive optimization subsystem embedded in the high-dynamic adaptive deep learning model subsystem includes an IRNN adaptive compensation branch, a multi-view data modeling module, a fusion module, and a separation module; the IRNN adaptive compensation branch dynamically updates the weight coefficients based on the influence of environmental noise and predicts the target signal based on the corrected features; the multi-view data modeling module performs convolutional processing on each view signal and embeds physical constraints; the fusion module uses multi-head attention to fuse multi-view features to generate a reconstructed signal; the separation module separates the target signal and the interference signal in the reconstructed signal through a fully connected network.

[0008] The online environmental parameter calibration subsystem is used to estimate environmental parameters through multi-source fusion data and dynamically calibrate system parameters based on the environmental parameters.

[0009] Optionally, the preprocessing module includes:

[0010] An amplitude extraction unit for extracting the maximum value of each channel sequence as an amplitude feature;

[0011] A frequency domain conversion unit for extracting frequency domain features based on the discrete Fourier transform;

[0012] An attenuation coefficient calculation unit for obtaining the signal attenuation coefficient by using the exponential decay fitting method;

[0013] A static reflection feature extraction unit for extracting static reflection features by using a one-dimensional convolutional layer with fixed weights;

[0014] A feature fusion unit for splicing and fusing the amplitude feature, the frequency domain feature, the signal attenuation coefficient, and the static reflection feature to generate the fusion feature vector.

[0015] Optionally, the multi-layer convolutional network includes a plurality of stacked convolutional layers, skip connections are adopted between the convolutional layers, and batch normalization and ReLU activation functions are sequentially connected after each convolutional layer.

[0016] Optionally, the multi-head attention Transformer module includes:

[0017] A linear layer for mapping the input features to a standard dimension and obtaining long-range temporal dependencies by using the multi-head attention mechanism;

[0018] A feed-forward network including a first-layer network composed of a first linear layer and a ReLU activation function and a second-layer network composed of a second linear layer and layer normalization for obtaining local spatial feature data and generating the extracted features.

[0019] Optionally, the adaptive physical constraint module includes:

[0020] A fully connected layer, configured to predict a scattering factor, a propagation coefficient, and a phase shift based on the extracted features;

[0021] A physical residual calculation module, configured to calculate theoretical values of a scattering factor, a propagation coefficient, and a phase shift based on a predefined physical model, and calculate the physical residual with respect to corresponding predicted values of the scattering factor, the propagation coefficient, and the phase shift;

[0022] A feature correction module, configured to correct the extracted features output by the front-end feature extraction module according to a preset physical loss coefficient and the physical residual to generate corrected features.

[0023] Optionally, the RNN adaptive compensation branch includes:

[0024] A time series feature extraction unit, configured to obtain the fused feature vector and generate time series data based on the fused feature vector;

[0025] A comprehensive feature generation unit, configured to map the fused feature vector, and splice the mapped fused feature vector and the time series data in the same dimension to generate a comprehensive feature;

[0026] A dynamic weight adjustment subunit, configured to calculate an environmental noise evaluation value in real time according to the physical residual, and dynamically update a weight coefficient based on the environmental noise evaluation value;

[0027] A signal prediction output unit, configured to map the comprehensive feature to a predicted target signal by using a fully connected layer.

[0028] Optionally, the online environmental parameter calibration subsystem includes:

[0029] A multi-source environmental data fusion module, configured to collect radar data, temperature and humidity data, and electromagnetic field data and perform fusion to generate multi-source fusion data;

[0030] An environmental parameter estimation module, configured to map the multi-source fusion data to obtain an environmental parameter estimation vector;

[0031] A Bayesian optimization calibration module, configured to perform iterative optimization of a detection threshold based on Bayes' theorem to obtain an optimal solution of the detection threshold;

[0032] A self-supervised contrastive learning module, configured to dynamically adjust a weight coefficient according to the optimal solution of the threshold and historical data, and calculate a cosine similarity of a feature space to perform discriminative feature learning in the case of no labels;

[0033] A closed-loop feedback execution module, configured to update the detection threshold through a parameter mapping network according to the current environmental parameter estimation vector and the optimal solution of the detection threshold.

[0034] Optionally, the adaptive optimization subsystem further includes a total loss function optimization module, and the expression of the optimized total loss function is:

[0035] L_total = L_data * w_dynamic + λ_physical * L_physical + L_separation;

[0036] In the formula, L_data is the data fitting loss; w_dynamic is the adaptive weight adjustment factor; λ_physical is the physical loss coefficient; L_physical is the physical loss; L_separation is the multi-view separation loss.

[0037] As can be seen from the above technical solutions, the present invention provides a high-dynamic adaptive environmental interference compensation and calibration system. Compared with the prior art, it has the following beneficial effects:

[0038] 1. By combining MIMO millimeter-wave radar, physical modeling, and deep learning technologies, it can capture and compensate environmental interference in real time in a complex environment, accurately detect weak life signals of living beings, such as physiological activities such as breathing and heartbeat of insects and cold-blooded animals, etc.; improve the accuracy of micro-motion signal detection;

[0039] 2. Using MIMO millimeter-wave radar to collect target data, and through an adaptive interference compensation and calibration system based on a physical neural network, it can effectively distinguish real life signals from environmental noise; improve the accuracy of target detection;

[0040] 3. It can monitor environmental parameters in real time, dynamically adjust the detection threshold, reduce the false alarm rate, and ensure accurate identification of weak target signals such as cold-blooded living bodies in a complex security inspection environment, providing strong guarantees for border security and public safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0042] Figure 1 It is a schematic diagram of the system structure provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] An embodiment of the present invention discloses a high-dynamic adaptive environmental interference compensation and calibration system, designs a high-dynamic adaptive deep learning model IRNN, embeds a radar scattering model, a wave propagation model, and an electromagnetic interference model for the collected weak life signal radar spectrum of the MINO radar, adds an adaptive physical constraint network structure, designs an adaptive loss function and uses an adversarial training strategy and a time series feature matching method to implement a target signal reconstruction and noise suppression mechanism; the structure is as Figure 1 shown, including:

[0045] A high-dynamic adaptive deep learning model subsystem, including a preprocessing module, a front-end feature extraction module, and an adaptive physical constraint module; the preprocessing module generates a fused feature vector in the original radar data through amplitude extraction, frequency domain conversion, attenuation coefficient calculation, and static reflection feature extraction; the front-end feature extraction module extracts features from the fused feature vector through a multi-layer convolutional network and a multi-head attention Transformer module; the adaptive physical constraint module predicts the scattering factor, propagation coefficient, and phase shift based on the extracted features through a fully connected network, calculates the physical residual with the corresponding theoretical values, and corrects the extracted features based on the physical residual to generate corrected features;

[0046] An adaptive optimization subsystem embedded in the high-dynamic adaptive deep learning model subsystem, including an IRNN adaptive compensation branch, a multi-view data modeling module, a fusion module, and a separation module; the IRNN adaptive compensation branch dynamically updates the weight coefficients based on the influence of environmental noise and predicts the target signal based on the corrected features; the multi-view data modeling module performs convolutional processing on the signals of each view and embeds physical constraints; the fusion module uses multi-head attention to fuse multi-view features to generate a reconstructed signal; the separation module separates the target signal and the interference signal in the reconstructed signal through a fully connected network;

[0047] An online environmental parameter calibration subsystem for estimating environmental parameters through multi-source fusion data and dynamically calibrating system parameters based on the environmental parameters.

[0048] In one embodiment, the preprocessing module includes:

[0049] An amplitude extraction unit for extracting the maximum value of each channel sequence as the amplitude feature;

[0050] A frequency domain conversion unit, configured to extract frequency domain features based on discrete Fourier transform;

[0051] An attenuation coefficient calculation unit, configured to obtain a signal attenuation coefficient by using an exponential decay fitting method;

[0052] A static reflection feature extraction unit, configured to extract static reflection features by using a one-dimensional convolutional layer with fixed weights;

[0053] A feature fusion unit, configured to splice and fuse amplitude features, frequency domain features, signal attenuation coefficients and static reflection features to generate a fused feature vector.

[0054] Further, the IRNN mainly includes a preprocessing module PreprocessingLayer, a front-end feature extraction module, an adaptive physical constraint module and a total loss function. Its network structure diagram is as Figure 1 shown. The IRNN combines physical heuristic information and deep learning techniques to ensure that it can effectively extract the weak life signal features in MIMO radar data.

[0055] Furthermore, in the PreprocessingLayer structure, there is an amplitude extraction unit which, through the AmplitudeExtraction function, uses torch.max(input, dim=-1) to take the maximum value for each channel sequence, and the output shape is (64, 16); a frequency domain conversion unit which, when performing FrequencyExtraction, calls torch.fft.fft(input) to convert the time domain signal into frequency domain data and extracts the main frequency components as frequency domain features; a decay coefficient calculation unit which, when performing DecayCoefficientExtraction, uses scipy.optimize.curve_fit to perform exponential decay fitting on each signal to obtain the decay coefficient; a static reflection feature extraction unit which uses a fixed weight filter to extract the static reflection component, including nn.Conv1d with a kernel size of 5 (kernel_size = 5), a stride of 1 (stride = 1), a padding of 2 (padding = 2), and the bias set to False, and presets an adjustable constant weight based on self.static_filter.weight.fill_(0.2). The amplitude, frequency domain, decay coefficient, and static reflection features are all subjected to global average pooling and concatenated in the channel dimension; the final output feature vector has a dimension of 4×16. In this embodiment, torch.cat([amplitude, frequency, decay_coefficient, static_filter], dim=-1) is used to fuse the above features to form a feature vector with a dimension of approximately 64 - 128 as the input to the subsequent network module.

[0056] In one embodiment, the multi-layer convolutional network includes multiple stacked convolutional layers, with skip connections between the respective convolutional layers, and batch normalization and ReLU activation functions are sequentially connected after each convolutional layer.

[0057] In one embodiment, the multi-head attention Transformer module includes:

[0058] A linear layer for mapping the input features to 64 dimensions and using the multi-head attention mechanism to obtain long-range temporal dependencies, where the number of attention heads is 8;

[0059] A feed-forward network including a first layer network composed of a first linear layer and a ReLU activation function, and a second layer network composed of a second linear layer and layer normalization, for obtaining local spatial feature data and generating extracted features.

[0060] Furthermore, the front-end feature extraction module adopts a multi-layer convolutional network and a multi-head attention Transformer module.

[0061] Among them, the multi-layer convolutional network includes a first convolutional layer with a kernel size of 3 (kernel_size = 3), a stride of 1 (stride = 1), a padding of 1 (padding = 1), and the bias is set to True. Subsequently, nn.BatchNorm1d(64) and nn.ReLU() are used for activation; a second convolutional layer with a kernel size of 3 (kernel_size = 3), a stride of 1 (stride = 1), a padding of 1 (padding = 1), and the bias is set to True. Subsequently, nn.BatchNorm1d(128) and nn.ReLU() are used; a third convolutional layer with a kernel size of 3 (kernel_size = 3), a stride of 1 (stride = 1), a padding of 1 (padding = 1), and the bias is set to True. Subsequently, nn.BatchNorm1d(256) and nn.ReLU() are used; a skip connection (Residual Connection) is adopted between each convolutional block. For example, the output of the first convolutional layer is directly added to the input of the second convolutional layer to ensure the smooth transmission of gradients and alleviate the gradient disappearance problem of deep networks.

[0062] The multi-head attention Transformer module maps the feature of each time step of the output data of the multi-layer convolutional network from 256 dimensions to 512 dimensions through a linear layer; nn.MultiheadAttention(embed_dim = 512, num_heads = 8) is used, where the dimension of each head is 64 and the total embedding dimension is 512; the first layer of the feed-forward network (FFN) is nn.Linear(512, 2048) with the activation function ReLU; the second layer is nn.Linear(2048, 512); nn.LayerNorm(512) is used to normalize the data to ensure training stability; the shape of the transformed output feature is (Sequence, Batch, 512), and then it is converted back to (Batch, Sequence, 512) for subsequent processing; among them, Sequence represents the sequence length; Batch represents the batch size.

[0063] In one embodiment, the adaptive physical constraint module includes:

[0064] A fully connected layer for predicting the scattering factor, propagation coefficient, and phase shift according to the extracted features;

[0065] The physical residual calculation module calculates the theoretical values of the scattering factor, propagation coefficient, and phase shift based on a predefined physical model, and calculates the physical residuals by comparing them with the corresponding predicted values of the scattering factor, propagation coefficient, and phase shift.

[0066] The feature correction module is used to correct the extracted features output by the front-end feature extraction module according to a preset physical loss coefficient and physical residuals, and generate corrected features.

[0067] Furthermore, the fully connected layer of the adaptive physical constraint module includes the fully connected layer FC1: with an input dimension of 512 and an output dimension of 256, using nn.Linear(512, 256), and the activation function is ReLU. The radar scattering, propagation coefficient, and phase shift are predicted through the fully connected layer FC1, and at the same time, the theoretical expected values are calculated using a predefined physical model; the fully connected layer FC2: with an input dimension of 256 and an output dimension of 128, using nn.Linear(256, 128), outputs the predicted scattering factor, propagation coefficient, and phase shift; define the physical formula that the interference quantity conforms to according to the radar scattering model, wave propagation model, and electromagnetic propagation model, and embed the physical constraints. Calculate the theoretical expected value P_theoretical using an external function, the FC2 outputs the network predicted value P_predicted, and calculate the physical residual, where the physical residual = MSE(P_theoretical, P_predicted); the preset physical loss coefficient λ_physical = 0.1, so the physical loss L_physical = 0.1 × MSE(P_theoretical, P_predicted); the corrected feature vector output by the adaptive physical constraint module is fused with the output features of the front-end feature extraction module to form the final feature representation for reconstructing the target signal.

[0068] In this embodiment, the high-dynamic adaptive deep learning model IRNN enables the network to follow physical laws while fitting data, thereby improving the model's stability and interpretability. IRNN adopts a multi-modal physical constraint framework and combines radar signal features to construct a set of dynamic physical modeling systems, including: a radar scattering model based on Mie Scattering and Rayleigh Scattering, which dynamically calculates the reflection, transmission, and attenuation characteristics of radar signals in different media, enabling the neural network to adaptively adjust weights; using the electromagnetic wave propagation equation and calculating the changes of radar signals in complex environments (such as multipath propagation and attenuation media) with Maxwell's equations, enabling the model to dynamically adjust the signal processing method to adapt to different target types and interference sources; based on the random field modeling of electromagnetic noise, using the Gaussian-Markov random process (GMRF) to model the noise distribution and embedding this model into the loss function of IRNN, enabling the network to automatically optimize the noise reduction strategy during training to adapt to different interference environments. The data used by the system is the original echo data collected in real time by the MIMO radar, and the data acquisition length of each signal channel is set to 1024 points.

[0069] In one embodiment, the total loss function includes the data fitting loss L_data: calculating the error between the network output signal and the true target signal using the mean square error (MSE); the physical loss L_physical: multiplying the above-calculated physical residual by λ__physical = 0.1; the regularization loss L_reg: preventing overfitting through L2 regularization and Dropout; and the final total loss L_total = L_data + L_physical + L_reg.

[0070] In one embodiment, the IRNN adaptive compensation branch includes:

[0071] A timing feature extraction unit for obtaining a fused feature vector and generating timing data based on the fused feature vector;

[0072] A comprehensive feature generation unit for mapping the fused feature vector, concatenating the mapped fused feature vector and the timing data in the same dimension to generate a comprehensive feature;

[0073] A dynamic weight adjustment subunit for calculating the environmental noise evaluation value in real time according to the physical residual and dynamically updating the weight coefficient based on the environmental noise evaluation value;

[0074] A signal prediction output unit for mapping the comprehensive feature to a predicted target signal using a fully connected layer.

[0075] In one embodiment, an adaptive optimization subsystem is designed to improve the network recognition accuracy by embedding IRNN.

[0076] Further, the adaptive optimization subsystem includes an IRNN adaptive compensation branch, a multi-view data modeling module, a fusion module, a signal separation module, and a loss function optimization module.

[0077] Among them, on the IRNN adaptive compensation branch, the IRNN_AdaptiveBranch(features, spatio_temporal_data) function is constructed. The input features are features: the fused feature vector output from the preprocessing module, with a shape of (256, feature_dim); spatio_temporal_data is the temporal information obtained through temporal convolution or a recurrent network, with a shape of (256, 1024, 64). The features are mapped to the same dimension as the temporal features through a fully connected layer and concatenated with the temporal data through element-wise addition and then reduced in dimension to form the comprehensive input features. In the dynamic weight adjustment mechanism, an adaptive weight w_dynamic is defined, with an initial value set to 1.0. The weight is updated with the real-time measured environmental noise calculated from the current residual: w_dynamic = 1.0 + alpha * noise_level, where alpha is the learning rate adjustment factor and noise_level is the output of data fitting and physical constraints. Finally, the comprehensive features are mapped to the target signal prediction using a fully connected layer.

[0078] In the multi-view data modeling module, a sub-module is established separately for each view (front, rear, left, right), and the function name is ViewAngleModel(input, angle_id). The input is the signal corresponding to the view in the original radar data, and angle_id is the view number (0, 1, 2, 3 corresponding to the front, rear, left, and right respectively). The first layer of convolution nn.Conv1d has a kernel size of 5 (kernel_size = 3), a stride of 1 (stride = 1), a padding of 1 (padding = 1), and the bias is set to True. Subsequently, it is connected to nn.BatchNorm1d(64) and the Relu activation; the second layer of convolution nn.Conv1d has a kernel size of 5 (kernel_size = 3), a stride of 1 (stride = 1), a padding of 1 (padding = 1), and the bias is set to True. Subsequently, it is connected to nn.BatchNorm1d(64) and the Relu activation; the average pooling is connected and embedded, and the physical constraint embedding is calculated. For the signal of this view, the theoretical prediction value is calculated through a preset physical formula, and the prediction value is output using a fully connected layer, and then the residual is calculated.

[0079] In the fusion module, the output feature lists from the view_features_list sub-modules of each perspective, each with a shape of (256, 128), are input into the function named FusionAndSeparation(view_features_list). Subsequently, multi-head attention fusion is performed. Using nn.MultiheadAttention(embed_dim = 128, num_heads = 4), the features of multiple perspectives are concatenated into a sequence, and the mean value is taken to achieve signal reconstruction.

[0080] In the signal separation module, a fully connected network is used to achieve signal separation. The fused features are decomposed into two parts: the target signal and the interference signal, ensuring that the reconstruction error of the target signal is controlled within 5%.

[0081] In the loss function optimization module, adaptive weight adjustment is used. Multiply the data fitting loss (L_data): MSE(predicted_signal, ground_truth), the physical loss (L_physical): MSE(P_theoretical, P_predicted) by the preset physical loss coefficient λ_physical = 0.1, and the multi-perspective separation loss (L_separation): for example, the MSE between the target signal and the true signal. At the same time, the interference signal is constrained to be the residual, and the total loss is optimized as L_total = L_data + λ_physical * L_physical + L_separation; design an adaptive weight adjustment factor w_dynamic, which is updated according to the real-time noise level (through an external sensor or the current residual). The update rule is: w_dynamic = 1.0 + 0.1 * noise_level, where noise_level is the calculated value after normalization. Therefore, the optimized total loss is L_total = L_data * w_dynamic + λ_physical * L_physical + L_separation.

[0082] In this embodiment, for adaptive interference compensation and multi-perspective data fusion, the original echo data is collected in real time using a MIMO radar. In the preprocessing stage, interference features such as liquid damping and static reflection are extracted through a built-in physical heuristic module, and the spatio-temporal features are input into the IRNN network. The weights of the data fitting loss and the physical constraint loss are dynamically adjusted in the network to achieve real-time update of the compensation parameters. For data from different perspectives, each sub-module independently models the interference features, and then the output of each module is constrained by the physical model. A fusion strategy is used to separate the target signal and the interference signal, ensuring that the reconstruction error of the target signal is controlled within 5%, thereby achieving high-precision detection.

[0083] In one embodiment, the online environmental parameter calibration subsystem includes:

[0084] A multi-source environmental data fusion module, which is used to collect radar data, temperature and humidity data, and electromagnetic field data, and perform fusion to generate multi-source fusion data;

[0085] An environmental parameter estimation module, which is used to map the multi-source fusion data to obtain an environmental parameter estimation vector;

[0086] A Bayesian optimization calibration module, which is used to iteratively optimize the detection threshold based on Bayes' theorem to obtain the optimal solution of the detection threshold;

[0087] A self-supervised contrastive learning module, which is used to dynamically adjust the weight coefficient according to the optimal threshold solution and historical data, calculate the cosine similarity of the feature space, and perform discriminative feature learning without labels;

[0088] A closed-loop feedback execution module, which is used to update the detection threshold through a parameter mapping network according to the current environmental parameter estimation vector and the optimal solution of the detection threshold.

[0089] Furthermore, the online environmental parameter calibration subsystem forms a closed-loop feedback with the main IRNN network through a multi-source data fusion algorithm and a Bayesian optimization calibration module to ensure the long-term stable operation of the system and maintain high detection accuracy.

[0090] Among them, in the multi-source environmental data fusion module, the temperature and humidity data are collected by the SHT31 device, the electromagnetic field data are collected by the HMC5883L, and after tensor conversion of the data, multi-source data fusion is performed. The function fuse_environment_data(radar_data, temp_humidity, em_field) is designed, where radar_data is the MIMO radar data, temp_humidity is the temperature and humidity data, and em_field is the electromagnetic field data. The temperature and humidity and electromagnetic field data are encoded to 128 dimensions through a two-layer fully connected network, and the features extracted from the radar data after preprocessing are spliced with the environmental features. Finally, the fused environmental feature vector fused_feature, that is, the multi-source fusion data, is output.

[0091] In the environmental parameter estimation module, the real-time environmental parameter calculation compute_environment_parameters(fused_feature) is designed, and the fused features are mapped to the environmental parameter space through a set of fully connected layers to output the environmental parameter estimation vector E (including current temperature and humidity, electromagnetic field, and other environmental characteristic information).

[0092] Specifically, in the Bayesian optimization calibration module, the function bayesian_optimize(calibration_loss_fn, param_range, iterations = 20) is designed, where calibration_loss_fn is a loss function that accepts candidate parameters and returns the calibration error, param_range is the calibration parameter search range, the detection threshold range is [0.1, 1.0], iterations is the number of iterations, with a default of 20 times, and the function outputs the best calibration parameter best_param.

[0093] In the self-supervised contrastive learning module, the function contrastive_loss(features_anchor, features_positive, temperature = 0.07) is designed. The input features_anchr are the main sample features, features_positive are the corresponding positive sample features in the same environment, and temperature is the temperature parameter with a default of 0.07. The initial contrast loss weight is set to 0.05 and dynamically adjusted to 0.1 based on feedback to calculate the cosine similarity to help the network learn discriminative feature representations in the absence of labels.

[0094] In the closed-loop feedback execution module, the function update_calibration(E, current_threshold) is designed, where E is the environmental parameter vector and current_threshold is the current detection threshold. A fully connected network is used for parameter mapping, and the update period is set to trigger once every 5 seconds. This function is used in combination with Bayesian optimization to form a closed-loop feedback mechanism that automatically adjusts the detection threshold and filtering strategy, and finally outputs the updated detection threshold new_threshold.

[0095] Furthermore, the combined action of the subsystems and the training process are as follows: (1) Data collection: Periodically collect MIMO radar data and auxiliary sensor data (temperature and humidity, electromagnetic field) to form multi-source data input; (2) Data fusion: Call fuse_environment_data() to fuse radar and sensor data into high-dimensional features; (3) Environmental parameter estimation: Use compute_environment_parameters() to obtain environmental parameter E; (4) Bayesian optimization calibration: Call Bayesian_optimize() every 5 seconds to optimize the detection threshold, and update the best parameters in combination with the calibration loss function (for example: the difference between the calibrated target signal and the true signal); (5) Self-supervised contrastive learning: Use contrastive_loss() to optimize the environmental feature representation and improve the robustness of the model to environmental changes; (6) Closed-loop feedback: Update the detection threshold through update_calibration() and feedback the update result to the main PINN network to achieve online calibration; (7) Training strategy and parameter settings: Use torch.utils.data.DataLoader to load the fused training data, set the batch size to 64; The optimizer uses Adam, the initial learning rate is 1×10-3, and the training period is set to 100 epochs; Trigger Bayesian optimization update every 5 seconds, and return the optimal calibration parameters after 20 iterations; The initial loss weight of self-supervised contrastive learning is 0.05, and the dynamic adjustment range is 0.05-0.1; The total loss function consists of data fitting loss, calibration loss, contrast loss and regularization loss. Use an adaptive normalization factor to balance each loss to ensure that the system can still maintain high detection accuracy when environmental factors such as temperature and humidity and electromagnetic interference change.

[0096] In this embodiment, through online environmental parameter estimation and calibration, the system simultaneously collects MIMO radar data and auxiliary sensor data, uses a multi-source data fusion algorithm to calculate the current environmental parameters in real time, and the calibration module uses the Bayesian optimization method and introduces a self-supervised contrastive learning strategy to ensure that the system always maintains a stable and high-precision detection effect when environmental factors such as temperature and humidity and electromagnetic interference change.

[0097] In this embodiment, radar data is obtained through a MIMO radar array. This embodiment adopts an 8-transmitter × 8-receiver design, with a total of 64 signals. The perspective data obtained by each antenna is preprocessed through a spatio-temporal alignment algorithm. The time window is set to 50 milliseconds, and the overlap rate is 50%. The operating frequency is 77 GHz, the bandwidth is 4 GHz, and the sampling rate of each channel is 10 Msps. 310 degrees of effective radar spectra are collected, 4200 groups of weak life signals are collected, and 1000 groups of background signals are used as mixed signals. A total of 50 types of customs biological entry categories of weak life samples are collected, and the target data and interference data collected under different background environmental conditions are stored as csv. 80% of the dataset is randomly selected as the training set, 10% as the validation set, and 10% as the test set for convenient subsequent data training.

[0098] The various embodiments in this specification are described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0099] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A high-dynamic adaptive environmental interference compensation and calibration system, characterized in that, Including: A high-dynamic adaptive deep learning model subsystem, including a preprocessing module, a front-end feature extraction module, and an adaptive physical constraint module; The preprocessing module generates a fused feature vector from the original radar data through amplitude extraction, frequency domain conversion, attenuation coefficient calculation, and static reflection feature extraction; the front-end feature extraction module extracts features from the fused feature vector through a multi-layer convolutional network and a multi-head attention Transformer module; the adaptive physical constraint module predicts the scattering factor, propagation coefficient, and phase shift based on the extracted features through a fully connected network, calculates the physical residual with the corresponding theoretical values, and corrects the extracted features based on the physical residual to generate corrected features; An adaptive optimization subsystem embedded in the high-dynamic adaptive deep learning model subsystem, including an IRNN adaptive compensation branch, a multi-view data modeling module, a fusion module, and a separation module; the IRNN adaptive compensation branch dynamically updates the weight coefficients based on the influence of environmental noise and predicts the target signal based on the corrected features; the multi-view data modeling module performs convolutional processing on the signals of each view and embeds physical constraints; the fusion module uses multi-head attention to fuse multi-view features to generate a reconstructed signal; the separation module separates the target signal and the interference signal in the reconstructed signal through a fully connected network; An online environmental parameter calibration subsystem for estimating environmental parameters through multi-source fused data and dynamically calibrating system parameters based on the environmental parameters.

2. The high-dynamic adaptive environmental interference compensation and calibration system according to claim 1, wherein The preprocessing module includes: An amplitude extraction unit for extracting the maximum value of each channel sequence as the amplitude feature; A frequency domain conversion unit for extracting frequency domain features based on the discrete Fourier transform; An attenuation coefficient calculation unit for obtaining the signal attenuation coefficient using the exponential decay fitting method; A static reflection feature extraction unit for extracting static reflection features using a one-dimensional convolutional layer with fixed weights; A feature fusion unit for splicing and fusing the amplitude feature, the frequency domain feature, the signal attenuation coefficient, and the static reflection feature to generate the fused feature vector.

3. A high-dynamic adaptive environmental interference compensation and calibration system according to claim 1, characterized in that The multi-layer convolutional network includes a plurality of stacked convolutional layers, with skip connections between each convolutional layer, and a batch normalization and a ReLU activation function are sequentially connected after each convolutional layer.

4. A high-dynamic adaptive environmental interference compensation and calibration system according to claim 1, characterized in that The multi-head attention Transformer module includes: A linear layer for mapping the input features to a standard dimension and obtaining long-range temporal dependencies using the multi-head attention mechanism; A feed-forward network, including a first layer network composed of a first linear layer and a ReLU activation function and a second layer network composed of a second linear layer and layer normalization, for obtaining local spatial feature data and generating the extracted features.

5. A high-dynamic adaptive environmental interference compensation and calibration system according to claim 1, characterized in that, The adaptive physical constraint module includes: A fully connected layer for predicting the scattering factor, propagation coefficient, and phase shift based on the extracted features; A physical residual calculation module for calculating the theoretical values of the scattering factor, propagation coefficient, and phase shift based on a predefined physical model and calculating the physical residual with the corresponding predicted values of the scattering factor, propagation coefficient, and phase shift; A feature correction module for correcting the extracted features output by the front-end feature extraction module according to a preset physical loss coefficient and the physical residual to generate corrected features.

6. A high-dynamic adaptive environmental interference compensation and calibration system according to claim 1, characterized in that The RNN adaptive compensation branch includes: A temporal feature extraction unit for obtaining the fused feature vector and generating temporal data based on the fused feature vector; A comprehensive feature generation unit for mapping the fused feature vector, and concatenating the mapped fused feature vector and the temporal data in the same dimension to generate a comprehensive feature; A dynamic weight adjustment sub-unit for calculating an environmental noise evaluation value in real time according to the physical residual and dynamically updating the weight coefficient based on the environmental noise evaluation value; A signal prediction output unit for mapping the comprehensive feature to a predicted target signal by using a fully connected layer.

7. A high-dynamic adaptive environmental interference compensation and calibration system according to claim 1, characterized in that, The online environmental parameter calibration subsystem includes: A multi-source environmental data fusion module for collecting radar data, temperature and humidity data, and electromagnetic field data and fusing them to generate multi-source fusion data; An environmental parameter estimation module for mapping the multi-source fusion data to obtain an environmental parameter estimation vector; A Bayesian optimization calibration module for iteratively optimizing the detection threshold based on Bayes' theorem to obtain an optimal solution for the detection threshold; A self-supervised contrastive learning module for dynamically adjusting the weight coefficient according to the optimal threshold solution and historical data, calculating the cosine similarity in the feature space, and learning discriminative features in the case of no labels; A closed-loop feedback execution module for updating the detection threshold through a parameter mapping network according to the current environmental parameter estimation vector and the optimal solution of the detection threshold.

8. A high-dynamic adaptive environmental interference compensation and calibration system according to claim 1, characterized in that The adaptive optimization subsystem further includes a total loss function optimization module, and the expression of the optimized total loss function is: L_total = L_data * w_dynamic + λ_physical * L_physical + L_separation; In the formula, L_data is the data fitting loss; w_dynamic is the adaptive weight adjustment factor; λ_physical is the physical loss coefficient; L_physical is the physical loss; L_separation is the multi-view separation loss.

Images