Sensor data denoising processing method in electromagnetic interference environment of flight control system

Abstract

This invention discloses a method for denoising sensor data in flight control systems under electromagnetic interference environments, belonging to the field of flight control system signal processing technology. The method includes: acquiring a flight control sensor noise dataset and dividing it into a training subset and a real-time processing subset; extracting time-domain and frequency-domain features of each channel and associating them with interference intensity signal features to generate noise feature vectors; constructing a hybrid noise recognition model to identify noise patterns; performing hierarchical misjudgment correction through cosine similarity matching; marking signal segments suspected of residual noise and establishing a noise interference mapping table; performing differentiated denoising processing on marked signal segments and normal signal segments; performing secondary denoising on signal segments suspected of residual noise; collecting denoised sensor data and calculating three-dimensional indicators to evaluate the denoising effect; if the indicator meets the standard, verifying the improvement in flight control accuracy; if the indicator does not meet the standard, adjusting the parameters of the hybrid noise recognition model and iteratively optimizing it until the standard is met, thus achieving high-precision identification and adaptive denoising of flight control system noise under electromagnetic interference environments.

Description

Technical Field

[0001] This invention belongs to the field of flight control system signal processing technology, and relates to a method for noise reduction processing of sensor data under electromagnetic interference environment of flight control system. Background Technology

[0002] Flight control systems are central to aircraft flight safety, relying on high-precision data collected by sensors such as gyroscopes and accelerometers for stable control. However, the dense array of onboard electronic equipment and the complex external electromagnetic environment make sensor data susceptible to electromagnetic interference, generating noise, pulses, and drift, which can lead to misinterpretations of flight control commands in severe cases. Traditional filtering methods are effective in normal scenarios, but in environments with strong electromagnetic interference, fixed parameters are difficult to adapt to changes, resulting in signal delays or distortions. They cannot effectively separate real motion from electromagnetic noise, especially when fusing multi-sensor data, which can exacerbate error accumulation, causing control oscillations or even loss of control. Existing solutions mostly rely on hardware shielding or fixed filtering algorithms, lacking intelligent identification and collaborative noise reduction capabilities for interference spectrum characteristics and sensor dynamic responses. This makes it difficult to achieve highly reliable, low-latency sensor data purification under complex electromagnetic conditions, affecting the safe application of flight control systems in highly interference scenarios such as industrial inspections and cluster operations.

[0003] Chinese Patent CN112487730B discloses a noise suppression method for multi-rotor aircraft based on phase angle control. The method includes: the system acquiring the three-dimensional coordinates of the target noise reduction position and sending the position information to a noise prediction module; the flight control module sending attitude angles and velocity as feature parameters to the noise prediction module; the noise prediction module calculating the noise information generated by the rotor at the target position using a fast noise prediction method and sending it to a phase angle optimization module; the phase angle optimization module obtaining the optimal phase angle combination based on an optimization algorithm and sending it to a phase synchronization control module; and the phase synchronization control module adjusting the phase angle positions of multiple rotors based on the received optimal phase angle combination and phase angle position information to achieve noise reduction at the target position. This technical solution can adjust and optimize the objective function according to noise reduction requirements. Compared with passive noise reduction methods, it can specifically control the components of noise and has a better suppression effect on low-frequency noise of multi-rotor aircraft.

[0004] Existing technical solutions are only applicable to stationary Gaussian noise, which is difficult to suppress time-varying non-stationary noise, pulse spikes and coupling noise caused by electromagnetic interference. They are prone to filtering out effective transient signals and cannot balance noise reduction effect and signal fidelity, thus reducing the accuracy of flight control attitude calculation and control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for denoising sensor data in flight control systems under electromagnetic interference environments. This method solves the problems of low accuracy in noise pattern recognition of flight control sensors and the distortion of effective signals due to the reliance on a single denoising method under complex electromagnetic interference. It achieves high-precision identification of noise patterns and differentiated and efficient denoising while preserving effective transient signals from the sensors, thus improving the quality of sensor data after denoising. The invention also verifies the effect of denoising on improving the accuracy of flight control attitude calculation. Through iterative optimization of model parameters, a closed-loop improvement in denoising effect is achieved, adapting to the high-precision operating requirements of flight control systems under electromagnetic interference environments.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Methods for noise reduction of sensor data in flight control systems under electromagnetic interference environments include:

[0008] Raw sensor data and interference signals are collected, preprocessed, and sensor data features and interference features within each time window are extracted. The data are then correlated with timestamps to generate a flight control sensor noise dataset, which is then divided into a training subset and a real-time processing subset.

[0009] Based on real-time processing subset, extract time-domain and frequency-domain features of each channel and associate them with interference intensity signal features to generate noise feature vectors. Construct a hybrid noise recognition model to identify different noise patterns. Perform hierarchical misjudgment correction through cosine similarity matching. Mark signal segments suspected of residual noise and associate them with the corresponding interference-dominant frequency band and interference intensity time-series features to establish a noise interference mapping table.

[0010] Based on the noise interference mapping table, a differentiated noise reduction process is performed on the marked signal segments and normal signal segments using an adaptive multi-mode fusion noise reduction algorithm; secondary noise reduction is performed on signal segments suspected of having residual noise by combining the historical noise reduction parameters of the training subset.

[0011] After collecting sensor data for noise reduction, calculate three-dimensional indicators, set evaluation thresholds to assess the noise reduction effect, and if the target is met, compare the attitude calculation errors before and after noise reduction to verify the improvement in flight control accuracy; if the target is not met, adjust the parameters of the mixed noise recognition model and iteratively optimize until the target is met.

[0012] Specifically, the steps for identifying different noise patterns include:

[0013] Based on the real-time processing subset of the flight control sensor noise dataset, the time-domain and frequency-domain features of each channel data within each time window are extracted.

[0014] Using the time window as the unique temporal index, the temporal domain features, frequency domain features and interference intensity signal features under the same time window are time-aligned and window-by-window bound and fused to generate a noise feature vector;

[0015] A hybrid noise recognition model is constructed by combining CNN and LSTM, including an input layer, a feature extraction layer, an LSTM temporal modeling layer, an attention feature fusion layer, and an output layer;

[0016] A mixed noise recognition model was trained using a subset of the flight control sensor noise dataset as the training set, a multi-class cross-entropy loss function as the loss function, and the Adam optimizer.

[0017] After training, input the noise feature vector generated in real time, and output the noise pattern classification result and the associated electromagnetic interference type.

[0018] Specifically, the steps for constructing a hybrid noise recognition model include:

[0019] The input layer receives noise feature vectors, maps them uniformly to 256 dimensions and standardizes them, and outputs a noise feature tensor.

[0020] The feature extraction layer is used to extract multi-scale spatial features from the noise feature tensor. It employs three progressive multi-scale convolutional kernels to capture local detail differences in a single channel, characterize the global spatial coupling relationship of multiple channels, and refine the correlation between noise features and interference intensity. The adaptive mean and max pooling layers preserve the extreme value mutations and overall distribution patterns of noise features, and output a spatial feature vector.

[0021] The LSTM temporal modeling layer is used to receive spatial feature vectors, construct bidirectional gated optimized LSTM temporal modeling units, capture noise evolution trends through forward temporal coding, associate historical interference features through reverse temporal coding, and incorporate a temporal weight decay gate to output temporal feature vectors.

[0022] Specifically, the steps for constructing a hybrid noise recognition model also include:

[0023] The attention feature fusion layer is used to construct a dual-branch adaptive attention mechanism to dynamically weight and fuse spatial feature vectors and temporal feature vectors. The first branch generates an initial confidence score by compressing and reducing the dimension through a fully connected layer, and then backpropagates the gradient by combining cross-entropy loss and normalizing it through global average pooling and Softmax to obtain the contribution of spatial features to the discrimination of noisy patterns.

[0024] The second branch calculates the cosine similarity between the temporal feature vector and the interference intensity signal feature vector of the preceding historical window, encodes it into an interference evolution state vector through LSTM, multiplies it with the temporal feature vector, and obtains the representational contribution of the temporal feature through Softmax normalization; balances the spatial and temporal feature weights for different interference types, and outputs a comprehensive feature vector.

[0025] The output layer receives the synthesized feature vector and maps it to the probability distributions of the first, second, and third noise modes through a fully connected layer. The Softmax function is used to output the noise mode label corresponding to the maximum probability and the associated electromagnetic interference type.

[0026] Specifically, the steps for correcting misjudgments in the stratification process include:

[0027] Based on noise pattern labels, historical samples of the same type are selected from the historical sample library. Timeliness weights are dynamically set and expired samples are removed to obtain a timeliness candidate sample set.

[0028] Obtain real-time feature vectors, calculate the cosine similarity between the real-time feature vectors and each sample in the timely candidate sample set, and calculate the mean and standard deviation of the cosine similarity. For the first, second, and third noise modes, set first-level, second-level, and third-level thresholds respectively.

[0029] The matching percentage is calculated by combining the number of samples with a statistical cosine similarity greater than or equal to the corresponding threshold with the total number of samples in the time-sensitive candidate sample set. ;

[0030] Set percentage threshold , To correct misjudgments in the hierarchical structure;

[0031] like If so, the noise pattern classification result is deemed valid;

[0032] like If so, a weighted voting correction is triggered, and the corrected noise mode label is determined by weighted voting.

[0033] like If the result is not found, it is determined to be a high-probability misjudgment. Based on all types of historical samples, a new matching is performed, and the noise pattern label of the sample corresponding to the maximum cosine similarity is selected as the correction result.

[0034] Specifically, the steps for establishing the noise interference mapping table include:

[0035] Based on the real-time signal segments corresponding to the corrected noise pattern labels, the deviation of the eigenvalues ​​of the time-domain features and frequency-domain features of each time window from the mean of historical samples is calculated.

[0036] Set an anomaly threshold. If the deviation of three consecutive windows is greater than the anomaly threshold, the signal segment is judged as a suspected residual noise, and the interference-dominant frequency band and interference intensity temporal features are extracted.

[0037] The temporal characteristics of interference intensity are normalized, and the interference intensity level is divided into low interference, medium interference and high interference intensity levels according to the statistical distribution of the interference intensity amplitude of the training subset.

[0038] A noise interference mapping table is established with the first, second, and third noise modes as the first-level dimension, the interference intensity level as the second-level dimension, and the interference dominant frequency band type as the third-level dimension.

[0039] Specifically, the steps for differentiated noise reduction processing include:

[0040] The core features of each noise pattern in the enhanced noise interference mapping table are quantized and bound to the noise interference mapping table to generate a mapping table feature library.

[0041] Adaptive DC component removal and amplitude normalization are performed on both labeled and unlabeled signal segments in the real-time processing subset. A classified signal set is generated according to noise mode labels, and differentiated noise reduction processing is then performed, including:

[0042] Extract the interference amplitude gradient of the first noise pattern signal segment from the feature library of the mapping table, and set the gradient threshold to determine the number of wavelet decomposition levels. The db4 wavelet basis was used for... Layered wavelet decomposition yields approximation coefficients and detail coefficients;

[0043] The noise energy of the first noise mode signal segment is extracted and the total signal energy is counted. The noise energy threshold is calculated by weighting and the detail coefficients are processed by improved soft thresholding to obtain the improved detail coefficients.

[0044] The time-domain characteristic preservation coefficients are determined, and the denoised signal is obtained by performing inverse wavelet transform using the approximation coefficients and the improved detail coefficients. This denoised signal is then multiplied by the time-domain characteristic preservation coefficients to generate the first-mode denoised signal.

[0045] Specifically, the steps for differentiated noise reduction processing also include:

[0046] Extract the pulse cluster interval mean and pulse cluster interval variation coefficient of the second noise mode signal segment from the noise interference map table, detect signal amplitude abrupt change points and cluster them to generate a pulse cluster position marker set;

[0047] The filter window length is dynamically set based on the average pulse cluster interval, and the pulse threshold is dynamically calculated based on the pulse cluster interval variation coefficient.

[0048] Extract the standard amplitude. If the standard amplitude is greater than the pulse threshold, use median filtering to reduce noise and obtain the pulse region signal; otherwise, retain the original amplitude.

[0049] The pulse region signal is multiplied by the pulse cluster interval variation coefficient, and the second mode noise reduction signal is generated through amplitude calibration.

[0050] Extract the dominant frequency band center frequency and dominant frequency band drift rate of the third noise mode signal segment from the noise interference map table, calculate the notch center frequency in real time, and adjust the notch bandwidth.

[0051] An IIR notch transfer function is constructed, and amplitude compensation is performed on the third noise mode signal segment after notch filtering. The amplitude-compensated filtered signal is then multiplied by the oscillation attenuation coefficient to generate the third mode noise-reduced signal.

[0052] Specifically, the steps for secondary noise reduction include:

[0053] For signal segments suspected of having residual noise, the amount of residual noise is obtained by calculating the ratio of the noise power after noise reduction to the noise power before noise reduction. ;

[0054] Set residual threshold , To determine the residual noise level;

[0055] like If it is high residue, it is judged as high residue; if If it is, then it is determined to be a residue; if If so, it is judged as low residue;

[0056] Based on the residual noise level, the weights of wavelet denoising, median filtering, and notch filtering are dynamically allocated according to the noise pattern.

[0057] For signal segments suspected of having residual noise, wavelet denoising, median filtering denoising, and notch filtering denoising are performed respectively to obtain three intermediate denoised signals. These signals are then weighted and summed using dynamically assigned weights to generate a signal after secondary denoising.

[0058] Specifically, the steps for evaluating the noise reduction effect include:

[0059] Real-time acquisition of sensor data after noise reduction; using the target transient features of the flight control sensor transient signal annotation library as the identification benchmark, the number of effective transient signals before and after noise reduction is counted to calculate the transient signal retention rate;

[0060] The initial noise reduction signal-to-noise ratio is calculated based on the noise pattern segmentation, and the average value of multiple channels is taken as the noise reduction signal-to-noise ratio; the root mean square error is calculated with a noise-free reference signal as a reference.

[0061] Set the first, second, and third thresholds to evaluate the noise reduction effect;

[0062] If the transient signal retention rate is greater than or equal to the first threshold, the noise reduction signal-to-noise ratio is greater than or equal to the second threshold, and the root mean square error is less than or equal to the third threshold, then the noise reduction effect is deemed to meet the standard.

[0063] The noise-reduced sensor data is input into the preset flight control attitude calculation module, and the attitude angle error and position error before and after noise reduction are compared to calculate the error improvement rate.

[0064] Otherwise, return to adjust the parameters of the mixed noise recognition model, calculate the classification error rate, and set the error rate threshold;

[0065] If the classification error rate is greater than the error rate threshold, the cosine similarity threshold of the corresponding noise pattern is reduced, and the optimization is iterated until the noise reduction effect meets the standard.

[0066] The beneficial effects of this invention are:

[0067] 1. This invention constructs a hybrid noise recognition model integrating CNN and LSTM. Through a hierarchical design of the input layer, feature extraction layer, LSTM temporal modeling layer, and attention feature fusion layer, it achieves multi-scale spatial extraction and bidirectional temporal modeling of noise features. Combined with a cosine similarity-based hierarchical misjudgment correction strategy, it accurately identifies three types of electromagnetic interference noise patterns in flight control systems, labels residual noise, and establishes a noise interference mapping table. It effectively captures the local details and global coupling relationships of noise features and adapts to the time-varying, non-stationary, and abrupt intermittent characteristics of electromagnetic interference noise, significantly improving the accuracy and robustness of noise pattern recognition. This solves the problems of low classification accuracy and high misjudgment rate of traditional recognition methods for complex electromagnetic interference noise.

[0068] 2. This invention employs an adaptive multi-mode fusion denoising algorithm based on a noise interference mapping table for differentiated denoising. It matches corresponding denoising strategies to different noise modes and dynamically adjusts denoising parameters. Suspected residual noise segments are subjected to secondary denoising using historical denoising parameters. After denoising, the transient signal retention rate, denoising signal-to-noise ratio, and root mean square error are calculated to comprehensively evaluate the denoising effect. If the target is not met, the parameters of the hybrid noise identification model are iteratively optimized. This achieves efficient suppression of different types of electromagnetic interference noise, maximizes the preservation of effective transient signals, avoids distortion of effective signals caused by denoising, and verifies the effect of denoising on improving the accuracy of flight control attitude calculation. It enhances the sensor data quality and control accuracy of the flight control system, adapting to the operational requirements of flight control systems in complex electromagnetic interference environments. Attached Figure Description

[0069] Figure 1 A schematic diagram of a sensor data noise reduction processing method for a flight control system under electromagnetic interference environment;

[0070] Figure 2 This is a flowchart illustrating the identification of different noise patterns in this invention;

[0071] Figure 3 This is a flowchart illustrating the layering misjudgment correction process in this invention;

[0072] Figure 4 This is a flowchart illustrating the differentiated noise reduction process in this invention. Detailed Implementation

[0073] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0074] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a method for noise reduction processing of sensor data in an electromagnetic interference environment of a flight control system, including the following steps:

[0075] Step S1: Collect raw sensor data from each sensor in the flight control system. Use a high-precision hardware clock to synchronize and align the timestamps of the data from each sensor channel, eliminating phase errors and data misalignment between channels caused by transmission delays. Preprocess the synchronized raw sensor data using 3D printing technology. Null values ​​and abnormal jump points caused by sensor failures and transmission interruptions are removed. Inherent system errors are corrected through sensor calibration parameters. Raw sensor data of different dimensions are normalized to eliminate the influence of dimensional differences. Combined with interference signals collected by the electromagnetic environment monitoring module, sensor data features and interference features within each time window are extracted. The synchronized data, interference features and corresponding timestamps are associated and integrated, and divided into training subsets and real-time processing subsets to generate a standardized flight control sensor noise dataset. Among them, sensors include but are not limited to gyroscopes, accelerometers, and magnetometers. Raw sensor data includes but is not limited to angular rate, acceleration, and attitude angle. Sensor data features include but are not limited to time domain kurtosis, frequency domain energy, and frequency domain spectral peak value.

[0076] Step S2: Based on the real-time processing subset of the flight control sensor noise dataset, extract the temporal and frequency domain features of each channel data within each time window, associate the interference intensity signal features of the corresponding time window, generate multi-dimensional noise feature vectors, and construct a hybrid noise recognition model using CNN and LSTM to classify and identify the noise feature vectors. Different noise modes are identified based on the electromagnetic interference type, including the first noise mode, the second noise mode, and the third noise mode. Cross-validation is performed on the identified noise modes, comparing them with historical noise feature samples from the training subset of the flight control sensor noise dataset. Layered misjudgment correction of noise modes is achieved through cosine similarity matching. Signal segments suspected of residual noise in the real-time processing subset are marked, and associated with the corresponding dominant interference frequency band and temporal features of interference intensity. A noise interference mapping table between noise modes and interference features is established. Temporal features include, but are not limited to, kurtosis, variance, impulse factor, and waveform factor; frequency domain features include, but are not limited to, frequency band duty cycle and harmonic distortion rate. The first noise mode is time-varying non-stationary noise caused by continuous broadband interference; the second noise mode is pulse spike noise caused by intermittent pulse clusters; and the third noise mode is coupling noise caused by low-frequency coupled oscillations.

[0077] Step S3: Based on the noise interference map, differentiated noise reduction processing is performed on the labeled signal segments and unlabeled normal signal segments in the real-time processing subset using an adaptive multi-mode fusion noise reduction algorithm. This includes: for the first noise mode, an adaptive wavelet threshold noise reduction algorithm is used, dynamically adjusting the wavelet decomposition level and threshold according to the interference amplitude in the noise interference map to suppress non-stationary noise and retain the time-domain characteristics of the effective signal; for the second noise mode, an improved adaptive median filtering algorithm is used, setting a dynamic pulse threshold based on the pulse cluster interval in the noise interference map to remove pulse spike noise to avoid misfiltering effective transient signals; for the third noise mode, an adaptive notch filtering algorithm is used, dynamically adjusting the notch frequency bandwidth according to the interference dominant frequency band in the noise interference map to suppress coupled noise and reduce the amplitude attenuation of the effective signal; for signal segments suspected of having residual noise, a secondary noise reduction is performed using a fusion noise reduction strategy weighted by the residual noise amount, combined with the historical noise reduction parameters of the training subset, to ensure complete suppression of residual noise.

[0078] Step S4: Real-time acquisition of denoised sensor data to calculate three-dimensional indicators, including transient signal retention rate, denoised signal-to-noise ratio, and root mean square error. Evaluation thresholds are set as a first threshold, a second threshold, and a third threshold to assess the denoising effect. If the transient signal retention rate is greater than or equal to the first threshold, the denoised signal-to-noise ratio is greater than or equal to the second threshold, and the root mean square error is less than or equal to the third threshold, the denoising effect is deemed satisfactory. The denoised sensor data is then input into a preset flight control attitude calculation module to compare the attitude calculation errors before and after denoising, verifying the improvement in flight control accuracy caused by denoising. If the noise reduction effect is not satisfactory, return to step S2 to adjust the parameters of the hybrid noise recognition model. The parameters include the feature weight coefficients and cosine similarity threshold used to weight the signal time domain, frequency domain and amplitude features. Then, noise pattern recognition and noise reduction are re-performed and iteratively optimized until the standard is met. Among them, the preset flight control attitude calculation module is determined by using the original data of acceleration, angular velocity and magnetic heading collected by the inertial sensor to complete the real-time calculation of the UAV attitude angle and angular velocity through Kalman filtering, which provides the core input for attitude control and navigation decision-making of the flight control system.

[0079] Specifically, the steps for identifying different noise patterns in step S2 include:

[0080] Based on the real-time processing subset of the flight control sensor noise dataset, the temporal and frequency domain features of each channel data within each time window are extracted. Using the time window as the unique time index, the temporal and frequency domain features and interference intensity signal features under the same time window are time-aligned and window-by-window bound and fused to generate a multi-dimensional noise feature vector.

[0081] A hybrid noise recognition model is constructed by combining CNN and LSTM, including an input layer, a feature extraction layer, an LSTM temporal modeling layer, an attention feature fusion layer, and an output layer;

[0082] The input layer receives noise feature vectors, maps the noise feature vectors to a uniform 256-dimensional dimension, and performs standardization to eliminate the problems of dimensional differences and feature value imbalance. The output is a noise feature tensor with regular dimensions.

[0083] The feature extraction layer is used to extract spatial features from the input noise feature tensor at multiple scales and granularities. Considering the distribution characteristics of the three heterogeneous features of the flight control sensor—time domain, frequency domain, and interference intensity—a three-layer progressive multi-scale convolutional kernel is used to extract feature associations: the first 3×3 convolutional layer captures local detail differences in single-channel noise features; the second 5×5 convolutional layer characterizes the global spatial coupling relationship between multi-channel features; and the third 3×3 convolutional layer refines the association between noise features and interference intensity. An adaptive mean and max pooling layer preserves the extreme value mutations and overall distribution patterns of noise features, compressing redundant feature dimensions while strengthening key features strongly correlated with electromagnetic interference, outputting a high-dimensional spatial feature vector with strong pattern discrimination. These key features include, but are not limited to, time domain mutation features, frequency domain spectral peak features, and interference intensity features.

[0084] The LSTM timing modeling layer is used to receive spatial feature vectors. To address the time-varying, non-stationary, and abrupt intermittent characteristics of electromagnetic interference noise, a bidirectional gated optimized LSTM timing modeling unit is constructed. Forward timing encoding captures the continuous evolution trend of noise features, while backward timing encoding traces the correlation effects of historical interference features. An embedded timing weight attenuation gate is also included to enhance the weight of noise features in recent time windows. This allows for high-precision adaptation to the gradual characteristics of continuous broadband interference, the abrupt jump characteristics of intermittent pulse clusters, and the periodic characteristics of low-frequency coupled oscillations. In turn, it captures the timing dependencies, abrupt change characteristics, and periodic patterns of the three types of noise modes, and outputs timing feature vectors adapted to complex interference scenarios in flight control.

[0085] The attention feature fusion layer is used to construct a two-branch adaptive attention mechanism to perform spatiotemporal collaborative dynamic weighted fusion of spatial feature vectors and temporal feature vectors for noise pattern perception. This includes: the first branch inputs the spatial feature vectors into two cascaded fully connected layers. The first fully connected layer compresses the dimension to 64, and the second fully connected layer outputs a dimension of 3, corresponding to three noise patterns. An initial confidence score for each noise pattern is generated using the Softmax function. The gradient of the cross-entropy loss between the initial confidence score and the true label is calculated with respect to the spatial feature vector. The absolute value of the gradient is taken to eliminate positive and negative influences. Global average pooling is performed on the spatial dimension to obtain a sensitivity coefficient representing the importance of each spatial feature dimension. The sensitivity coefficient is normalized using Softmax to finally obtain the contribution of the spatial feature to the discrimination of the noise pattern. The second branch calculates the cosine similarity between the temporal feature vector and the interference intensity signal feature vectors of the five consecutive historical windows preceding the current time window. A five-dimensional similarity sequence is generated and encoded into an interference evolution state vector using a single-layer LSTM. This vector is then multiplied element-wise with the temporal feature vector and mapped to 128-dimensional original weights through a fully connected layer. After Softmax normalization, the contribution of temporal features to the representation of interference evolution is obtained. Spatial feature weights are strengthened for continuous broadband interference, temporal abrupt change feature weights are strengthened for intermittent pulse interference, and spatial and temporal feature weights are strengthened for low-frequency coupled oscillation equilibrium. Weakly correlated features with low contribution to noise pattern differentiation and interference evolution representation are eliminated through weighted normalization. The complementary spatiotemporal features are deeply coupled to output a comprehensive feature vector with both spatial discriminativeness and temporal representation, thereby fundamentally improving the accuracy and robustness of noise pattern recognition under complex electromagnetic interference. Among them, weakly correlated features refer to spatiotemporal feature components that do not have efficient ability to distinguish different noise patterns, lack representational value for interference temporal evolution, and are prone to introducing redundant information or even interfering with the discrimination results.

[0086] The output layer receives the comprehensive feature vector, maps it to the probability distributions of the first noise mode, the second noise mode, and the third noise mode through a fully connected layer, and uses the Softmax activation function to output the noise mode label corresponding to the maximum probability and associate it with the corresponding electromagnetic interference type.

[0087] Using a subset of the flight control sensor noise dataset as the training set, and a multi-class cross-entropy loss function adapted to three types of noise pattern classification tasks as the loss function, a hybrid noise recognition model is trained using the Adam optimizer. After training, the multi-dimensional noise feature vector generated in real time is input, and the noise pattern classification results and associated electromagnetic interference types are output.

[0088] For example, taking the noise identification of a total of 6 channels of MEMS gyroscopes and accelerometers in a fixed-wing UAV flight control system as an example, a real-time processing subset of the flight control sensor noise dataset is extracted, and the time window length is set to 20. The sliding step size is 10. Extract the time-domain and frequency-domain features of each channel data within the 500th time window. For example, the mean value of the gyroscope X-axis channel data is 0.018. The variance of the channel data is 0.0008. The peak value was 0.035. The mean value of the accelerometer Z-axis channel data is 9.805. The variance of the channel data is 0.0015. The kurtosis is 2.9, and a total of 24 temporal features are obtained from 4 classes × 6 channels; 0-1000 are extracted through FFT transformation. The main frequency within the frequency band is 280. The peak power spectral density is 0.075. Frequency band energy ratio 0-200 The proportion is 52%, with a total of 3 categories × 6 channels = 18 frequency domain features; and using the time window as the unique time sequence index, the time domain features, frequency domain features, and electromagnetic interference intensity signal features under the same window are time-aligned and window-by-window bound and fused to generate a 44-dimensional noise feature vector; among which, the electromagnetic interference intensity signal features include electric field strength 10.5 The magnetic field strength is 0.048. ;

[0089] The input layer maps the 44-dimensional noise feature vector to a 256-dimensional vector, and after Z-score normalization, outputs a noise feature tensor of size 1×256.

[0090] The feature extraction layer employs a three-layer progressive multi-scale convolutional kernel layer. The first layer is a 3×3 convolution with 64 kernels, a stride of 1, and 1 pixel padding to capture local details in a single channel. The second layer is a 5×5 convolution with 128 kernels, a stride of 1, and 2 pixels padding to characterize the global coupling relationship of multiple channels. The third layer is a 3×3 convolution with 256 kernels, a stride of 1, and 1 pixel padding to refine the correlation between noise and interference intensity. Combined with a 2×2 adaptive mean and max pooling layer to compress redundant features, the output is a 128-dimensional spatial feature vector with strong pattern discriminativeness.

[0091] The LSTM temporal modeling layer receives spatial feature vectors and optimizes the LSTM temporal modeling unit based on a bidirectional gating system with 128 units and 8 consecutive time steps. Forward temporal coding captures the continuous evolution trend of noise, and reverse temporal coding traces back the correlation effects of historical interference. A built-in temporal weight attenuation gate with an attenuation coefficient of 0.75 increases the weights of the four most recent windows, which are 1.0, 0.9, 0.8, and 0.7, respectively. After adapting to three types of noise characteristics, a 128-dimensional temporal feature vector is output.

[0092] The attention feature fusion layer constructs a dual-branch adaptive attention mechanism, using the mutual information value between the feature and the intermittent pulse cluster interference class as the core to calculate the contribution of each branch: The first branch targets the 128-dimensional spatial feature vector, calculating the mutual information value between the spatial feature vector and the target noise class to be 0.15, and obtaining a discrimination contribution of 0.18 after min-max normalization; The second branch targets the 128-dimensional temporal feature vector, calculating the mutual information value between the temporal feature vector and the target noise class to be 0.62, and obtaining a representation contribution of 0.78 using the same normalization rule; The contribution of the two branches is weighted according to the normalization formula, and the spatial feature weight is calculated to be 0.2 and the temporal mutation feature weight is calculated to be 0.8. The two types of feature vectors are weighted and fused dimension by dimension by combining the spatial feature weight and the temporal mutation feature weight, and weakly correlated dimensions with feature values ​​less than 0.05 are removed, outputting a 256-dimensional comprehensive feature vector with regular dimensions;

[0093] The output layer receives a 256-dimensional comprehensive feature vector, which is mapped to the probability distributions of the first noise mode, the second noise mode, and the third noise mode through a fully connected layer. The Softmax activation function is used to output the noise mode label corresponding to the maximum probability and associate it with the corresponding electromagnetic interference type.

[0094] Using a training subset of 100,000 data points as the training set, with a batch size of 64 and an Adam optimizer learning rate of 1e-4, a mixed noise recognition model was trained using a multi-class cross-entropy loss function. After 200 rounds of iterative training, the training loss converged to 0.02, and the recognition accuracy on the validation set reached 98.5%, meeting the noise pattern recognition requirements in complex electromagnetic interference scenarios of flight control.

[0095] After training, the input is a 256-dimensional noise feature vector generated in real time. The output outputs the probability distributions of three noise modes: the first noise mode is 0.95, the second noise mode is 0.03, and the third noise mode is 0.02. The label corresponding to the highest probability value is the first noise mode, associated with electromagnetic interference type 150. Continuous broadband electromagnetic interference.

[0096] Specifically, the steps for establishing the noise interference mapping table in step S2 include:

[0097] Based on the noise pattern label corresponding to the maximum probability, historical samples of the same type are selected from the historical sample library to form a candidate sample set;

[0098] According to the time interval between sample collection time and the current time Dynamically set timeliness weights and set interval thresholds based on the timeliness test results of flight control sensor noise data. , ,like If so, then set the timeliness weight to 1.0; if If so, the timeliness weight will be linearly decayed to 0.7; if If the timeliness weight is set to 0.5, expired samples with a timeliness weight less than 0.5 are removed to obtain a timeliness candidate sample set, thus solving the matching bias problem caused by the decay of the timeliness of historical samples.

[0099] Real-time feature vectors are obtained from the currently acquired flight control sensor data after noise reduction preprocessing, and the cosine similarity between the real-time feature vectors and each sample in the time-sensitive candidate sample set is calculated. This forms a cosine similarity sequence, expressed as follows:

[0100]

[0101] in, For real-time feature vectors, The feature vectors of historical samples are obtained from the time-sensitive candidate sample set. The dot product of the real-time feature vector and the historical sample feature vector, || || represents the L2 norm. This is the sample index in the candidate sample set for timeliness. , The total number of samples in the time-sensitive candidate sample set;

[0102] According to the statistical formula for calculating the arithmetic mean Calculate the mean of the cosine similarity sequence. According to the statistical sample standard deviation formula Calculate the standard deviation of cosine similarity sequences ;in, To sum the cosine similarities between all samples in the time-sensitive candidate sample set and the real-time feature vector, and then divide by... Get the mean , Cosine similarity sequence and mean The sum of squared deviations is calculated as the sum of squared deviations relative to the total number of samples. The standard deviation is obtained by subtracting 1 from the square root. ;

[0103] For the first noise pattern, a first-level threshold is set based on the sum of the mean and 1.2 times the standard deviation of the cosine similarity sequence. It adapts to time-varying, non-stationary fluctuations; for the second noise mode, a secondary threshold is set based on the sum of the mean and 0.8 times the standard deviation of the cosine similarity sequence. It adapts to the characteristics of sudden pulse spikes; for the third noise mode, a three-level threshold is set based on the sum of the mean and 0.5 times the standard deviation of the cosine similarity sequence. Adapt to the stability of periodic characteristics;

[0104] Calculate the cosine similarity separately The number of samples greater than or equal to the corresponding threshold is counted, and the total number of samples in the time-sensitive candidate sample set is counted. The ratio of the number of samples meeting the corresponding threshold to the total number of samples in the time-sensitive candidate sample set is calculated to obtain the matching ratio. The corresponding threshold is the first-level threshold. Secondary threshold With Level 3 Threshold ;

[0105] Set the percentage threshold based on the statistical results of the cross-validation misclassification rate of the training subset. , ,and Perform layered misjudgment correction, if If the noise pattern classification result is valid, the noise pattern classification result and associated electromagnetic interference type output by the mixed noise recognition model are retained.

[0106] like If the cosine similarity is among the top 5 historical samples, the product of the cosine similarity and the timeliness weight is calculated to obtain the voting weight. The weighted voting determines the corrected noise pattern label, and the reason for the correction and the weight distribution are recorded.

[0107] like If the result is not found, it is determined to be a high-probability misjudgment. All historical samples of all types are retrieved for rematching. The noise pattern label of the sample corresponding to the maximum cosine similarity is selected as the correction result and marked as a high-confidence correction.

[0108] Based on the real-time signal segments corresponding to the corrected noise pattern labels, the deviation of the eigenvalues ​​of the time-domain features and frequency-domain features of each time window from the historical sample mean is calculated according to the formula |real-time feature value - historical sample mean| / historical sample standard deviation.

[0109] An anomaly threshold is set based on the 95th percentile of the deviation of the noise pattern features corresponding to the training subset. If the deviation is greater than or equal to 3 consecutive windows, it is determined to be a signal segment suspected of residual noise. The interference-dominant frequency band and interference intensity temporal features of the signal segment are extracted.

[0110] The temporal characteristics of interference intensity are normalized, and the interference intensity levels are classified according to the statistical distribution of the interference intensity amplitude of the training subset: the mean of the normalized interference intensity amplitude is divided into three levels according to the 25th percentile and 75th percentile. If the mean of the interference intensity amplitude is less than or equal to the 25th percentile, it is classified as low interference intensity level; if the mean of the interference intensity amplitude is between the 25th percentile and the 75th percentile, it is classified as medium interference intensity level; if the mean of the interference intensity amplitude is greater than or equal to the 75th percentile, it is classified as high interference intensity level.

[0111] A hierarchical noise interference mapping table is established, with the first-level dimension being the first noise mode, the second noise mode, and the third noise mode; the second-level dimension being the interference intensity level; and the third-level dimension being the dominant frequency band type of interference, namely low frequency, mid frequency, and high frequency.

[0112] For example, assuming the maximum probability corresponds to the second noise mode label, 100 historical samples of the second noise mode are selected from the historical sample library to form a candidate sample set; according to the time interval between the sample collection time and the current time... Dynamically set timeliness weights and set interval thresholds based on the timeliness test results of flight control sensor noise data. , The collection time of 20 samples However, if the timeliness weight is less than 0.5, 20 samples will be removed, and the collection time of 5 samples will be considered. With a timeliness weight of 1.0, the collection time of the 75 samples was determined as follows: Calculate the time-related weights using the linear decay formula. , among which, when At that time, the timeliness weight is 1.0. At that time, the timeliness weight is 1.0 - 0.3 = 0.7, and the timeliness weight is in to Linear variation between them; There are 75 samples within the interval, although the specific time for each sample is unknown. The differences cause their respective timeliness weights to vary between [0.7, 1.0], but when calculating the average timeliness, it is usually assumed that these samples are within [0.7, 1.0]. The interval is uniformly distributed. For the linearly changing interval [0.7, 1.0], the average timeliness weight is calculated as (1.0 + 0.7) / 2 = 0.85 by (maximum value + minimum value) / 2. Finally, 80 timeliness candidate sample sets are obtained, which solves the matching bias problem caused by the timeliness decay of historical samples.

[0113] Calculate real-time feature vectors ,For example The time-domain and frequency-domain characteristics of pulse spike noise and the feature vectors of 80 time-dependent candidate samples are used to characterize the noise. cosine similarity This forms a cosine similarity sequence, for example With the first The dot product is , , Calculate the cosine similarity of the first sample. Based on the cosine similarity sequence of 80 samples, for example The mean is calculated as Standard deviation is Set a secondary threshold for the second noise mode. To adapt to the characteristics of sudden pulse spikes; the number of samples with a statistical cosine similarity greater than or equal to 0.92 was 60, and the total number of samples in the time-sensitive candidate sample set was 80. The matching ratio was calculated. Set the percentage threshold based on the statistical results of the cross-validation misclassification rate of the training subset. , ,because If the classification result of the second noise mode is deemed valid, the second noise mode label output by the mixed noise identification model and the associated intermittent pulse cluster electromagnetic interference type are retained. Based on the real-time signal segment corresponding to the corrected second noise mode label, taking the pulse factor as an example, the historical sample mean of the second noise mode pulse factor is 5.2, the historical sample standard deviation is 1.1, and the pulse factor of real-time window 1 is 6.8. The formula for calculating the deviation is: Deviation The impulse factor for window 2 is 7.0, and the calculated deviation is approximately equal to... The impulse factor for window 3 is 6.9, and the calculated deviation is approximately equal to... The anomaly threshold was set to 1.5 based on the 95th quantile of the deviation of the second noise mode features in the training subset. Since the deviation of all three consecutive windows was greater than 1.5, the signal segment was identified as potentially residual noise. The interference-dominant frequency band and temporal features of the interference intensity were extracted from the signal segment. The interference-dominant frequency band features included a center frequency of... bandwidth is The temporal characteristics of the interference intensity are normalized to obtain the normalized interference intensity amplitude sequence. The mean value was calculated to be 0.8. Combined with the 25th and 75th quantiles of the interference intensity amplitude of the second noise mode in the training subset, since 0.8 is greater than the 75th quantile, the signal segment was classified as a high interference intensity level. Finally, a hierarchical noise interference mapping table was established. The first dimension of the noise interference mapping table corresponding to the signal segment is the second noise mode, the second dimension is the high interference intensity level, and the third dimension corresponds to the mid-frequency interference dominant band type and the dominant frequency band. In The mid-frequency range.

[0114] Specifically, the differentiated noise reduction process in step S3 includes the following steps:

[0115] The core features of each noise mode in the noise interference mapping table are enhanced by quantization, including: for the first noise mode, extracting the interference amplitude gradient and the proportion of broadband coverage frequency band, and performing normalized quantization in the [0,1] interval to enhance the feature discrimination of non-stationary noise; wherein, the interference amplitude gradient is calculated by: selecting the time-series amplitude data of the broadband interference signal, calculating the amplitude difference between two adjacent moments according to the time step, taking the average of the absolute values ​​of all amplitude differences to obtain the interference amplitude gradient, which is used to quantify the rate of change of noise amplitude; the proportion of broadband coverage frequency band is calculated by: determining the actual frequency band range covered by the broadband interference, calculating the ratio of the frequency band range to the operating frequency band range of the flight control sensor, and obtaining the proportion of broadband coverage frequency band, which is used to quantify the degree of noise coverage of the sensor's operating frequency band;

[0116] For the second noise mode, the pulse cluster interval variation coefficient is calculated by the ratio of the standard deviation of the pulse cluster interval to the mean. Calculate the ratio of peak amplitude to background signal to quantify pulse density;

[0117] For the third noise mode, the dominant frequency band drift rate and oscillation attenuation coefficient are statistically analyzed to quantify the stability of the coupled noise. The dominant frequency band drift rate is calculated as follows: selecting time-series spectral data of low-frequency coupled oscillations, extracting the dominant frequency band within each time window according to a set time window, calculating the absolute difference between the dominant frequency bands of adjacent time windows, taking the average of all absolute differences, and calculating the ratio of the average value to the initial dominant frequency band frequency value to obtain the dominant frequency band drift rate, which is used to quantify the degree of drift of the dominant frequency band. The oscillation attenuation coefficient is calculated as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] An exponential fitting is performed, and the coefficient of the fitted exponential term is the oscillation decay coefficient. The larger the oscillation decay coefficient, the faster the oscillation decays and the stronger the stability of the coupled noise. In this embodiment, the time window is set to 5 minutes to adapt to the slow drift characteristics of low-frequency coupled oscillation. This effectively captures the dynamic changes of the dominant frequency band while avoiding calculation errors caused by an excessively short window and omissions of drift features caused by an excessively long window. The frequency value of the initial dominant frequency band is determined by extracting the frequency point with the largest spectral amplitude within the initial time window (which is consistent with the set time window) of the low-frequency coupled oscillation time-series spectrum data.

[0118] The enhanced core features are bound to the noise interference mapping table, a feature parameter linkage index is constructed, and a parsed mapping table feature library is generated.

[0119] The labeled signal segments and unlabeled normal signal segments in the real-time processing subset undergo adaptive DC component removal processing to obtain pure original signal segments after DC component removal. Max-min normalization is used to map the signal amplitude to the [-1,1] interval to eliminate dimensional differences between different sensor channels. The standardized signal segments are then classified according to noise mode labels to generate a classified signal set, including first noise mode signal segments, second noise mode signal segments, and third noise mode signal segments. Differential noise reduction processing is then applied to the classified signal set, including:

[0120] The wavelet decomposition level is dynamically calculated based on the adaptive wavelet threshold denoising algorithm, including: extracting the interference amplitude gradient of the first noise pattern signal segment from the feature library of the mapping table. Based on the statistical mean of the interference amplitude gradient of the first noise mode in the feature library of the mapping table, with standard deviation Set gradient threshold , ;like If the non-stationarity is strong, then the number of wavelet decomposition levels should be set. ;like If the non-stationarity is moderate, then the number of wavelet decomposition levels is set. ;like If the non-stationarity is weak, then the number of wavelet decomposition levels is set. The first noise mode signal segment was processed using a db4 wavelet basis adapted to the non-stationary noise characteristics of the flight control system. Layer wavelet decomposition, specifically the decomposition process includes: using the original signal as the initial layer, performing high-pass filtering and low-pass filtering on the original signal sequentially using the db4 wavelet basis to separate high-frequency and low-frequency components; after each filtering process, performing a downsampling operation to reduce dimensionality, and repeating the filtering and downsampling operations. Finally, the low-frequency approximation coefficients at the bottom layer of the decomposition are obtained. and high-frequency detail coefficients distributed across each layer ;

[0121] Extract the noise energy of the first noise mode signal segment from the noise interference map table. And count the total signal energy ,according to Weighted calculation of noise energy threshold ,in, To determine the optimal initial noise threshold for the training subset, the SURE thresholding method is used based on the statistical characteristics of the detail coefficients of the training subset. An improved soft thresholding process is then applied to the detail coefficients, based on... Improved detail coefficients ,in, , The wavelet decomposition level is denoted as . For the first The original detail coefficients obtained from layer wavelet decomposition For the first Improved detail coefficients obtained from layer wavelet decomposition Based on the absolute value of detail coefficients and noise energy threshold The exponential decay term, constructed from the ratio of [value], is used to quantify the degree of attenuation of the noise component in the detail coefficients. To improve the attenuation correction factor for the soft threshold;

[0122] according to Determine the time-domain characteristic retention coefficients and use approximation coefficients. With improved detail coefficient The denoised signal is obtained by performing an inverse wavelet transform. This denoised signal is then multiplied by the time-domain characteristic preservation coefficients to generate the first-mode denoised signal. To preserve the fundamental coefficients for time-domain characteristics, This is a non-stationarity fit correction term;

[0123] Extract the pulse cluster interval mean of the second noise mode signal segment from the noise interference map table. Pulse cluster interval variation coefficient ;use Detecting abrupt changes in signal amplitude, if If , then it is marked as a pulse candidate point; where, This represents the amplitude difference between the current sampling point and the previous sampling point. The standard amplitude of the current sampling point. The standard amplitude of the previous sampling point. The standard deviation of the background signal is calculated based on the amplitude of the background region of the current second noise mode signal segment.

[0124] Combined with the average pulse cluster interval Cluster the continuous pulse candidate points into pulse clusters and generate a pulse cluster location marker set;

[0125] Based on the average pulse cluster interval Dynamically set the filter window length The mean threshold is set based on the statistical interval of the mean interval of the second noise mode pulse cluster in the feature library of the mapping table. , ,and ,like Then set the filter window length. ;like Then set the filter window length. ;like Then set the filter window length. ;

[0126] according to Dynamic calculation of pulse threshold ,in, To determine the optimal initial pulse threshold for the training subset, the maximum inter-class variance method is used, based on the pulse amplitude distribution characteristics of the second noise mode signal in the training subset. This is a pulse density correction factor; the standard amplitude of each pulse candidate point is extracted, and if the standard amplitude is greater than... If the signal is positive, median filtering is used for noise reduction to obtain the pulse region signal after median filtering; otherwise, it is determined to be a valid transient signal and the original amplitude is retained.

[0127] The pulse region signal is multiplied by the pulse cluster interval variation coefficient, and the second-mode noise reduction signal is dynamically adjusted using an amplitude calibration factor; wherein, the amplitude calibration factor is based on... Sure, To calibrate the baseline value for amplitude, This is a pulse interval stability correction term;

[0128] Extract the center frequency of the dominant frequency band of the third noise mode signal segment from the noise interference map table. Dominant frequency band drift rate ;according to Real-time calculation of notch center frequency ,in, This is the current time window number. It serves as the dominant frequency band drift correction factor, adapting to the time-varying characteristics of the coupled noise frequency band;

[0129] The drift rate threshold is set based on the statistical threshold of the drift rate of the dominant frequency band of the third noise mode in the feature library of the mapping table. To adjust the notch bandwidth ,like Then through Calculate the notch bandwidth; otherwise, pass through The notch bandwidth is calculated to balance noise reduction effect and signal fidelity; among which, It is a broadband symmetry coefficient, set according to the characteristic of the notch filter to extend symmetrically on both sides of the center frequency; The high drift rate bandwidth coefficient was determined through experimental statistics based on the requirement of covering the main noise frequency band under high frequency drift conditions. This is the low drift rate bandwidth coefficient, set according to the trade-off principle of ensuring signal fidelity and narrowband notch filtering under low drift conditions;

[0130] Adaptive notch filtering noise reduction is performed using an IIR notch filter. For normalized angular frequency, For the first coefficient term, combine the first coefficient term with... Summing the terms gives the first combinatorial term; subtracting the first combinatorial term from 1 gives the numerator polynomial. For the second coefficient term, combine the second coefficient term with... Summing yields the second combination term, and subtracting the second combination term from 1 gives the denominator polynomial. The IIR notch transfer function is constructed by calculating the ratio of the numerator polynomial to the denominator polynomial, as shown in the following expression:

[0131]

[0132] in, For IIR notch filters The domain transfer function is used to characterize the amplitude-frequency response of a filter to signals of different frequencies. The damping coefficient and , The sampling rate of the flight control sensor. The signal value at the previous sampling time; These are the signal values ​​at the first two sampling times, respectively;

[0133] Notch filtering is performed on the third noise mode signal segment to obtain the filtered normalized noise signal. The amplitude attenuation rate is obtained by calculating the ratio of the average amplitude after filtering to the average amplitude before filtering. The normalized noise signal is divided by the amplitude attenuation rate to obtain the amplitude-compensated filtered signal, so as to reduce the amplitude attenuation of the effective signal.

[0134] The product of the amplitude-compensated filtered signal and the oscillation attenuation coefficient is calculated to generate the third-mode noise-reduced signal;

[0135] For signal segments suspected of having residual noise, the amount of residual noise is obtained by calculating the ratio of the noise power after noise reduction to the noise power before noise reduction. The residual threshold is set based on the statistical results of residual noise in the training subset and the noise reduction accuracy requirements of the flight control sensors. , ,and Perform residual noise classification and determination, if If it is high residue, it is judged as high residue; if If it is, then it is determined to be a residue; if If so, it is judged as low residue;

[0136] The weights for wavelet denoising, median filtering, and notch filtering are assigned based on the residual noise level. , and ,and If the remaining noise is mode 1, then set... , , The first mode noise is continuous broadband interference, and the appropriate denoising algorithm is wavelet denoising. Therefore, the base weight is set to 0.6, which has the highest proportion, and the weight of the median filter is... Set the base weight of the notch filter to 0.2, and introduce... Dynamic correction The larger the value, the higher the weight. Synchronous improvement, weight Synchronous reduction, while ensuring Under the premise of strengthening targeted suppression of strong non-stationary broadband noise;

[0137] If the remaining noise is mode 2, then set , , The second mode noise is intermittent impulse interference, and the appropriate denoising algorithm is median filtering. Therefore, the base weight is set to 0.6, which has the highest proportion, and the weight for wavelet denoising is... Set the base weight of the notch filter to 0.2, and introduce... Dynamic correction The larger the value, the higher the weight. Synchronous improvement, weight Synchronous reduction, while ensuring Under the premise of strengthening the targeted filtering of dense pulse interference;

[0138] If the remaining noise is third-mode noise, then set , , The third mode noise is low-frequency coupled oscillation, and the appropriate denoising algorithm is notch filtering. Therefore, the base weight is set to 0.6, which has the highest proportion, and the weight of wavelet denoising is... Set the base weight of the median filter to 0.2, and introduce... Dynamic correction The larger the value, the higher the weight. Synchronous reduction, weight Synchronous improvement, while ensuring Under the premise of strengthening targeted notch filtering for strongly coupled oscillation noise;

[0139] For signal segments suspected of having residual noise, wavelet denoising, median filtering denoising, and notch filtering denoising were performed respectively, resulting in three intermediate denoised signals. , and Combined with weights , and A weighted summation is performed to generate a signal after secondary noise reduction, ensuring that residual noise is completely suppressed.

[0140] For example, in a noise reduction scenario involving a subset of real-time processing of flight control sensors, the core features of the noise interference mapping table are quantized and enhanced: the first noise mode extracts the interference amplitude gradient. The broadband coverage frequency band ratio is 0.8 and normalized; the pulse cluster interval variation coefficient is calculated for the second noise mode. The ratio of peak amplitude to background signal is 2.5; the third noise mode statistical dominant frequency band drift rate is... The oscillation attenuation coefficient is 0.98, and a parsed mapping table feature library is generated.

[0141] After removing the DC component and performing maximum-minimum normalization on the real-time signal segments, they are classified according to noise mode: the interference amplitude gradient of the first noise mode signal segment. Statistical mean in the noise interference mapping table Standard deviation Therefore, gradient threshold , ,because If the non-stationarity is strong, then the number of wavelet decomposition levels should be set. The approximate coefficients are obtained by using db4 wavelet basis decomposition. With detail coefficient Noise energy Total signal energy Optimal initial noise threshold for training subset Therefore, the noise threshold For detail coefficients The improved soft thresholding process was used to obtain The time-domain characteristic retention factor is 1. The wavelet inverse transform is multiplied by the time-domain characteristic preservation coefficients to generate the first mode denoised signal;

[0142] Mean pulse cluster interval of the second noise mode signal segment Pulse cluster interval variation coefficient Background signal standard deviation Detected If it is, then it is marked as a pulse candidate point in the noise interference mapping table. Statistical interval , Then the filter window length Optimal initial pulse threshold for training subset Pulse threshold Amplitude greater than The pulse region uses median filtering, and the amplitude calibration factor is [value missing]. The pulse region signal and After multiplication, the amplitude calibration factor is adjusted to generate the second-mode noise reduction signal;

[0143] The dominant frequency band center frequency of the third noise mode signal segment Dominant frequency band drift rate Current time window number Therefore, the center frequency of the notch filter is , The notch bandwidth is... Flight control sensor sampling rate Damping coefficient Construct the IIR notch transfer function, with the numerator being... The denominator is After filtering, the amplitude attenuation rate is 0.95. After amplitude compensation, it is multiplied by the oscillation attenuation coefficient of 0.98 to generate the third mode noise reduction signal.

[0144] For suspected residual noise segments, calculate the residual noise level. Based on the statistical results of residual noise in the training subset and the noise reduction accuracy requirements of the flight control sensors, a residual threshold is set. , ,because If it is found to be residual, it is determined to be second-mode noise, and weights are assigned accordingly. , , For signal segments suspected of having residual noise, wavelet denoising, median filtering denoising, and notch filtering denoising were performed respectively, resulting in three intermediate denoised signals. , and The weighted summation yields the second-order denoised signal. This ensures that residual noise is completely suppressed.

[0145] Specifically, the steps for evaluating the noise reduction effect in step S4 include:

[0146] Real-time acquisition of denoised sensor data; extraction of target transient features from a flight control sensor transient signal annotation library as an identification benchmark; and statistical analysis of valid transient signals meeting the identification benchmark before and after denoising. The number of valid transient signals before and after denoising is then calculated. , ,pass Calculate the transient signal retention rate;

[0147] Based on noise pattern, the area is divided into sections. Calculate the initial noise reduction signal-to-noise ratio, where, The effective energy after noise reduction is obtained by summing the squares of the amplitudes of the effective signal segments. The residual noise energy after noise reduction is obtained by summing the squares of the amplitudes of the residual noise segments; the average initial noise reduction signal-to-noise ratio of the multi-channel sensor data is taken as the noise reduction signal-to-noise ratio.

[0148] Using the noise-free reference signal calibrated by the flight control sensor as a reference, according to Calculate the root mean square error, where For the noise reduction Frame amplitude, For noise-free reference signal Frame amplitude, The total number of valid data frames. This is the sum of squared errors between the denoised signal and the noise-free reference signal;

[0149] The evaluation thresholds are set according to the flight control system attitude calculation accuracy requirements, historical noise reduction experimental data and industry standards for noise reduction indicators of aerospace sensors. The first threshold is 95%, the second threshold is 25dB and the third threshold is 0.02V. The noise reduction effect is evaluated by combining the three-dimensional indicators of transient signal retention rate, noise reduction signal-to-noise ratio and root mean square error.

[0150] If the transient signal retention rate is greater than or equal to the first threshold, the noise reduction signal-to-noise ratio is greater than or equal to the second threshold, and the root mean square error is less than or equal to the third threshold, then the noise reduction effect is deemed satisfactory. The noise-reduced sensor data is then input into the flight control attitude calculation module to calculate the attitude angles. Calculate the attitude angle error before and after noise reduction ,in, For the solution value, The true value is the result of the fusion of GPS and inertial navigation; based on Calculate position error ;in, , , These are the calculated 3D position values ​​of the UAV output by the flight control attitude calculation module. , , These are the actual 3D positions of the UAV obtained after fusing GPS and inertial navigation;

[0151] Compare the attitude calculation errors before and after noise reduction, including based on Calculate the attitude angle error improvement rate based on The position error improvement rate was calculated to verify the effect of noise reduction on improving flight control accuracy.

[0152] Otherwise, the noise reduction effect is deemed unsatisfactory, and the unsatisfactory index type is marked, including unsatisfactory transient signal retention rate, unsatisfactory noise reduction signal-to-noise ratio, and unsatisfactory root mean square error. Return to step S2 and adjust the feature weight coefficients of the hybrid noise recognition model by a step size of 0.05. If the transient signal retention rate is unsatisfactory, reduce the pulse amplitude feature weight by 0.05 and increase the pulse interval feature weight by 0.05. Statistically count the number of erroneous signal segments that do not match the actual noise pattern and the total number of signal segments divided by the real-time processing subset according to the time window. Calculate the ratio of the number of erroneous signal segments to the total number of signal segments to obtain the classification error rate. Based on the historical accuracy requirements of flight control noise pattern recognition and the noise reduction effect compliance rate target, set the error rate threshold to 10%. If the classification error rate is greater than 10%, reduce the cosine similarity threshold of the corresponding noise pattern by 0.03 to expand the matching range, and re-perform noise pattern recognition and differentiated noise reduction processing until the noise reduction effect meets the standard.

[0153] In summary, this invention aims to address the technical problems of existing noise reduction methods, which are only suitable for stationary Gaussian noise, have difficulty suppressing time-varying non-stationary noise, and are prone to falsely filtering effective transient signals, failing to balance noise reduction effectiveness and signal fidelity. It generates a flight control sensor noise dataset through multi-sensor data synchronization and alignment, preprocessing, and correlation with electromagnetic interference features. A CNN-LSTM-based hybrid noise recognition model combined with a spatiotemporal attention mechanism achieves high-precision recognition of three types of electromagnetic noise patterns, establishing a dynamic mapping relationship between noise patterns and interference features. A mode-specific adaptive noise reduction strategy is employed, using adaptive wavelet thresholding for continuous broadband interference and improved adaptive median filtering for intermittent pulse spikes. Adaptive notch filtering is used for low-frequency coupled oscillations, and weighted fusion secondary noise reduction is performed on suspected residual noise segments to suppress complex electromagnetic noise while preserving attitude transient features to the greatest extent. The noise reduction effect is evaluated in a closed loop using three-dimensional indicators: transient signal retention rate, noise reduction signal-to-noise ratio, and root mean square error. Based on the evaluation results, the feature weights and cosine similarity threshold of the mixed noise identification model are adaptively adjusted to form a complete iterative mechanism of identification, noise reduction, evaluation, and optimization. This invention is effectively adapted to non-stationary flight control sensor scenarios with multiple types of coupled noise under strong electromagnetic interference, improving noise suppression capability and effective signal fidelity, reducing attitude calculation errors, and improving the control accuracy and operational reliability of the flight control system in complex electromagnetic environments.

[0154] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for noise reduction processing of sensor data under electromagnetic interference environment in flight control system, characterized in that, include: Raw sensor data and interference signals are collected, preprocessed, and sensor data features and interference features within each time window are extracted. The data are then correlated with timestamps to generate a flight control sensor noise dataset, which is then divided into a training subset and a real-time processing subset. Based on real-time processing subset, extract time-domain and frequency-domain features of each channel and associate them with interference intensity signal features to generate noise feature vectors. Construct a hybrid noise recognition model to identify different noise patterns. Perform hierarchical misjudgment correction through cosine similarity matching. Mark signal segments suspected of residual noise and associate them with the corresponding interference-dominant frequency band and interference intensity time-series features to establish a noise interference mapping table. Based on the noise interference mapping table, the marked signal segments and normal signal segments are distinguished by an adaptive multi-mode fusion noise reduction algorithm for differentiated noise reduction processing. Secondary denoising is performed on signal segments suspected of having residual noise by combining historical denoising parameters from the training subset; After collecting sensor data for noise reduction, calculate three-dimensional indicators, set evaluation thresholds to assess the noise reduction effect, and if the target is met, compare the attitude calculation errors before and after noise reduction to verify the improvement in flight control accuracy; if the target is not met, adjust the parameters of the mixed noise recognition model and iteratively optimize until the target is met.

2. The method for noise reduction of sensor data under electromagnetic interference environment in a flight control system according to claim 1, characterized in that, The specific steps for identifying different noise patterns include: Based on the real-time processing subset of the flight control sensor noise dataset, the time-domain and frequency-domain features of each channel data within each time window are extracted. Using the time window as the unique temporal index, the temporal domain features, frequency domain features and interference intensity signal features under the same time window are time-aligned and window-by-window bound and fused to generate a noise feature vector; A hybrid noise recognition model is constructed by combining CNN and LSTM, including an input layer, a feature extraction layer, an LSTM temporal modeling layer, an attention feature fusion layer, and an output layer; A mixed noise recognition model was trained using a subset of the flight control sensor noise dataset as the training set, a multi-class cross-entropy loss function as the loss function, and the Adam optimizer. After training, input the noise feature vector generated in real time, and output the noise pattern classification result and the associated electromagnetic interference type.

3. The method for noise reduction of sensor data under electromagnetic interference environment in a flight control system according to claim 2, characterized in that, The specific steps for constructing a hybrid noise recognition model include: The input layer receives noise feature vectors, maps them uniformly to 256 dimensions and standardizes them, and outputs a noise feature tensor. The feature extraction layer is used to extract multi-scale spatial features from the noise feature tensor. It employs three progressive multi-scale convolutional kernels to capture local detail differences in a single channel, characterize the global spatial coupling relationship of multiple channels, and refine the correlation between noise features and interference intensity. The adaptive mean and max pooling layers preserve the extreme value mutations and overall distribution patterns of noise features, and output a spatial feature vector. The LSTM temporal modeling layer is used to receive spatial feature vectors, construct bidirectional gated optimized LSTM temporal modeling units, capture noise evolution trends through forward temporal coding, associate historical interference features through reverse temporal coding, and incorporate a temporal weight decay gate to output temporal feature vectors.

4. The method for noise reduction of sensor data under electromagnetic interference environment in a flight control system according to claim 3, characterized in that, The specific steps for constructing a hybrid noise recognition model also include: The attention feature fusion layer is used to construct a dual-branch adaptive attention mechanism to dynamically weight and fuse spatial feature vectors and temporal feature vectors. The first branch generates an initial confidence score by compressing and reducing the dimension through a fully connected layer, and then backpropagates the gradient by combining cross-entropy loss and normalizing it through global average pooling and Softmax to obtain the contribution of spatial features to the discrimination of noisy patterns. The second branch calculates the cosine similarity between the temporal feature vector and the interference intensity signal feature vector of the preceding historical window, encodes it into an interference evolution state vector through LSTM, multiplies it with the temporal feature vector, and obtains the representational contribution of the temporal feature through Softmax normalization; balances the spatial and temporal feature weights for different interference types, and outputs a comprehensive feature vector. The output layer receives the synthesized feature vector and maps it to the probability distributions of the first, second, and third noise modes through a fully connected layer. The Softmax function is used to output the noise mode label corresponding to the maximum probability and the associated electromagnetic interference type.

5. The method for noise reduction of sensor data under electromagnetic interference environment in a flight control system according to claim 4, characterized in that, The specific steps for correcting misjudgments in stratification include: Based on noise pattern labels, historical samples of the same type are selected from the historical sample library. Timeliness weights are dynamically set and expired samples are removed to obtain a timeliness candidate sample set. Obtain real-time feature vectors, calculate the cosine similarity between the real-time feature vectors and each sample in the timely candidate sample set, and calculate the mean and standard deviation of the cosine similarity. For the first, second, and third noise modes, set first-level, second-level, and third-level thresholds respectively. The matching percentage is calculated by combining the number of samples with a statistical cosine similarity greater than or equal to the corresponding threshold with the total number of samples in the time-sensitive candidate sample set. ; Set percentage threshold , To correct misjudgments in the hierarchical structure; like If so, the noise pattern classification result is deemed valid; like If so, a weighted voting correction is triggered, and the corrected noise mode label is determined by weighted voting. like If the result is not found, it is determined to be a high-probability misjudgment. Based on all types of historical samples, a new matching is performed, and the noise pattern label of the sample corresponding to the maximum cosine similarity is selected as the correction result.

6. The method for noise reduction processing of sensor data under electromagnetic interference environment in a flight control system according to claim 5, characterized in that, The specific steps for establishing a noise interference mapping table include: Based on the real-time signal segments corresponding to the corrected noise pattern labels, the deviation of the eigenvalues ​​of the time-domain features and frequency-domain features of each time window from the mean of historical samples is calculated. Set an anomaly threshold. If the deviation of three consecutive windows is greater than the anomaly threshold, the signal segment is judged as a suspected residual noise, and the interference-dominant frequency band and interference intensity temporal features are extracted. The temporal characteristics of interference intensity are normalized, and the interference intensity level is divided into low interference, medium interference and high interference intensity levels according to the statistical distribution of the interference intensity amplitude of the training subset. A noise interference mapping table is established with the first, second, and third noise modes as the first-level dimension, the interference intensity level as the second-level dimension, and the interference dominant frequency band type as the third-level dimension.

7. The method for noise reduction processing of sensor data under electromagnetic interference environment in a flight control system according to claim 6, characterized in that, The specific steps for performing differentiated noise reduction processing include: The core features of each noise pattern in the enhanced noise interference mapping table are quantized and bound to the noise interference mapping table to generate a mapping table feature library. Adaptive DC component removal and amplitude normalization are performed on both labeled and unlabeled signal segments in the real-time processing subset. A classified signal set is generated according to noise mode labels, and differentiated noise reduction processing is then performed, including: Extract the interference amplitude gradient of the first noise pattern signal segment from the feature library of the mapping table, and set the gradient threshold to determine the number of wavelet decomposition levels. The db4 wavelet basis was used for... Layered wavelet decomposition yields approximation coefficients and detail coefficients; The noise energy of the first noise mode signal segment is extracted and the total signal energy is counted. The noise energy threshold is calculated by weighting and the detail coefficients are processed by improved soft thresholding to obtain the improved detail coefficients. The time-domain characteristic preservation coefficients are determined, and the denoised signal is obtained by performing inverse wavelet transform using the approximation coefficients and the improved detail coefficients. This denoised signal is then multiplied by the time-domain characteristic preservation coefficients to generate the first-mode denoised signal.

8. The method for noise reduction processing of sensor data under electromagnetic interference environment in a flight control system according to claim 7, characterized in that, The specific steps for performing differentiated noise reduction processing also include: Extract the pulse cluster interval mean and pulse cluster interval variation coefficient of the second noise mode signal segment from the noise interference map table, detect signal amplitude abrupt change points and cluster them to generate a pulse cluster position marker set; The filter window length is dynamically set based on the average pulse cluster interval, and the pulse threshold is dynamically calculated based on the pulse cluster interval variation coefficient. Extract the standard amplitude. If the standard amplitude is greater than the pulse threshold, use median filtering to reduce noise and obtain the pulse region signal; otherwise, retain the original amplitude. The pulse region signal is multiplied by the pulse cluster interval variation coefficient, and the second mode noise reduction signal is generated through amplitude calibration. Extract the dominant frequency band center frequency and dominant frequency band drift rate of the third noise mode signal segment from the noise interference map table, calculate the notch center frequency in real time, and adjust the notch bandwidth. An IIR notch transfer function is constructed, and amplitude compensation is performed on the third noise mode signal segment after notch filtering. The amplitude-compensated filtered signal is then multiplied by the oscillation attenuation coefficient to generate the third mode noise-reduced signal.

9. The method for noise reduction processing of sensor data under electromagnetic interference environment in a flight control system according to claim 8, characterized in that, The specific steps for secondary noise reduction include: For signal segments suspected of having residual noise, the amount of residual noise is obtained by calculating the ratio of the noise power after noise reduction to the noise power before noise reduction. ; Set residual threshold , To determine the residual noise level; like If it is high residue, it is judged as high residue; if If it is, then it is determined to be a residue; if If so, it is judged as low residue; Based on the residual noise level, the weights of wavelet denoising, median filtering, and notch filtering are dynamically allocated according to the noise pattern. For signal segments suspected of having residual noise, wavelet denoising, median filtering denoising, and notch filtering denoising are performed respectively to obtain three intermediate denoised signals. These signals are then weighted and summed using dynamically assigned weights to generate a signal after secondary denoising.

10. The method for noise reduction processing of sensor data under electromagnetic interference environment in a flight control system according to claim 9, characterized in that, The specific steps for evaluating noise reduction effectiveness include: Real-time acquisition of sensor data after noise reduction; using the target transient features of the flight control sensor transient signal annotation library as the identification benchmark, the number of effective transient signals before and after noise reduction is counted to calculate the transient signal retention rate; The initial noise reduction signal-to-noise ratio is calculated based on the noise pattern segmentation, and the average value of multiple channels is taken as the noise reduction signal-to-noise ratio; the root mean square error is calculated with a noise-free reference signal as a reference. Set the first, second, and third thresholds to evaluate the noise reduction effect; If the transient signal retention rate is greater than or equal to the first threshold, the noise reduction signal-to-noise ratio is greater than or equal to the second threshold, and the root mean square error is less than or equal to the third threshold, then the noise reduction effect is deemed to meet the standard. The noise-reduced sensor data is input into the preset flight control attitude calculation module, and the attitude angle error and position error before and after noise reduction are compared to calculate the error improvement rate. Otherwise, return to adjust the parameters of the mixed noise recognition model, calculate the classification error rate, and set the error rate threshold; If the classification error rate is greater than the error rate threshold, the cosine similarity threshold of the corresponding noise pattern is reduced, and the optimization is iterated until the noise reduction effect meets the standard.

Images