MEMS IMU sensor anti-interference method and system based on auto-encoder

Abstract

The invention requests to protect an MEMS IMU sensor anti-interference method based on an auto-encoder. The method relates to the technical field of MEMS IMU sensor anti-interference, and comprises the following steps: acquiring normal working data of an MEMS IMU sensor and interference data of different types and different intensities, preprocessing the data to construct a data set, and training an AE self-encoding model fusing feature extraction and classification evaluation functions; in practical application, the model receives sensor output data in real time, judges whether interference exists or not through reconstruction errors, outputs interference types and interference intensity levels through a built-in classification head and an intensity evaluation head, and provides an accurate decision basis for subsequent anti-interference coping strategies. According to the method, the features do not need to be manually designed, rapid identification, accurate classification and intensity quantification of interference are realized by virtue of the deep feature automatic extraction capability of AE self-coding, the method has the characteristics of strong real-time performance, excellent generalization capability and wide adaptability, and the technical problems of weak pertinence and low identification precision of an existing MEMS sensor anti-interference method are effectively solved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of MEMS (Micro-Electro-Mechanical Systems) sensor technology and deep learning. Specifically, it relates to an anti-interference method for MEMS IMU sensors based on self-encoders (AEs), which is applicable to MEMS IMU sensor application scenarios that require high-precision and high-reliability data output, such as inertial navigation, industrial inspection, and smart wearable devices. Background Technology

[0002] MEMS IMU sensors have been widely used in aerospace, automotive electronics, consumer electronics, and other fields due to their advantages such as small size, light weight, low cost, and low power consumption. However, the operating environment of MEMS IMU sensors often involves complex external interferences such as electromagnetic interference and mechanical vibration interference. These interferences can cause distortion of the sensor output data, seriously affecting the accuracy of subsequent decision-making and control based on sensor data.

[0003] Existing MEMS IMU sensor anti-interference technologies are mainly divided into two categories: hardware anti-interference and software anti-interference. Hardware anti-interference methods are achieved by optimizing the sensor packaging structure, adding shielding layers or filtering circuits, but their design cycle is long, the cost is high, and the adaptability to different types of interference is poor. Software anti-interference methods mainly include digital filtering (such as Kalman filtering and mean filtering), feature engineering combined with traditional machine learning (such as SVM and decision trees), etc. Among them, digital filtering can only suppress interference in specific frequency bands, and the effect on handling unknown types of interference is limited. Traditional machine learning relies on manually designing time-domain (such as peak value and variance) or frequency-domain (such as spectral peak value and bandwidth) features. Feature engineering is time-consuming and has weak generalization ability, making it difficult to adapt to complex interference scenarios with multiple types and varying intensities.

[0004] In recent years, deep learning technology has demonstrated powerful automatic feature extraction capabilities in the field of signal recognition. For example, recurrent neural networks (RNNs) and one-dimensional convolutional neural networks (1D-CNNs) have been attempted for sensor interference identification. However, RNNs suffer from the vanishing gradient problem and are inefficient for processing long sequences of data. While 1D-CNNs can extract local features, they lack the ability to capture global features of interference. Autoencoders (AEs), as an unsupervised / semi-supervised learning model, achieve feature extraction and reconstruction through the symmetrical structure of encoders and decoders, possessing natural advantages in anomaly detection and feature learning. However, their application in the entire process of "existence judgment - type classification - intensity assessment" of sensor interference in MEMSMU has not yet yielded a mature solution.

[0005] Therefore, there is an urgent need for a MEMS IMU sensor anti-interference method that can automatically extract interference features, accurately identify interference states, and output decision-making basis in real time, so as to solve the technical defects of existing technologies such as weak targeting, poor generalization ability, and insufficient real-time performance.

[0006] A search revealed application publication number CN116599624A, entitled "An Adaptive Anti-interference Algorithm Based on Autoencoder and Meta-learning," which relates to an adaptive anti-interference algorithm. This invention aims to address the problem that traditional anti-interference algorithms cannot automatically adjust parameters to adapt to different types of interference, whether known or unknown. The steps of this invention are: Step 1, offline meta-learning; Step 2, online learning and fine-tuning.

[0007] 1. Overview of the technical content of the prior art document CN116599624A

[0008] A search revealed CN116599624A, which discloses an adaptive anti-interference scheme for wireless communication systems. This scheme combines an autoencoder and meta-learning, updating the parameters of the transceiver link through offline training and online fine-tuning to adapt to different channel and interference conditions, thereby improving the data transmission reliability of the communication link. The technical object of this prior art is the transceiver process of a communication system, and the target of this processing is the transmission performance under interference conditions in the communication link.

[0009] 2. Shortcomings of the prior art and differences from the present invention

[0010] (1) The comparison document mainly focuses on channel and interference adaptation of wireless communication links. It does not establish a data acquisition, sample construction and sensor-side inference process for interference identification and suppression scenarios of MEMS IMU sensor output data, and is difficult to be directly used for interference processing of MEMS IMU sensors.

[0011] (2) The comparison document mainly uses link parameter updates, does not provide an output method for interference status in response to sensor interference, and does not generate output results on interference existence, interference type and interference intensity, which is not conducive to selecting the corresponding handling method on the sensor side based on the interference status.

[0012] (3) The comparison document involves a joint mechanism of offline training and online fine-tuning, and relies on the collaborative process at both ends of the communication link, which is difficult to meet the requirements of MEMS IMU sensor on inference latency, power consumption and resource consumption under embedded conditions.

[0013] (4) The comparative document does not disclose the sensor-side response strategy output and execution mechanism corresponding to the interference state. Its anti-interference effect is mainly achieved indirectly through communication parameter updates, which is difficult to use for the sensor-side active processing of interference and strategy execution.

[0014] 3. The technical solution adopted by the present invention to address the above-mentioned shortcomings.

[0015] (1) This invention is aimed at the interference processing scenario of MEMS IMU sensor. It collects normal working data and data under different interference types and intensities, and constructs a labeled dataset through preprocessing for training and verification of anti-interference model.

[0016] (2) The present invention constructs a model structure including an encoder, a decoder and an output of interference type and intensity, and uses the reconstruction error to determine the existence of interference, while outputting the interference type and interference intensity level to provide a basis for subsequent strategy selection.

[0017] (3) The present invention adopts a one-dimensional convolution and deconvolution structure suitable for embedded inference, avoiding the online fine-tuning method of co-updating the sending and receiving ends, so that the model can complete real-time inference on an embedded processor or FPGA platform.

[0018] (4) The present invention sets up a response strategy output mechanism that matches the type and intensity of interference, and selects processing methods such as filtering, frequency band switching or data fusion according to the reasoning results, thereby realizing anti-interference processing and strategy execution at the sensor end. Summary of the Invention

[0019] This invention aims to solve the problems of the prior art mentioned above. It proposes a method for anti-interference of MEMS IMU sensors based on a self-encoder. The technical solution of this invention is as follows:

[0020] An anti-interference method for MEMS IMU sensors based on self-encoders includes the following steps:

[0021] Step 1: Data Acquisition and Preprocessing: Acquire the output data of the MEMS IMU sensor under normal operating conditions, as well as the output data under different interference types and intensities. Clean, synchronize, and standardize all acquired data to obtain a standardized dataset.

[0022] Step 2: Dataset Construction: Divide the standardized dataset into training set, validation set and test set according to a preset ratio. The dataset contains sample labels, which include "no interference", interference type and interference intensity labels.

[0023] Step 3: AE autoencoder model construction and training: Construct an AE autoencoder model that includes an encoder, decoder, interference classification head, and interference intensity evaluation head; train the model using the training set, optimize the encoder's feature extraction capability through the decoder's reconstruction error, and optimize the prediction accuracy of the classification head and evaluation head through classification loss and intensity regression loss. Adjust the model's hyperparameters using the validation set until the model converges.

[0024] Step 4: Real-time anti-interference reasoning: After processing the real-time output data of the MEMS IMU sensor according to the preprocessing rules in Step 1, input it into the trained AE model. The model outputs the interference existence judgment result, interference type classification result, and interference intensity evaluation result.

[0025] Step 5: Output of Countermeasure Strategy: Based on the model output results, match and output the preset anti-interference countermeasure strategy to realize intelligent anti-interference of MEMS IMU sensor.

[0026] Furthermore, the interference types in step 1 include single-frequency interference, frequency sweeping interference, and broadband noise interference. The interference intensity is divided into three levels: weak, medium, and strong, according to a preset power threshold. The preprocessing includes: removing outliers and missing values ​​from the data, performing time synchronization calibration on the data, mapping the data to a preset interval using Z-score standardization, and obtaining fixed-length input samples by segmenting the data through a sliding window.

[0027] Furthermore, in step 3, the encoder includes a one-dimensional convolutional layer, an activation function layer, a pooling layer, and a fully connected layer connected in sequence, used to map the input sample into a low-dimensional feature vector; the decoder includes a fully connected layer, a deconvolutional layer, and an activation function layer connected in sequence, used to reconstruct the low-dimensional feature vector into a reconstructed sample with the same dimension as the input sample; the interference classification head is a stacked structure of fully connected layers, outputting the probability distribution of various interferences and interference-free states; the interference intensity evaluation head is a fully connected regression layer, outputting the quantized value or level label of the interference intensity.

[0028] Furthermore, the encoder has multiple one-dimensional convolutional layers, with the kernel size, stride, and padding method configured as needed; the activation function is the ReLU function; the pooling layer uses max pooling or average pooling, with the kernel size and stride configured as needed; the decoder's deconvolutional layer has a symmetrical structure with the encoder's convolutional layer, and the output layer activation function is the Sigmoid function.

[0029] Furthermore, in step 3, the total loss function for model training is a weighted sum of reconstruction loss, classification loss, and intensity regression loss; wherein, the reconstruction loss is calculated based on the mean squared error criterion, the classification loss is calculated based on the cross-entropy criterion, and the intensity regression loss is calculated based on the mean absolute error criterion, and the weighting coefficients of each loss term are configured according to a preset ratio.

[0030] Furthermore, the interference existence judgment rule in step 4 is as follows: when the reconstruction error of the model output is greater than a preset threshold, interference is determined to exist; otherwise, it is determined to be without interference; the reconstruction error is the mean square error or cosine distance between the input sample and the reconstructed sample.

[0031] Furthermore, the anti-interference strategies in step 5 include: maintaining the original operating parameters of the sensor when there is no interference; using digital filtering algorithms to suppress interference when there is weak interference; switching the sensor operating frequency band or adjusting the sampling rate when there is medium interference; and starting redundant sensor data fusion or pausing non-critical data output when there is strong interference.

[0032] A MEMS IMU sensor anti-interference system includes a data acquisition module, a data preprocessing module, a model inference module, and a strategy execution module. The data acquisition module is used to acquire real-time output data from the MEMS IMU sensor. The data preprocessing module is used to perform the preprocessing operation in step 1. The model inference module has an AE autoencoder model trained by any of the methods described above, and outputs interference-related judgment results. The strategy execution module is used to execute the corresponding anti-interference strategy based on the inference results.

[0033] Furthermore, the data acquisition module includes a MEMS IMU sensor body and a data cache unit, the cache capacity of which meets the continuous input requirements of the sliding window samples; the model inference module is deployed on an embedded processor or FPGA chip, and the inference latency meets the real-time processing requirements.

[0034] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the self-encoder-based MEMS IMU sensor anti-interference methods.

[0035] The advantages and beneficial effects of this invention are as follows:

[0036] 1. Automated feature extraction: Adopting the AE autoencoder model, there is no need for manual design of time-domain / frequency-domain features. The encoder automatically extracts the deep abstract features of the interference signal, avoiding the subjectivity and limitations of feature engineering in traditional methods and improving the adaptability to complex interference.

[0037] 2. End-to-end interference perception: It integrates interference presence determination, type classification, and intensity assessment, overcoming the shortcomings of existing methods that can only determine "whether there is interference" or identify a single type of interference, and providing a comprehensive and accurate decision-making basis for anti-interference strategies;

[0038] 3. Excellent real-time performance and generalization ability: The model adopts a lightweight one-dimensional convolution and deconvolution structure, with low inference latency, meeting the requirements for real-time anti-interference; through training with multi-type and multi-intensity interference data, the model has strong generalization ability and can be adapted to various sensors such as MEMS accelerometers and gyroscopes and different application scenarios.

[0039] 4. High practicality: The anti-interference response strategy is precisely matched with the interference intensity and type, avoiding the "one-size-fits-all" anti-interference approach. While ensuring data accuracy, it reduces the power consumption and computing resource consumption of the sensor.

[0040] 5. Flexible deployment: The model can be deployed on various hardware platforms such as embedded processors and FPGAs without relying on high-performance computing devices, and is adapted to the miniaturization and low power consumption application requirements of MEMS IMU sensors.

[0041] The innovations of this invention are mainly concentrated in the methods and system architectures defined by claims 1, 2, 3, 5, 6, 7, 8, and 9: Claim 1 proposes an anti-interference processing flow based on an autoencoder, including at least data acquisition and preprocessing, dataset construction, model construction and training, online inference, and output of response strategies; Claim 2 defines the types and intensity levels of interference and the preprocessing flow, wherein the preprocessing includes at least outlier handling, time synchronization, Z-score normalization, and sliding window segmentation; Claim 3 defines the structure of the AE autoencoder model, which includes an encoder, a decoder, and a functional module for outputting the type and intensity of interference; Claim 5 defines the model training method, employing a joint loss function consisting of reconstruction loss, classification loss, and intensity regression loss; Claim 6 defines the method for determining the existence of interference, based on reconstruction error or similarity measurement and a threshold determined through a validation set; Claim 7 defines the anti-interference response strategy, enabling the response method to adjust according to changes in the type and intensity level of interference; Claim 8 provides the system module composition and its functional connections; Claim 9 defines the system deployment method, ensuring that model inference and data caching meet the real-time processing requirements of embedded processors or FPGA platforms.

[0042] Compared with existing technologies, this invention does not rely on manually constructed time-domain or frequency-domain features. Instead, it utilizes an autoencoder to learn features from the input sequence and provides interference criteria through reconstruction errors. Simultaneously, it combines classification and intensity assessment outputs, enabling the system to provide information on the existence, type, and intensity of interference. Furthermore, by constructing datasets with interference type and intensity labels and using a joint loss training method, the model's adaptability to multi-type and multi-intensity interference scenarios is improved. Through a modular system structure and embedded deployment design, the above method can achieve online inference and policy execution on resource-constrained platforms, thereby improving the real-time performance and applicability of anti-interference processing. Attached Figure Description

[0043] Figure 1 This is a comparison chart showing the distribution of reconstruction errors.

[0044] Figure 2 This is a scatter plot showing the predicted interference intensity.

[0045] Figure 3 This is a comparison chart showing the reconstruction results;

[0046] Figure 4 This represents the training loss curve and the testing loss curve of the autoencoder;

[0047] Figure 5 This is a flowchart illustrating the entire process of anti-interference processing based on an autoencoder.

[0048] Figure 6 This diagram illustrates the structure of an anti-interference processing model based on an autoencoder. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0050] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0051] The purpose of this invention is to provide an anti-interference method for MEMS IMU sensors based on self-encoders, achieving the following technical effects:

[0052] 1. No manual feature design is required; deep time-domain / frequency-domain features of interference signals are automatically extracted through the AE autoencoder model.

[0053] 2. Simultaneously, it completes the determination of the existence of interference, the classification of interference types, and the assessment of interference intensity, providing a comprehensive decision-making basis for anti-interference strategies;

[0054] 3. Improve the model's real-time performance and generalization ability to adapt to different types of MEMS IMU sensors and complex interference scenarios.

[0055] Technical solution

[0056] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0057] Overall process

[0058] The overall process of the self-encoder-based MEMS IMU sensor anti-interference method of the present invention is as follows: Figure 1 As shown, the process includes five core steps: data acquisition and preprocessing, dataset construction, AE autoencoder model construction and training, real-time anti-interference inference, and response strategy output, as detailed below:

[0059] Step 1: Data Acquisition and Preprocessing

[0060] 1. Data Acquisition: Select the target MEMS IMU sensor and acquire normal output data without interference under standard working conditions; apply single-frequency interference (frequency range 1kHz-1MHz), sweep frequency interference (start and stop frequency 1kHz-10MHz, sweep speed 100Hz / s-1kHz / s), and broadband noise interference (bandwidth 10kHz-100MHz) through a signal generator or simulate actual interference scenarios, and divide them into three intensity levels: weak, medium, and strong according to preset power thresholds (such as -60dBm, -30dBm, 0dBm), and acquire sensor output data under each type and intensity.

[0061] 2. Data preprocessing:

[0062] Outlier and missing value handling: Outliers are removed using the 3σ criterion, and missing values ​​are filled using linear interpolation;

[0063] Time synchronization: Timestamp calibration of multi-channel sensor data ensures data timing consistency;

[0064] Standardization: Z-score standardization is adopted.

[0065]

[0066] Where μ is the mean and σ is the standard deviation, the data is mapped to the interval [-1, 1] to eliminate the influence of dimensions;

[0067] Sample segmentation: The sliding window method is used to segment the data. The window size and step size are configured according to the actual data characteristics and processing requirements to obtain input samples of fixed length.

[0068] Step 2: Dataset Construction

[0069] Label each input sample with a triplet:

[0070] Interference presence indicator (0=no interference, 1=interference present), interference type indicator (0=single frequency, 1=sweep frequency, 2=wideband noise), interference intensity indicator (0=weak, 1=medium, 2=strong).

[0071] The labeled samples were divided into training, validation, and test sets in a 7:2:1 ratio for model training, hyperparameter tuning, and performance verification.

[0072] Step 3: AE Autoencoder Model Construction and Training

[0073] Model structure as follows Figure 2 As shown, an integrated architecture of "encoder-decoder-classification head-strength evaluation head" is adopted:

[0074] 1. Encoder: Used to extract deep features from input samples, it adopts a stacked structure of "one-dimensional convolutional layer → activation layer → pooling layer", and the number of convolutional layers can be set as needed; the number of input and output channels, convolutional kernel size, stride and padding method of each layer are configured according to the feature extraction requirements; low-dimensional feature vectors are obtained by mapping through fully connected layers.

[0075] 2. Decoder: Used to reconstruct the input samples. Its structure is symmetrical to that of the encoder, and it adopts a stacked structure of "fully connected layer → deconvolution layer → activation layer". The fully connected layer maps the low-dimensional feature vector to the dimension that is suitable for deconvolution processing. The parameters of each deconvolution layer are configured as needed, and the final output is a reconstructed sample with the same dimension as the input sample.

[0076] 3. Interference Classification Header: It adopts a fully connected layer stacking structure, and the number of fully connected layers and the output dimension of each layer can be set as needed; the hidden layers use activation functions such as ReLU, and the output layers use the Softmax activation function to output the probability distribution of each type of interference and no-interference category.

[0077] 4. Interference Intensity Assessment Head: It adopts a fully connected layer structure, and the output dimension is adapted to the interference intensity quantization requirements. It uses a linear activation function to output the quantized value of the interference intensity, which can be mapped to a preset intensity level.

[0078] 5. Model training process:

[0079] (1) Loss function: The weighted sum of reconstruction loss, classification loss and intensity regression loss is adopted; wherein, the reconstruction loss is calculated based on the mean square error criterion, the classification loss is calculated based on the cross-entropy criterion, and the intensity regression loss is calculated based on the mean absolute error criterion. The weighting coefficients of each loss term are configured as needed according to the training effect.

[0080]

[0081] in, The reconstruction loss (MSE) is α = 0.7; The classification loss (cross-entropy loss) is β=0.2; The magnitude regression loss (MAE) is γ=0.1;

[0082] (2) Optimizer: The Adam optimizer is used, and the learning rate and decay strategy are configured according to the training convergence requirements;

[0083] (3) Training parameters: Batch size, number of training rounds and decision criteria for stopping training are set as needed to ensure that the model converges to the optimal performance and saves the optimal model.

[0084] Step 4: Real-time anti-interference inference

[0085] 1. Real-time data input: After the real-time output data of the MEMS sensor is preprocessed in step 1 (outlier handling, standardization, sliding window segmentation), real-time samples are generated;

[0086] 2. Model Inference: Input real-time samples into the trained AE autoencoder model to obtain reconstructed samples, classification probability distribution, and intensity quantization values;

[0087] 3. Result determination:

[0088] (1) Existence of interference: Calculate the mean square error between the real-time sample and the reconstructed sample. If the mean square error is greater than the preset threshold (which is determined by statistical analysis of the validation set data), it is judged as "interference exists"; otherwise, it is "no interference".

[0089] (2) Interference type: Select the category with the highest probability in the classification probability distribution as the interference type;

[0090] (3) Interference intensity: Based on the intensity quantification value, it is mapped to three levels: weak, medium and strong.

[0091] Step 5: Outputting Response Strategies

[0092] A pre-defined anti-interference strategy mapping table is used to match the corresponding strategy based on the model output.

[0093] 1. No interference: Maintains the original operating parameters of the MEMS IMU sensor (sampling rate, filtering coefficient, etc.) and outputs data normally;

[0094] 2. Weak-intensity interference: Activate the adaptive Kalman filter algorithm to suppress interference signals;

[0095] 3. Medium-intensity interference: Switch the sensor's operating frequency band or adjust the sampling rate;

[0096] 4. Strong interference: Initiate redundant sensor data fusion or suspend non-critical data output, retaining only core monitoring data.

[0097] System Architecture

[0098] Based on the above method, the present invention also provides a MEMS IMU sensor anti-interference system, comprising:

[0099] 1. Data Acquisition Module: Composed of a MEMS IMU sensor body and a data buffer unit; the sensor body acquires raw data, and the buffer capacity of the buffer unit meets the continuous input requirements of the sliding window samples, ensuring the continuity and stability of data transmission;

[0100] 2. Data preprocessing module: Implemented based on FPGA or MCU, it quickly completes outlier handling, standardization, and sample segmentation, with processing latency meeting real-time requirements;

[0101] 3. Model Inference Module: Deployed on an embedded processor or FPGA chip, with a built-in trained AE autoencoder model, and inference latency meets real-time processing requirements;

[0102] 4. Strategy Execution Module: Connected to the control unit of the MEMS IMU sensor, it adjusts the sensor's operating parameters or activates the corresponding anti-interference algorithm based on the model inference results.

[0103] The following describes the implementation process of the present invention in detail with reference to the accompanying drawings and parameter examples:

[0104] 1. Data Acquisition and Preprocessing: A certain type of MEMS IMU (sampling rate 1kHz, range ±16g) was selected to collect interference-free normal data and three types of interference data (including different intensity levels): single frequency, sweep frequency, and broadband noise. Outliers were removed using the 3σ criterion, and Z-score was standardized to the [-1,1] interval. Data was segmented with 512 sampling points as the sliding window and a step size of 256 to construct a dataset containing approximately 1200 training samples, 340 validation samples, and 170 test samples. The sample labels included three-element information: "interference presence, type, and intensity".

[0105] 2. Model Training: (According to...) Figure 6 The structure shown constructs an AE autoencoder model, trained for 50 epochs using the Adam optimizer (initial learning rate 0.001, batch size = 32). The loss converges rapidly during training, as... Figure 4 As shown, the model significantly distinguishes between reconstruction errors of normal data and interference data, such as... Figure 1 As shown.

[0106] 3. Performance Evaluation: On the test set, the accuracy rate for determining the presence of interference reached 99.2% (based on a reconstruction error threshold of 0.045); the accuracy rate for evaluating interference intensity was 97.8%. Figure 2 As shown, the model accurately reconstructs normal signals and various types of interference signals, as indicated. Figure 3 As shown.

[0107] 4. Real-time Inference and Policy Execution: After the model is deployed on an ARM Cortex-A7 processor (1GHz), it can analyze sensor outputs in real time. When "medium-intensity single-frequency interference" is detected, the adaptive Kalman filter algorithm is activated, and the root mean square error (RMSE) of the filtered data is reduced from 0.25g to 0.03g. When "strong-intensity broadband noise interference" is detected, dual-sensor data fusion is activated, improving data reliability by more than 40%.

[0108] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0109] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0110] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0111] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for anti-interference of MEMS IMU sensors based on self-encoders, characterized in that, Includes the following steps: Step 1: Data Acquisition and Preprocessing: Acquire the output data of the MEMS IMU sensor under normal operating conditions, as well as the output data under different interference types and intensities. Clean, synchronize, and standardize all acquired data to obtain a standardized dataset. Step 2: Dataset Construction: Divide the standardized dataset into training set, validation set and test set according to a preset ratio. The dataset contains sample labels, which include "no interference", interference type and interference intensity labels. Step 3: AE autoencoder model construction and training: Construct an AE autoencoder model that includes an encoder, decoder, interference classification head, and interference intensity evaluation head; train the model using the training set, optimize the encoder's feature extraction capability through the decoder's reconstruction error, and optimize the prediction accuracy of the classification head and evaluation head through classification loss and intensity regression loss. Adjust the model's hyperparameters using the validation set until the model converges. Step 4: Real-time anti-interference reasoning: After processing the real-time output data of the MEMS IMU sensor according to the preprocessing rules in Step 1, input it into the trained AE model. The model outputs the interference existence judgment result, interference type classification result, and interference intensity evaluation result. Step 5: Output of Countermeasure Strategy: Based on the model output results, match and output the preset anti-interference countermeasure strategy to realize intelligent anti-interference of MEMS IMU sensor.

2. The anti-interference method for MEMS IMU sensors based on self-encoders according to claim 1, characterized in that, The interference types in step 1 include single-frequency interference, frequency sweeping interference, and broadband noise interference. The interference intensity is divided into three levels: weak, medium, and strong, according to a preset power threshold. The preprocessing includes: removing outliers and missing values ​​from the data, performing time synchronization calibration on the data, mapping the data to a preset interval using Z-score standardization, and obtaining fixed-length input samples by segmenting the data through a sliding window.

3. The anti-interference method for MEMS IMU sensors based on self-encoders according to claim 1, characterized in that, In step 3, the encoder includes a one-dimensional convolutional layer, an activation function layer, a pooling layer, and a fully connected layer connected in sequence, used to map the input sample into a low-dimensional feature vector; the decoder includes a fully connected layer, a deconvolutional layer, and an activation function layer connected in sequence, used to reconstruct the low-dimensional feature vector into a reconstructed sample with the same dimension as the input sample; the interference classification head is a stacked structure of fully connected layers, outputting the probability distribution of various interferences and interference-free states; the interference intensity evaluation head is a fully connected regression layer, outputting the quantized value or level label of the interference intensity.

4. The anti-interference method for MEMS IMU sensors based on self-encoders according to claim 3, characterized in that, The encoder has multiple one-dimensional convolutional layers, with the kernel size, stride, and padding configured as needed; the activation function is ReLU; the pooling layers use max pooling or average pooling, with the kernel size and stride configured as needed; the decoder's deconvolutional layer is symmetrical to the encoder's convolutional layer, and the output layer activation function is Sigmoid.

5. The anti-interference method for MEMS IMU sensors based on self-encoders according to claim 1, characterized in that... In step 3, the total loss function for model training is a weighted sum of reconstruction loss, classification loss, and intensity regression loss. The reconstruction loss is calculated based on the mean squared error criterion, the classification loss is calculated based on the cross-entropy criterion, and the intensity regression loss is calculated based on the mean absolute error criterion. The weighting coefficients of each loss term are configured according to a preset ratio.

6. The anti-interference method for MEMS sensors based on self-encoders according to claim 1, characterized in that, The interference existence judgment rule in step 4 is as follows: when the reconstruction error of the model output is greater than a preset threshold, interference is determined to exist; otherwise, it is determined to be without interference; the reconstruction error is the mean square error or cosine distance between the input sample and the reconstructed sample.

7. The anti-interference method for MEMS IMU sensors based on self-encoders according to claim 1, characterized in that, The anti-interference strategies in step 5 include: maintaining the original operating parameters of the sensor when there is no interference; using digital filtering algorithms to suppress interference when there is weak interference; switching the sensor operating frequency band or adjusting the sampling rate when there is medium interference; and starting redundant sensor data fusion or pausing non-critical data output when there is strong interference.

8. A MEMS IMU sensor anti-interference system, characterized in that, It includes a data acquisition module, a data preprocessing module, a model inference module, and a strategy execution module; the data acquisition module is used to acquire real-time output data from MEMS IMU sensors; the data preprocessing module is used to perform the preprocessing operation in step 1 of claim 1; the model inference module has a built-in AE autoencoder model trained by any of the methods described in claims 1-7, and outputs interference-related judgment results; the strategy execution module is used to execute the corresponding anti-interference strategy based on the inference results.

9. The MEMS IMU sensor anti-interference system according to claim 8, characterized in that, The data acquisition module includes a MEMS IMU sensor body and a data cache unit. The cache capacity of the data cache unit meets the continuous input requirements of the sliding window samples. The model inference module is deployed on an embedded processor or FPGA chip, and the inference latency meets the real-time processing requirements.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the anti-interference method for a MEMS IMU sensor based on an autoencoder as described in any one of claims 1-7.

Images