A blind denoising method of MEMS multi-sensor self-contrast learning

Abstract

This invention provides a blind noise reduction method based on self-contrast learning for MEMS multi-sensor systems, belonging to the field of MEMS technology. The method includes receiving observation data from multiple MEMS sensors, constructing a whitening matrix to obtain whitened data; initializing a separation matrix with random unit vectors, solving for the optimal separation vector, and outputting normalized signal and noise sources after orthogonalization; allowing individual observation signals, signal sources, and noise sources to share encoder weights to extract features, which are then input into the decoder to reconstruct the signal; calculating the total loss, and outputting a high-precision denoised signal when the total loss is minimized. This invention employs the aforementioned blind noise reduction method based on self-contrast learning for MEMS multi-sensor systems, solving the dilemma of modeling the non-Gaussianity of interference in unknown environments in traditional methods, and addressing the amplitude uncertainty problem in blind signal separation when there is no prior information about the measured physical quantity.

Description

Technical Field

[0001] This invention relates to the field of microelectromechanical systems (MEMS) technology, and in particular to a blind noise reduction method for MEMS multi-sensor self-comparative learning. Background Technology

[0002] With the widespread application of Micro-Electro-Mechanical Systems (MEMS) sensors in high-end fields such as aerospace flight testing, precision guidance of weapon systems, and intelligent assisted driving, MEMS sensors have become indispensable key components in modern engineering systems due to their significant advantages such as low power consumption, small size, light weight, and easy integration. However, in practical engineering applications, MEMS sensors face complex and unknown environmental interference challenges, such as high and low temperature changes, electromagnetic interference, and high-frequency vibrations. These interferences directly affect the sensors, causing a large amount of noise to be coupled into their observed signals, which seriously affects the stability and high-precision output capability of the sensors.

[0003] Traditionally, to improve the performance of MEMS sensors in interference environments, researchers have proposed various methods, such as using adaptive weighted averaging algorithms to fuse measurements from inertial device arrays, designing feedback weighted fusion algorithms and dynamic sensor bias correction mechanisms, and characterizing unknown disturbances with white noise and designing disturbance decoupling fusion strategies. However, most of these methods are based on a key assumption: the statistical characteristics of unknown disturbance parameters follow certain patterns, and the statistical properties of Gaussian white noise can be used to model environmental interference. But in practical engineering applications, environmental interference is often complex and unknown, with no prior information or specific patterns to follow. This severely limits the method of simply equating unknown disturbances to known zero-mean noise for state estimation, which not only reduces estimation accuracy but may even introduce new error factors, thus restricting the application expansion of MEMS sensors in a wider range of scenarios.

[0004] Therefore, it is particularly important to develop a new method that can achieve high-precision sensing by MEMS sensors under unknown environmental interference. Summary of the Invention

[0005] The purpose of this invention is to provide a blind noise reduction method based on self-comparative learning for MEMS multi-sensor systems, which solves the dilemma of modeling the non-Gaussianity of interference in unknown environments in traditional methods, and addresses the amplitude uncertainty of signal blind separation in the absence of prior information about the physical quantity to be measured.

[0006] To achieve the above objectives, this invention provides a blind noise reduction method for MEMS multi-sensor self-comparative learning, comprising:

[0007] S1. Receive observation data from MEMS multi-sensor flight test that is affected by unknown environmental interference, and reduce the dimensionality of the observation data through principal component analysis to obtain dimensionality-reduced data that retains the core features.

[0008] The dimensionality-reduced data is centered to obtain standardized data. The covariance matrix of the standardized data is calculated and eigenvalue decomposition is performed. The whitening matrix is ​​constructed using the eigenvalue diagonal matrix and the eigenvector matrix, and the whitened data is output.

[0009] S2. Receive the whitening data transmitted by S1 and select a random unit vector as the initial value of each separation vector in the separation matrix;

[0010] Negative entropy is used to measure the non-Gaussianity of the signal, and a Lagrangian function is constructed. The optimal separation vector is solved by Newton's iteration method. After iteration, a normalized separation matrix is ​​obtained, and the separation vector is orthogonalized. The whitened data is projected onto the independent components, and the normalized signal source and noise source are output.

[0011] S3 receives a single observed signal from S1, and the signal source and noise source transmitted from S2. The three share the encoder network weights. After feature extraction processing, they output low-dimensional feature vectors respectively. On the one hand, the similarity between the single observed signal and the signal source and noise source is calculated respectively. On the other hand, the low-dimensional feature vector of the observed signal is input into the decoder. After signal reconstruction processing, the reconstructed signal is output.

[0012] S4 receives the reconstructed signal transmitted by S3 and compares the similarity; it calculates the reconstruction loss and the comparison loss respectively to form the total loss, and iteratively optimizes the network parameters through the backpropagation algorithm and the Adam optimizer; when the total loss is minimized, the reconstructed signal is as close as possible to the real signal and far away from the noise, and outputs a high-precision denoised signal.

[0013] Preferably, the observation data is generated by coupling source signals, and the coupling model is as follows:

[0014] ;

[0015] in, For observational data; , , , which are the observation signals of sensor 1, sensor 2, and sensor n, respectively; It is a real signal source for the physical quantity to be measured; As a source of interference in multi-physics environments, This is the coupling matrix; These are the elements of the coupling matrix;

[0016] The dimensionality reduction data expression is:

[0017] ;

[0018] ;

[0019] ;

[0020] in, For dimensionality reduction data; , These are two-dimensional data representations of 1D and 2D observation data after dimensionality reduction projection, respectively.

[0021] The preferred standardized data expression is:

[0022] ;

[0023] in, To standardize data, , These are the dimensionality-reduced mean values ​​of the 1D and 2D observation data, respectively.

[0024] Construct a whitening matrix and output the whitening data as follows:

[0025] ;

[0026] in, for Covariance matrix; The eigenvector matrix; It is the eigenvalue matrix; Whitened data; For expectation calculation;

[0027] Whitened data satisfies:

[0028] .

[0029] The preferred expression for the non-Gaussianity metric is:

[0030] ;

[0031] in, yes The differential entropy; Is with Homoscedastic Gaussian variables; for Negative entropy; This represents the signal being measured;

[0032] ;

[0033] in, It is a nonlinear function; This represents the expected function.

[0034] Preferably, in S2, the Lagrangian function is constructed to solve for the optimal separation matrix. The expression is:

[0035] ;

[0036] in, For Lagrange multipliers; For separate signals;

[0037] The expression for the optimal separating vector obtained by Newton's iteration method is:

[0038] ;

[0039] in, To update the value, The initial value is determined randomly.

[0040] The normalized expressions for the signal source and noise source are:

[0041] ;

[0042] in, A normalized signal source; The noise source is normalized.

[0043] Preferably, the encoder specifically includes an input layer, three convolutional layers, three activation layers, one pooling layer, one flattening layer, and one normalization layer;

[0044] The expression for the low-dimensional feature vector is:

[0045] ;

[0046] in, , , and , , These represent the kernel weights and biases of the three convolutional layers; ReLU is the activation function of the three activation layers. , , The feature outputs are from three convolutional layers; It is an adaptive average pooling function; It is a flattening function; This is the final low-dimensional feature vector;

[0047] The decoder specifically includes an input layer, a fully connected layer, a deflattening layer, three transposed convolutional layers, and two activation layers;

[0048] The reconstructed signal expression is:

[0049] ;

[0050] in, , , , These are intermediate features in the decoding and reconstruction process. , These are the weight matrix and bias vector of the fully connected layer, respectively. For the sample size, The number of sample channels, For sample length, [] is the transpose operator, [] is the indexing operator, and : represents all indices. For sample channel number, ReLU is the summation operator, and it is the activation function. , , and , , These represent the weights and biases of the three transposed convolutional layers. This is the transpose convolution operator. For reconstructing the signal.

[0051] Preferably, the total loss includes reconstruction loss and contrast loss.

[0052] The reconstruction loss expression is:

[0053] ;

[0054] The contrastive loss expression is:

[0055] ;

[0056] in, For signal source The encoding representation, noise source The encoding representation, Encoded representation of signals observed by any sensor. The similarity between the signal source and the observed signal. The similarity between the noise source and the observed signal. For temperature coefficient, This is the dot product operator. This is the matrix multiplication operator. For logarithmic operators, For exponentiation operator, For summation operators, To compare the losses;

[0057] Total loss expression:

[0058] ;

[0059] in, As a weighting coefficient.

[0060] Therefore, the present invention employs the above-mentioned blind noise reduction method based on self-comparative learning for MEMS multi-sensor systems, and the technical effects are as follows:

[0061] 1. Overcoming the dilemma of non-Gaussian modeling: It can overcome the dilemma of non-Gaussian modeling of unknown environmental interference in traditional signal estimation methods, without the need to make Gaussian assumptions about the interference, thereby improving adaptability and robustness in complex and ever-changing environments.

[0062] 2. Solving the amplitude uncertainty problem of blind signal separation: By designing a positive and negative sample separation method based on blind separation, this method can achieve blind separation of source signals without any prior information, relying solely on observation data, and effectively solves the problem of signal amplitude uncertainty in blind source separation.

[0063] 3. Achieve high-precision signal output: By designing a convolutional autoencoder based on contrastive learning, the similarity between the observed signal and the positive sample (signal component) can be maximized, while the similarity between the observed signal and the negative sample (noise component) can be minimized. This effectively eliminates the influence of unknown environmental interference and achieves high-precision output of MEMS multi-sensor signals.

[0064] 4. Improve sensor stability and expand application scenarios: It can significantly reduce the impact of environmental interference on MEMS sensors, improve the stability and high-precision output capability of sensors in extreme environments, which not only helps the MEMS sensors to deepen their development in existing application fields, but also provides the possibility for their application expansion in a wider range of scenarios. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating a blind noise reduction method for self-comparative learning in MEMS multi-sensor systems according to the present invention.

[0066] Figure 2 This is a comparison image before and after filtering in a blind noise reduction method for self-comparative learning of MEMS multi-sensor according to the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0068] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0069] Example 1

[0070] like Figure 1 As shown, this invention provides a blind noise reduction method for self-comparative learning in MEMS multi-sensor systems, comprising the following steps: After sensing external angular velocity excitation, each sensor in the MEMS multi-sensor system obtains observation data through interface circuits, data acquisition modules, etc. :

[0071] ;

[0072] in, , , , which are the observation signals of sensor 1, sensor 2, and sensor n, respectively.

[0073] Observational data From source signal It is formed by coupling, and the coupling model is:

[0074] ;

[0075] in, It is a real signal source for the physical quantity to be measured. As a source of interference in multi-physics environments, Let be the coupling matrix. The elements of the coupling matrix represent the coupling signals formed when multiple sensors are simultaneously affected to varying degrees by interference sources during their respective signal sensing processes.

[0076] Without any prior information, relying solely on observation data, we analyze independent source signals and their coupling methods to achieve blind separation of source signals and obtain the signal source and interference source components.

[0077] First, Principal Component Analysis (PCA) is used to reduce the dimensionality of the observed data, preserving the main features of the signal while reducing the complexity of data analysis, resulting in dimensionality-reduced data. :

[0078] ;

[0079] ;

[0080] ;

[0081] in, For dimensionality reduction data; , These are two-dimensional data representations of 1D and 2D observation data after dimensionality reduction projection, respectively.

[0082] Secondly, the dimensionality-reduced data is standardized to obtain standardized data. :

[0083] ;

[0084] in, To standardize data, , These are the dimensionality-reduced mean values ​​of the 1D and 2D observation data, respectively.

[0085] right Perform whitening treatment. Calculate. The covariance matrix is ​​calculated, and eigenvalue decomposition is performed on it. Based on the obtained eigenvalue diagonal matrix and eigenvector matrix, a whitening matrix is ​​calculated to eliminate the correlation between the observed signals from each sensor, as detailed below:

[0086] ;

[0087] in, for covariance matrix The eigenvector matrix, For the eigenvalue matrix, For whitening data, For the expected operation, the following conditions must be met:

[0088] .

[0089] The separation matrix is ​​solved using the principle of maximizing non-Gaussianity to separate independent sources, and non-Gaussianity is measured using negative entropy.

[0090] ;

[0091] in, This represents the signal being measured. yes The differential entropy, Is with Homoscedastic Gaussian variables, for The larger the negative entropy, the stronger its non-Gaussianity. An approximate representation of negative entropy is:

[0092] ;

[0093] in For a nonlinear function, choose any one of the following equations:

[0094] ;

[0095] Secondly, find the optimal separation matrix. This makes the separated signal To achieve maximum non-Gaussianity, we need to find the following expression:

[0096] ;

[0097] To prevent the separation matrix from becoming infinitely large, each vector in the separation matrix is ​​constrained to be a unit vector:

[0098]

[0099] Under this constraint, construct the Lagrangian function and solve for it. Optimal solution:

[0100] ;

[0101] Differentiate the above equation:

[0102] ;

[0103] The optimal value can be obtained by taking 0 for the above expression. Using Newton's iteration method to Solving for the problem yields:

[0104] ;

[0105] in, For Lagrange multipliers; For the separation of signals, To update the value, The initial values ​​are randomly determined. To ensure that each vector in the separation matrix is ​​a unit vector, it is necessary to... After normalization, and since the denominator in the above formula is an approximate scalar, the final iterative formula is:

[0106] ;

[0107] Therefore, the separation result of each independent component can be determined based on the separation matrix:

[0108] ;

[0109] The two signals obtained at this time , With source signal , There are significant differences in amplitude, but other characteristics such as waveform remain highly consistent.

[0110] Design a deep convolutional autoencoder based on contrastive learning, comprising an encoder based on a convolutional neural network and a decoder based on a transposed convolutional neural network. This is applied to the source signal obtained from the aforementioned blind separation. , and observation data ,make To compare learning anchors This makes the Gaussian property stronger. negative samples (Noise component), making the Gaussianity weaker Positive samples (signal component).

[0111] First, signal feature extraction is performed. The original input signal is encoded into its respective features by the decoder. The encoder specifically includes an input layer, three convolutional layers, three activation layers, one pooling layer, one flattening layer, and one normalization layer. The input layer receives the original data, each convolutional layer extracts signal features at multiple levels while simultaneously downsampling, each activation layer introduces nonlinearity to increase the model's expressive power, the pooling layer performs global information compression while preserving channel-wide features, the flattening layer removes redundant dimensions, and the normalization layer constrains the features to a unit sphere. The specific calculation process is shown below:

[0112] ;

[0113] in, , , and , , These represent the kernel weights and biases of the three convolutional layers, and ReLU is the activation function of the three activation layers. , , For the feature input of the three convolutional layers, It is an adaptive average pooling function. For flattening functions, This is the final low-dimensional feature vector. , , The weights of each layer of the shared encoder are used to obtain their respective low-dimensional feature vectors after passing through the encoder. , , .

[0114] Next, signal reconstruction is performed. The encoded feature vector of the anchor point is reconstructed by the decoder, which specifically includes an input layer, a fully connected layer, a deflattening layer, three transposed convolutional layers, and two activation layers. The input layer receives a low-dimensional vector as the starting point for reconstruction. The fully connected layer maps the low-dimensional vector to a high-dimensional space, providing data for the transposed convolution. The deflattening layer reshapes the output of the fully connected layer into the format required for convolution. Each transposed convolutional layer performs upsampling and channel restoration. Each activation layer introduces nonlinearity to enhance feature reconstruction capabilities and avoid gradient vanishing. The specific calculation process is shown below:

[0115] ;

[0116] in, , , , These are intermediate features in the decoding and reconstruction process. , These are the weight matrix and bias vector of the fully connected layer, respectively. For the sample size, The number of sample channels, For sample length, [] is the transpose operator, [] is the indexing operator, and : represents all indices. For sample channel number, ReLU is the summation operator, and it is the activation function. , , and , , These represent the weights and biases of the three transposed convolutional layers. This is the transpose convolution operator. For reconstructing the signal.

[0117] Subsequently, the loss of the contrastive convolutional autoencoder is calculated. The loss consists of two parts: reconstruction loss and contrastive loss.

[0118] The mean squared error between the reconstructed value and the observed value is used as the reconstruction loss function. The smaller the loss value, the smaller the difference between the reconstructed value and the anchor value, indicating that the reconstruction is approximately close to the original value. Calculate the reconstruction loss:

[0119] ;

[0120] A contrastive loss function is constructed by comparing the similarity between positive samples and anchor points, and between negative samples and anchor points. A smaller loss value indicates a higher similarity between the anchor point and a positive sample, and a lower similarity between the anchor point and a negative sample. The contrastive loss is calculated as follows:

[0121] ;

[0122] in, For signal source The encoding representation, noise source The encoding representation, Encoded representation of signals observed by any sensor. The similarity between the signal source and the observed signal. The similarity between the noise source and the observed signal. For temperature coefficient, This is the dot product operator. This is the matrix multiplication operator. For logarithmic operators, For exponentiation operator, For summation operators, To compare the losses;

[0123] Finally, by combining the reconstruction loss and the contrastive loss, the loss of the convolutional autoencoder based on contrastive learning is calculated:

[0124] ;

[0125] in, As a weighting coefficient, it is used to balance the contributions of the two types of loss. When the autoencoder loss reaches its minimum, the reconstructed signal... By closely mimicking real signals and minimizing noise interference, high-precision output from MEMS multi-sensor systems with strong anti-interference capabilities is achieved. .

[0126] Finally, following the chain rule, through the above steps, the weights, biases, and other parameters of each stage of the convolutional neural network and the transposed convolutional neural network are continuously optimized using the backpropagation algorithm and the Adam optimizer to update the network parameters and minimize the loss of the autoencoder. On the one hand, this maximizes the reconstruction of the amplitude features of the original observed signal; on the other hand, by using a positive-negative sample contrast loss function, the anchor points are made as far away as possible from negative samples (noise components) and as close as possible to positive samples (signal features highly consistent with the real signal), thus achieving blind noise reduction of MEMS multi-sensor observation signals.

[0127] Figure 2 The diagram shows a signal comparison before and after applying the blind denoising method based on self-contradictory learning for MEMS multi-sensor proposed in this invention. The blue portion represents the original observation signal without denoising processing. These signals are acquired by MEMS sensors under complex and unknown environmental interference, thus containing a large amount of noise, resulting in blurred signal waveforms and making it difficult to accurately identify the true source characteristics of the measured physical quantity. In this situation, the stability and high-precision output capability of the sensor are severely affected.

[0128] The red portion shows the signal processed by the blind denoising method of this invention. This method first performs dimensionality reduction and standardization on the observed data using principal component analysis, constructing a whitening matrix to eliminate correlations between signals observed from different sensors. Subsequently, negative entropy is used to measure the non-Gaussianity of the signal, and the optimal separation vector is solved using Newton's iteration method, achieving effective separation of the signal source and noise source. Furthermore, a deep convolutional autoencoder based on contrastive learning is designed to maximize the reconstructed signal's approximation to the true signal components while minimizing the influence of noise components. Specifically, the encoder extracts signal features through multi-layer convolution and activation operations, while the decoder reconstructs the signal through transposed convolution.

[0129] from Figure 2 It is evident that the signal waveform after noise reduction is clearer, with a significant reduction in noise components, thus highlighting the true signal characteristics. This change demonstrates that the method of this invention successfully solves the dilemma of modeling the non-Gaussianity of interference in unknown environments in traditional methods, as well as the amplitude uncertainty problem of blind signal separation in the absence of prior information about the measured physical quantity. Ultimately, it achieves high-precision signal output from MEMS multi-sensors in unknown interference environments, significantly improving sensor stability and the scalability of application scenarios.

[0130] Therefore, this invention adopts the above-mentioned blind noise reduction method of self-contrast learning for MEMS multi-sensor. It proposes for the first time that, under the condition of no prior information, the signal source and noise source extracted by blind source separation are used as positive and negative samples for contrast learning. The similarity between the observed signal and the positive sample is enhanced and the similarity with the negative sample is reduced by contrast learning autoencoder. This solves the technical problem of signal amplitude uncertainty in blind source separation, reduces the influence of interference on MEMS sensors, and realizes high-precision output of MEMS sensors under unknown interference.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A blind noise reduction method for MEMS multi-sensor self-comparative learning, characterized in that, include: S1. Receive observation data from MEMS multi-sensor flight test that is affected by unknown environmental interference, and reduce the dimensionality of the observation data through principal component analysis to obtain dimensionality-reduced data that retains the core features. In a MEMS multi-sensor system, each sensor, after sensing external angular velocity excitation, obtains observation data through interface circuits and a data acquisition module. : ； in, , , , which are the observation signals of sensor 1, sensor 2, and sensor n, respectively; Observational data From source signal It is formed by coupling, and the coupling model is: ； in, It is a real signal source for the physical quantity to be measured. As a source of interference in multi-physics environments, Let be the coupling matrix. The elements of the coupling matrix represent the coupling signals formed when multiple sensors are simultaneously affected to varying degrees by interference sources during their respective signal sensing processes. The dimensionality-reduced data is centered to obtain standardized data. The covariance matrix of the standardized data is calculated and eigenvalue decomposition is performed. The whitening matrix is ​​constructed using the eigenvalue diagonal matrix and the eigenvector matrix, and the whitened data is output. S2. Receive the whitening data transmitted by S1 and select a random unit vector as the initial value of each separation vector in the separation matrix; Negative entropy is used to measure the non-Gaussianity of the signal, and a Lagrangian function is constructed. The optimal separation vector is solved by Newton's iteration method. After iteration, a normalized separation matrix is ​​obtained, and the separation vector is orthogonalized. The whitened data is projected onto the independent components, and the normalized signal source and noise source are output. S3 receives a single observed signal from S1, and the signal source and noise source transmitted from S2. The three share the encoder network weights. After feature extraction processing, they output low-dimensional feature vectors respectively. On the one hand, the similarity between the single observed signal and the signal source and noise source is calculated respectively. On the other hand, the low-dimensional feature vector of the observed signal is input into the decoder. After signal reconstruction processing, the reconstructed signal is output. S4 receives the reconstructed signal transmitted by S3 and compares the similarity; it calculates the reconstruction loss and the comparison loss respectively to form the total loss, and iteratively optimizes the network parameters through the backpropagation algorithm and the Adam optimizer; when the total loss is minimized, the reconstructed signal is as close as possible to the real signal and far away from the noise, and outputs a high-precision denoised signal.

2. The blind noise reduction method for MEMS multi-sensor self-comparative learning according to claim 1, characterized in that, The dimensionality reduction data expression is: ； ； ； in, For dimensionality reduction data; , These are two-dimensional data representations of 1D and 2D observation data after dimensionality reduction projection, respectively.

3. The blind noise reduction method for MEMS multi-sensor self-comparative learning according to claim 2, characterized in that, The standardized data expression is: ； in, To standardize data, , These are the dimensionality-reduced mean values ​​of the 1D and 2D observation data, respectively. Construct a whitening matrix and output the whitening data as follows: ； in, for Covariance matrix; The eigenvector matrix; It is the eigenvalue matrix; Whitened data; For expectation calculation; Whitened data satisfies: 。 4. The blind noise reduction method for MEMS multi-sensor self-comparative learning according to claim 3, characterized in that, The expression for the non-Gaussianity metric is: ； in, yes The differential entropy; Is with Homoscedastic Gaussian variables; for Negative entropy; This represents the signal being measured; ； in, It is a nonlinear function; This represents the expected function.

5. The blind noise reduction method for self-comparative learning of MEMS multi-sensor according to claim 4, characterized in that, In S2, construct the Lagrange function and solve for the optimal separation matrix. The expression is: ； in, For Lagrange multipliers; For separate signals; The expression for the optimal separating vector obtained by Newton's iteration method is: ； in, To update the value, The initial value is determined randomly. The normalized expressions for the signal source and noise source are: ； in, A normalized signal source; The noise source is normalized.

6. The blind noise reduction method for self-comparative learning of MEMS multi-sensor according to claim 5, characterized in that, The encoder specifically includes an input layer, three convolutional layers, three activation layers, one pooling layer, one flattening layer, and one normalization layer; The expression for the low-dimensional feature vector is: ； in, , , and , , These represent the kernel weights and biases of the three convolutional layers; ReLU is the activation function of the three activation layers. , , The feature outputs are from three convolutional layers; It is an adaptive average pooling function; It is a flattening function; This is the final low-dimensional feature vector; The decoder specifically includes an input layer, a fully connected layer, a deflattening layer, three transposed convolutional layers, and two activation layers; The reconstructed signal expression is: ； in, , , , These are intermediate features in the decoding and reconstruction process. , These are the weight matrix and bias vector of the fully connected layer, respectively. For the sample size, The number of sample channels, For sample length, [] is the transpose operator, [] is the indexing operator, and : represents all indices. For sample channel number, ReLU is the summation operator, and it is the activation function. , , and , , These represent the weights and biases of the three transposed convolutional layers. This is the transpose convolution operator. For reconstructing the signal.

7. The blind noise reduction method for MEMS multi-sensor self-comparative learning according to claim 6, characterized in that, The total loss includes reconstruction loss and contrast loss. The reconstruction loss expression is: ； The contrastive loss expression is: ； in, For signal source The encoding representation, noise source The encoding representation, Encoded representation of signals observed by any sensor. The similarity between the signal source and the observed signal. The similarity between the noise source and the observed signal. For temperature coefficient, This is the dot product operator. This is the matrix multiplication operator. For logarithmic operators, For exponentiation operator, For summation operators, To compare the losses; Total loss expression: ； in, As a weighting coefficient.

Images