MEMS inertial sensor reliability analysis method based on long and short term memory deep learning

Abstract

The invention discloses an MEMS inertial sensor reliability analysis method based on long and short term memory deep learning, and belongs to the technical field of microelectronic reliability evaluation, and the method comprises the steps: collecting an original output signal and environmental parameter data of an MEMS inertial sensor in a target scene, and recording a collection timestamp; then preprocessing is carried out to obtain a time sequence processing data set; performing feature extraction on the time sequence processing data set to form a multi-dimensional feature vector; carrying out correlation analysis on the multi-dimensional feature vectors, and screening out a key feature set; constructing an LSTM model based on the key feature set, and performing training and parameter optimization on the model to enable the model to capture a time sequence dependency relationship of sensor performance degradation; inputting the real-time key feature set subjected to preprocessing and feature extraction into the trained LSTM model, and outputting probability distribution of the current health state of the sensor; and then reliability evaluation is carried out and an early warning signal is output through the system terminal to prompt maintenance personnel to carry out timely intervention.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of microelectronic reliability evaluation, in particular to a MEMS inertial sensor reliability analysis method based on long short-term memory deep learning. BACKGROUND

[0002] MEMS (Micro Electro Mechanical System) inertial sensor is a kind of miniaturized sensor based on micro-nano technology research and development, which is widely used in various dynamic measurement scenes due to its small size, low cost and low power consumption.

[0003] But the stability and reliability of its working state directly determine the running safety and task execution accuracy of the terminal device, especially in complex scenes such as extreme environment and long-time continuous work, the sensor is easy to be affected by environmental factors such as temperature, vibration and humidity, and performance degradation occurs, and even causes failure, and the traditional reliability analysis method depends on the threshold set by artificial or simple statistical model, which cannot capture the timing correlation and nonlinear characteristics in the performance degradation process of the sensor, and has problems of early warning lag and high misjudgment rate, which cannot meet the real-time monitoring and accurate evaluation demand of sensor reliability in complex scenes.

[0004] Based on this, the present application provides a MEMS inertial sensor reliability analysis method based on long short-term memory deep learning, which can effectively capture the timing dependence, accurately extract the key features, and realize the real-time monitoring and reliability prediction of the sensor running state. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art and provide a MEMS inertial sensor reliability analysis method based on long short-term memory deep learning to solve the problems raised in the background.

[0006] The purpose of the present application can be achieved by the following technical scheme: a MEMS inertial sensor reliability analysis method based on long short-term memory deep learning, comprising:

[0007] Step one: collect the original output signal of the MEMS inertial sensor in the target scene and the corresponding environmental parameter data, and record the collection timestamp; then pre-process the collected original data to obtain a timing processing data set;

[0008] Step two: feature extraction is performed on the timing processing data set to form a multi-dimensional feature vector; then correlation analysis is performed on the multi-dimensional feature vector to select a key feature set sensitive to the reliability of the sensor;

[0009] Step three: based on the key feature set, a long short-term memory deep learning model is constructed, and the model is trained and parameter optimized to enable the model to capture the timing dependence of sensor performance degradation;

[0010] Step 4: Input the preprocessed and feature-extracted real-time key feature set into the trained LSTM model, and output the probability distribution of the current health status of the sensor; perform reliability assessment based on the probability distribution output by the model, including status determination, degradation trend prediction and fault warning, and output warning signals through the system terminal to prompt maintenance personnel to intervene in a timely manner.

[0011] Preferably, the collected raw data is preprocessed, including:

[0012] Wavelet transform or variational mode decomposition is used to separate high-frequency noise and low-frequency drift in the original output signal, and moving average filtering is combined to remove sudden interference.

[0013] The 3σ criterion is used to remove outliers from the original data, such as acceleration jumps caused by electromagnetic interference; and linear interpolation is used to fill in missing values.

[0014] Min-Max normalization was used to map all cleaned data to the [0,1] interval to eliminate the influence of units between data.

[0015] The standardized data are aligned according to their corresponding collection timestamps to obtain a time-series processed dataset.

[0016] Preferably, the process of feature extraction from the time-series processing dataset includes:

[0017] The raw output signals of the MEMS inertial sensor are obtained by traversing the time-series processing dataset; the time-series processing dataset is segmented using a sliding window of preset length to obtain several sliding window data.

[0018] Based on historical fault records, a corresponding sensor health status label is bound to each sliding window; where the health status label is Y = {y1, y2, y3}, y1 = normal, y2 = degraded, and y3 = faulty;

[0019] For each sliding window of data, calculate the time-domain and frequency-domain characteristics of the original output signal; by calculating the mean, variance, peak value, and root mean square of the original output signal within the sliding window, obtain the time-domain characteristics of the original output signal.

[0020] The frequency domain amplitude spectrum is obtained by performing a fast Fourier transform on the original output signal within each sliding window; based on the frequency domain amplitude spectrum, the spectral peak, peak frequency energy ratio, and spectral centroid are extracted to obtain the frequency domain characteristics of the original output signal.

[0021] For environmental parameter data, calculate the mean, fluctuation range, and rate of change of each parameter within the sliding window, such as the temperature change per minute; thus obtaining the characteristics of the environmental parameters.

[0022] A multidimensional feature vector is formed based on time-domain features, frequency-domain features, and environmental parameter features.

[0023] Preferably, the method for obtaining the key feature set is as follows:

[0024] Based on the multidimensional feature vector, obtain the corresponding sensor health status label bound to each sliding window; calculate the mutual information between each type feature X and the health status label Y in the multidimensional feature vector;

[0025] The mutual information between each type of feature and the health status label is obtained. If the calculated mutual information is greater than or equal to a preset threshold, it indicates that the type of feature is significantly related to the sensor's health status, and the type of feature is marked as a key feature; otherwise, the type of feature is removed.

[0026] After statistical screening, a set of key features sensitive to sensor reliability is obtained.

[0027] Preferably, the expression for the mutual information is:

[0028]

[0029] In the formula, MI(X,Y) represents mutual information; P(X=xi,Y=yj) represents the joint probability of feature X=xi and health status label Y=yj; i=1,2,...,k, k is the total number of sliding windows; j=1,2,3; xi is the multidimensional feature vector of the i-th sliding window, P(X=xi) represents the marginal probability of feature X=xi; P(Y=yj) represents the marginal probability of Y=yj.

[0030] Preferably, the LSTM model construction and training steps include:

[0031] S31: Divide the key feature set obtained in step two into training set, validation set and test set in chronological order;

[0032] S32: Construct the LSTM model structure, which includes an input layer, a hidden layer, a fully connected layer, and an output layer;

[0033] S33: Sparse classification cross-entropy loss is used as the loss function, and the Adam optimizer is used to accelerate convergence by adaptively adjusting the learning rate.

[0034] S34: The model is iteratively trained using the training set. After each training round, the loss value and classification accuracy are calculated using the validation set. An early stopping mechanism is introduced: when the loss value on the validation set does not decrease for 10 consecutive rounds, training is stopped to avoid model overfitting.

[0035] S35: The hyperparameters are optimized using a grid search method. The hyperparameters include: number of LSTM units, dropout rate, sequence length, and learning rate. The optimal parameter combination is selected when the validation set accuracy is greater than or equal to the preset standard, and the corresponding LSTM model is obtained.

[0036] Preferably, S32 further includes:

[0037] Input layer: The input dimension is the number of features in the key feature set. Each input sample is a time sequence containing k consecutive sliding window key features.

[0038] Hidden layers: Multiple LSTM units are set up to enhance the model's ability to capture long-term dependencies by stacking them; a Dropout layer is added after each LSTM layer with a dropout rate of 0.2-0.3 to randomly drop some neurons to prevent overfitting;

[0039] Fully connected layer: The output of the last LSTM layer is flattened through a Flatten layer and then connected to a fully connected layer. The ReLU function is used as the activation function to enhance the non-linear fitting ability of the model.

[0040] Output layer: Based on the health status labels of 3 categories, the output layer is set with 3 neurons, and the activation function is softmax. The output layer outputs the probability distribution corresponding to each health status.

[0041] Preferably, the warning signals include a first-level warning signal, a second-level warning signal, and a third-level warning signal.

[0042] Preferably, the specific implementation process of step four is as follows:

[0043] S41: Real-time acquisition of raw output signals and environmental parameter data from MEMS inertial sensors, recording the acquisition timestamp, and performing the same preprocessing and feature extraction process as in steps one and two on the real-time data to obtain a real-time key feature set;

[0044] S42: Organize the real-time key feature set into input samples according to the time series and input them into the trained LSTM model; the model output is the health state probability distribution [Y1,Y2,Y3] corresponding to the current sliding window; where Y1 is the normal state probability, Y2 is the degraded state probability, Y3 is the fault state probability, and Y1+Y2+Y3=1;

[0045] S43: Based on the probability distribution output by the model, perform multi-dimensional reliability assessment and output early warning signals through the system terminal to prompt maintenance personnel to intervene in a timely manner.

[0046] Preferably, S43 further includes:

[0047] State determination: Select the state with the highest probability as the current sensor health state;

[0048] Degradation trend prediction: When a degradation state is determined, the probability of degradation state is predicted by combining historical data of degradation state output and linear fitting, and the remaining lifetime (RUL) from the current state to the failure state is estimated.

[0049] The formula for calculating the remaining useful life (RUL) is as follows:

[0050]

[0051] In the formula, b is the proportionality coefficient, Y2 is the current degradation probability, and ΔY2 is the average increase in degradation probability per unit time.

[0052] Fault warning: Based on the current sensor health status, a multi-level warning mechanism is set up;

[0053] If the output is in a normal state, the trend of degradation probability will be continuously monitored.

[0054] If the output is in a degraded state, set a graded warning based on the remaining lifetime (RUL):

[0055] When RUL>T1, a first-level warning signal is output to strengthen monitoring;

[0056] When RUL≤T1, a level 2 warning signal is output, indicating that the remaining lifespan is close to the safety threshold and it is recommended to prepare maintenance resources.

[0057] When RUL≤T2, a level 3 warning signal is output, indicating that the remaining lifespan is insufficient and the sensor is about to enter a fault state. It is recommended to arrange an immediate shutdown for maintenance.

[0058] If the output indicates a fault, a level 3 warning signal will be output and an audible and visual alarm will be issued.

[0059] Compared to existing solutions, the beneficial effects achieved by this invention are:

[0060] This invention is based on the comprehensive extraction of time and frequency domain features, combined with the correlation analysis of environmental parameters and sensor performance parameters, to screen out key feature sets that are sensitive to reliability, effectively eliminate redundant information, focus on core influencing factors, avoid evaluation distortion caused by information overload or missing key information in feature engineering, and improve the targeting and accuracy of reliability analysis.

[0061] This invention employs an LSTM deep learning model, which, through training with historical fault data and parameter optimization, enables the model to accurately capture the dynamic temporal dependencies of sensor performance degradation, thereby achieving a refined characterization of the sensor performance degradation process, significantly improving the reliability of health scores, and providing a reliable basis for preventive maintenance.

[0062] This invention outputs a health status probability distribution through an LSTM model, which can simultaneously determine the current status, quantitatively predict future degradation trends, and provide early warnings of faults. The timely warning signals output by the system terminal guide maintenance personnel to intervene in advance, avoiding system downtime or risk accidents caused by sensor failures, and reducing maintenance costs and losses. Attached Figure Description

[0063] The invention will now be further described with reference to the accompanying drawings.

[0064] Figure 1 This is a flowchart of the reliability analysis method for MEMS inertial sensors based on long short-term memory deep learning proposed in this invention. Detailed Implementation

[0065] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0066] like Figure 1 As shown, this invention is a reliability analysis method for MEMS inertial sensors based on long short-term memory deep learning, comprising the following steps:

[0067] Step 1: Collect the raw output signals of the MEMS inertial sensor and the corresponding environmental parameter data in the target scene, and record the collection timestamp; then preprocess the collected raw data to obtain the time-series processing dataset.

[0068] The original output signal includes, but is not limited to, three-dimensional acceleration and three-dimensional angular velocity; the environmental parameter data includes ambient temperature, humidity, and power supply voltage; in this embodiment, the target scenario refers to a specific application scenario in which the MEMS inertial sensor is actually deployed and performs core tasks such as motion state detection and attitude perception. This scenario must have the motion excitation and environmental parameter fluctuations required for the sensor to work, and can reflect the actual working conditions of the sensor.

[0069] It should be further explained that the collected raw data is preprocessed, including signal denoising, data cleaning, standardization, and time-series alignment.

[0070] Signal denoising: Wavelet transform or variational mode decomposition is used to separate high-frequency noise and low-frequency drift in the original output signal, and moving average filtering is combined to remove sudden interference.

[0071] Data cleaning: The 3σ criterion is used to remove outliers from the original data, such as acceleration jumps caused by electromagnetic interference; and linear interpolation is used to fill in missing values.

[0072] Standardization: Min-Max normalization is used to map all cleaned data to the [0,1] interval to eliminate the influence of units between data.

[0073] Time-series alignment: The standardized data are aligned according to their corresponding collection timestamps to obtain a time-series processed dataset;

[0074] In this embodiment of the invention, the time-series processing dataset obtained by using timestamps as indexes includes at least multi-dimensional feature values ​​such as three-dimensional acceleration, three-dimensional angular velocity and environmental parameters of MEMS inertial sensors, and all feature values ​​have been mapped to the [0,1] interval, eliminating the interference of dimensional differences on subsequent model training.

[0075] Step 2: Extract features from the time-series processing dataset to form a multi-dimensional feature vector; then perform correlation analysis on the multi-dimensional feature vector to screen out the key feature set that is sensitive to sensor reliability;

[0076] It should be further explained that the feature extraction process for time-series processing datasets includes:

[0077] The raw output signals of the MEMS inertial sensor are obtained by traversing the time-series processing dataset.

[0078] The time-series processing dataset is segmented using a preset-length sliding window to obtain several sliding window data. The window size of the preset-length sliding window is set according to the sensor sampling frequency and dynamic response characteristics. For example, when the sampling frequency is 1000Hz, the window length is set to 1000 data points, corresponding to 1 second; the step size is set to 500 data points, i.e., the overlap rate is 50%, balancing the timeliness of features and computational efficiency.

[0079] Based on historical fault records, a corresponding sensor health status label is bound to each sliding window; where the health status label is Y = {y1, y2, y3}, y1 = normal, y2 = degraded, and y3 = faulty;

[0080] For each sliding window of data, calculate the time-domain and frequency-domain characteristics of the original output signal;

[0081] The time-domain characteristics of the original output signal are obtained by calculating the mean, variance, peak value, and root mean square of the original output signal within the sliding window. Furthermore, the mean is used to reflect the overall offset trend of the original output signal; the variance is used to reflect the dispersion of the original output signal; the peak value is used to reflect the impact intensity of the original output signal; and the root mean square is used to reflect the energy level of the original output signal.

[0082] By performing a Fast Fourier Transform on the original output signal within each sliding window, the frequency domain amplitude spectrum is obtained. Based on the frequency domain amplitude spectrum, the peak frequency, peak frequency energy ratio, and spectral centroid are extracted to obtain the frequency domain characteristics of the original output signal. Furthermore, the peak frequency is the frequency corresponding to the maximum peak value extracted from the frequency domain amplitude spectrum, reflecting the main vibration or noise frequency components of the original output signal; the peak frequency energy ratio is the proportion of energy at the peak frequency to the total frequency domain energy, reflecting the energy contribution of the dominant frequency; and the spectral centroid is used to reflect the frequency location where energy is concentrated.

[0083] For environmental parameter data, calculate the mean, fluctuation range, and rate of change of each parameter within the sliding window, such as the temperature change per minute; thus obtaining the characteristics of the environmental parameters.

[0084] A multidimensional feature vector is formed based on time-domain features, frequency-domain features, and environmental parameter features.

[0085] It should be further explained that the method for obtaining the key feature set is as follows:

[0086] Based on the multidimensional feature vector, obtain the corresponding sensor health status label bound to each sliding window; calculate the mutual information between each type feature X and the health status label Y in the multidimensional feature vector;

[0087] The expression for mutual information is:

[0088]

[0089] In the formula, MI(X,Y) represents mutual information, used to measure the non-linear dependency between any type of feature and the health status label; P(X=xi,Y=yj) represents the joint probability of feature X=xi and health status label Y=yj; i=1,2,...,k, k is the total number of sliding windows; j=1,2,3; xi is the multi-dimensional feature vector of the i-th sample, P(X=xi) represents the marginal probability of feature X=xi, that is, the proportion of samples with feature xi; P(Y=yj) represents the marginal probability of Y=yj, that is, the proportion of samples with health status label yj;

[0090] The mutual information between each type of feature and the health status label is obtained. If the calculated mutual information is greater than or equal to a preset threshold, it indicates that the type of feature is significantly related to the sensor's health status, and the type of feature is marked as a key feature; otherwise, the type of feature is removed.

[0091] After statistical screening, a set of key features sensitive to sensor reliability is obtained;

[0092] In this embodiment of the invention, features that can effectively reflect the performance degradation of MEMS inertial sensors are extracted from the original signals, and a set of key features that are sensitive to sensor reliability is obtained through correlation analysis and screening, laying the foundation for subsequent reliability assessment.

[0093] Step 3: Based on the key feature set, construct a Long Short-Term Memory (LSTM) deep learning model, and through training and parameter optimization, enable the model to capture the temporal dependencies of sensor performance degradation;

[0094] It should be further explained that the LSTM model construction and training steps include:

[0095] S31: Divide the key feature set obtained in step two into training set, validation set and test set according to time order; further, since the sensor performance degradation has temporal continuity, it is necessary to maintain the temporal relationship of the data to avoid information leakage caused by random partitioning; for example, select the first 70% of the sliding window data as training set, the middle 10% as validation set and the last 20% as test set to simulate the model's prediction of future states in actual applications;

[0096] S32: Construct the LSTM model structure, which includes an input layer, a hidden layer, a fully connected layer, and an output layer;

[0097] Input layer: The input dimension is the number of features in the key feature set (let's say m), and each input sample is a time sequence containing k consecutive sliding window key features;

[0098] Hidden layers: Multiple LSTM units are set up to enhance the model's ability to capture long-term dependencies by stacking them; a Dropout layer is added after each LSTM layer with a dropout rate of 0.2-0.3 to randomly drop some neurons to prevent overfitting;

[0099] Fully connected layer: The output of the last LSTM layer is flattened through a Flatten layer and then connected to a fully connected layer. The activation function is ReLU to enhance the model's non-linear fitting ability.

[0100] Output layer: Based on the health status labels of 3 categories (normal, deterioration, and failure), the output layer is set with 3 neurons, and the activation function is softmax. The output layer outputs the probability distribution corresponding to each health status.

[0101] S33: Sparse classification cross-entropy loss is used as the loss function, and the Adam optimizer is used to accelerate convergence by adaptively adjusting the learning rate.

[0102] S34: Iteratively train the model using the training set, and calculate the loss value and classification accuracy using the validation set after each training round; furthermore, introduce an early stopping mechanism, and stop training when the loss value of the validation set has not decreased for 10 consecutive rounds to avoid model overfitting;

[0103] S35: The hyperparameters are optimized using a grid search method. The hyperparameters include: number of LSTM units, Dropout rate, sequence length, and learning rate. The optimal parameter combination is selected when the validation set accuracy is greater than or equal to the preset standard, and the corresponding LSTM model is obtained.

[0104] In this embodiment of the invention, by constructing an LSTM model, the temporal information in the key feature set can be fully utilized to accurately characterize the performance evolution process of MEMS inertial sensors from normal to degradation and then to failure, providing a model basis for subsequent reliability assessment.

[0105] Step 4: Input the preprocessed and feature-extracted real-time key feature set into the trained LSTM model, and output the probability distribution of the current health status of the sensor; perform reliability assessment based on the probability distribution output by the model, including status determination, degradation trend prediction and fault warning, and output warning signals through the system terminal to prompt maintenance personnel to intervene in a timely manner.

[0106] It should be further explained that the specific implementation process of step four is as follows:

[0107] S41: Real-time acquisition of raw output signals and environmental parameter data from MEMS inertial sensors, recording the acquisition timestamp, and performing preprocessing and feature extraction processes on the real-time data that are completely consistent with steps one and two to obtain a real-time key feature set.

[0108] S42: Organize the real-time key feature set into input samples according to the time series and input them into the trained LSTM model; the model output is the health state probability distribution [Y1,Y2,Y3] corresponding to the current sliding window; where Y1 is the normal state probability, Y2 is the degraded state probability, Y3 is the fault state probability, and Y1+Y2+Y3=1;

[0109] S43: Based on the probability distribution output by the model, perform multi-dimensional reliability assessment, including:

[0110] State determination: Select the state with the highest probability as the current sensor health state;

[0111] Degradation trend prediction: When a degradation state is determined, the probability of degradation state is predicted by combining historical data of degradation state output and linear fitting, and the remaining lifetime (RUL) from the current state to the failure state is estimated.

[0112] The formula for calculating the remaining useful life (RUL) is as follows:

[0113]

[0114] In the formula, b is the proportional coefficient, and the specific value is set by experts in the field based on historical fault data. Y2 is the current degradation probability, and ΔY2 is the average increase of the degradation probability per unit time. The specific value is calculated based on the time series data of the sensor's historical degradation process and is used to reflect the average rate of change of the degradation probability over time.

[0115] Fault warning: Based on the current sensor health status, a multi-level warning mechanism is set up;

[0116] If the output is in a normal state, the trend of degradation probability will be continuously monitored.

[0117] If the output is in a degraded state, set a graded warning based on the remaining lifetime (RUL):

[0118] When RUL>T1, a first-level warning signal is output to strengthen monitoring;

[0119] When RUL≤T1, a level 2 warning signal is output, indicating that the remaining lifespan is close to the safety threshold and it is recommended to prepare maintenance resources.

[0120] When RUL≤T2, a level 3 warning signal is output, indicating that the remaining lifespan is insufficient and the sensor is about to enter a fault state. It is recommended to arrange an immediate shutdown for maintenance.

[0121] It should be noted that T1 and T2 are the safety threshold and maintenance threshold, respectively. The specific values ​​should be calibrated based on the task reliability requirements of the target scenario (such as continuous running time and fault tolerance) and historical fault repair data to ensure the timeliness and practicality of the warning.

[0122] If the output is in a fault state, a level three warning signal will be output and an audible and visual alarm will be issued;

[0123] In this embodiment of the invention, by combining real-time data with models, dynamic assessment and early warning of the reliability of MEMS inertial sensors are realized. This enables timely capture of performance degradation trends, provides data support for equipment maintenance decisions, and ultimately improves the stable operation capability of sensors in target scenarios.

[0124] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0125] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0127] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A MEMS inertial sensor reliability analysis method based on long short-term memory deep learning, characterized in that, The method comprises the following steps: Step 1: Collecting the original output signal of the MEMS inertial sensor in the target scene and the corresponding environmental parameter data, and recording the collection timestamp; Then, the collected original data is preprocessed to obtain a time series processing dataset; Step 2: Feature extraction is performed on the time series processing dataset to form a multi-dimensional feature vector; Then, correlation analysis is performed on the multi-dimensional feature vector to screen out a key feature set sensitive to sensor reliability; Step 3: Based on the key feature set, a long short-term memory deep learning model is constructed, and the model is trained and parameter-optimized to enable the model to capture the time series dependence of sensor performance degradation; Step 4: The real-time key feature set obtained through preprocessing and feature extraction is input into the trained LSTM model, and the probability distribution of the current health state of the sensor is output; based on the probability distribution output by the model, reliability evaluation is performed, including state judgment, degradation trend prediction and fault warning, and a warning signal is output through the system terminal to prompt the maintenance personnel to intervene in time.

2. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 1, wherein, The preprocessing of the collected original data includes: Wavelet transform or variational mode decomposition is used to separate high-frequency noise and low-frequency drift in the original output signal, and sliding average filtering is used to remove sudden interference; The 3σ rule is used to remove outliers in the original data, such as acceleration jump data caused by electromagnetic interference; and linear interpolation is used to complete the missing values; Min-Max normalization is used to map all cleaned data to the [0, 1] interval to eliminate the dimensional influence between data; The standardized data is aligned according to the corresponding collection timestamp to obtain a time series processing dataset.

3. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 2, characterized in that, The feature extraction process of the time series processing dataset includes: Traverse the time series processing dataset to obtain the original output signal of the MEMS inertial sensor; use a pre-set length sliding window to segment the time series processing dataset to obtain a plurality of sliding window data; Based on the historical fault record, a corresponding sensor health state label is bound for each sliding window; wherein the health state label is Y={y1, y2, y3}, y1=normal, y2=degradation, and y3=fault; For each sliding window data, the time domain features and frequency domain features of the original output signal are calculated; the mean, variance, peak value and root mean square of the original output signal in the sliding window are calculated to obtain the time domain features of the original output signal; The frequency domain amplitude spectrum is obtained by performing fast Fourier transform on the original output signal in each sliding window; based on the frequency domain amplitude spectrum, the spectral peak value, peak frequency energy proportion and spectral barycenter are extracted to obtain the frequency domain features of the original output signal; For the environmental parameter data, the mean, fluctuation amplitude and change rate of each parameter in the sliding window are calculated, such as the temperature change per minute; the environmental parameter features are obtained; Based on the time domain features, frequency domain features and environmental parameter features, a multi-dimensional feature vector is formed.

4. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 3, characterized in that, The method for obtaining the key feature set is as follows: Based on the multi-dimensional feature vector, the corresponding sensor health state label of each sliding window is obtained; the mutual information between each type of feature X in the multi-dimensional feature vector and the health state label Y is calculated; Obtaining mutual information between each type feature and the health state label, if the calculated mutual information is greater than or equal to a preset threshold, it indicates that the type feature is significantly related to the sensor health state, and the type feature is marked as a key feature; otherwise, the type feature is discarded; Statistically screening the key features to obtain a key feature set sensitive to sensor reliability.

5. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 4, characterized in that, The expression of the mutual information is: In the formula, MI(X, Y) represents mutual information; P(X=xi, Y=yj) represents the joint probability of the feature X=xi and the health state label Y=yj; i=1, 2,..., k, k is the total number of sliding windows; j=1, 2, 3; xi is a multi-dimensional feature vector of the i-th sliding window, P(X=xi) represents the marginal probability of the feature X=xi; and P(Y=yj) represents the marginal probability of Y=yj.

6. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 5, characterized in that, The LSTM model construction and training steps include: S31: dividing the key feature set obtained in step two into a training set, a validation set and a test set in chronological order; S32: constructing an LSTM model structure, the model structure including an input layer, a hidden layer, a fully connected layer and an output layer; S33: using a sparse classification cross-entropy loss as a loss function, and using an Adam optimizer to accelerate convergence by adaptively adjusting the learning rate; S34: iteratively training the model using the training set, calculating the loss value and the classification accuracy after each round of training through the validation set, and introducing an early stopping mechanism to stop training when the loss value of the validation set does not decrease for 10 consecutive rounds, thereby avoiding model overfitting; S35: using a grid search method to optimize the hyperparameters, including the number of LSTM units, the Dropout rate, the sequence length and the learning rate; taking the validation set accuracy as the evaluation index, and selecting the optimal parameter combination when the validation set accuracy is greater than or equal to a preset standard to obtain the corresponding LSTM model.

7. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 6, characterized in that, S32 further includes: The input layer: the input dimension is the number of features of the key feature set, and each input sample is a time sequence sequence containing k consecutive sliding window key features; The hidden layer: a plurality of LSTM units are set to enhance the model's ability to capture long-term sequence dependencies by stacking; wherein a Dropout layer is added after each LSTM, and the dropout rate is set to 0.2-0.3 to randomly discard some neurons to prevent overfitting; The fully connected layer: the output of the last LSTM layer is flattened through a Flatten layer, then connected to a fully connected layer, and the activation function uses a ReLU function to enhance the model's non-linear fitting ability; The output layer: based on the health state label being 3 categories, the output layer is set to have 3 neurons, and the activation function uses a softmax function to output the probability distribution of each health state.

8. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 1, characterized in that, The early warning signal includes a first-level early warning signal, a second-level early warning signal and a third-level early warning signal.

9. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method according to claim 8, characterized in that, The specific implementation process of step four is as follows: S41: real-time acquisition of the original output signal of the MEMS inertial sensor and the environmental parameter data, and recording the acquisition timestamp, performing the same preprocessing and feature extraction process as steps one and two on the real-time data to obtain a real-time key feature set; S42: Organize the real-time key feature set into input samples in time sequence, input the trained LSTM model; the model output is the health state probability distribution [Y1, Y2, Y3] corresponding to the current sliding window; wherein Y1 is the normal state probability, Y2 is the degradation state probability, Y3 is the fault state probability, and Y1+Y2+Y3=1; S43: Based on the probability distribution output by the model, perform multi-dimensional reliability evaluation and output warning signals through the system terminal to prompt maintenance personnel to intervene in time.

10. The long short-term memory deep learning based MEMS inertial sensor reliability analysis method of claim 9, wherein, S43 also includes: State determination: select the state with the maximum probability as the current sensor health state; Degradation trend prediction: when the degradation state is determined, combine the historical data output as the degradation state, and use linear fitting to predict the degradation state probability to estimate the remaining useful life RUL from the current state to the failure state; Wherein, the calculation formula of the remaining useful life RUL is: In the formula, b is the proportional coefficient, Y2 is the current degradation probability, and ΔY2 is the average increase of the degradation probability per unit time; Fault warning: based on the current sensor health state, set up a multi-level warning mechanism; If the output is normal, continue to monitor the change trend of the degradation probability; If the output is a degradation state, set a hierarchical warning based on the remaining useful life RUL: When RUL>T1, output a first-level warning signal and strengthen monitoring; When RUL≤T1, output a second-level warning signal, indicating that the remaining useful life is close to the safety threshold, and maintenance resources are recommended to be prepared; When RUL≤T2, output a third-level warning signal, indicating that the remaining useful life is insufficient, and the sensor will soon enter a fault state, suggesting that immediate shutdown maintenance be arranged; If the output is a fault state, output a third-level warning signal and issue an audible and visual alarm.

Images