A fiber optic gyroscope fault diagnosis method based on MiniRocket network model

Abstract

The present invention discloses a fiber optic gyroscope fault diagnosis method based on MiniRocket network model, comprising the following steps: calibrating multiple fiber optic gyroscopes of the same model using a standard angular velocity source, obtaining the output signal of each fiber optic gyroscope and the corresponding temperature stress signal; marking the output signal of each fiber optic gyroscope, and performing noise reduction processing through the EEMD algorithm to construct a data set; training the MiniRocket network model, and tuning the model hyperparameters in combination with the sparrow optimization algorithm; and performing fiber optic gyroscope fault diagnosis of the same model based on the optimized MiniRocket network model. The method can effectively improve the accuracy and reliability of fiber optic gyroscope fault diagnosis through precise data acquisition, effective preprocessing means, efficient feature extraction and model optimization, and real-time fault diagnosis, especially being able to accurately distinguish intermittent faults from permanent faults, thereby achieving better maintenance of the inertial navigation system.

Description

Technical Field

[0001] The invention relates to the field of fiber optic gyroscopes of inertial navigation systems, and more particularly to a fiber optic gyroscope fault diagnosis method based on a MiniRocket network model. Background Art

[0002] As a high-precision angular velocity sensor, fiber optic gyroscope (FOG) plays an important role in aerospace, navigation, and national defense. However, under long-term working conditions, fiber optic gyroscopes are easily affected by environmental factors such as temperature stress, resulting in performance degradation and possible intermittent failures. Intermittent failures usually manifest as random and uncertain behaviors, which pose a challenge to the reliability and stability of inertial navigation systems.

[0003] The current fiber optic gyroscope fault diagnosis technology mainly relies on the original output signal of the fiber optic gyroscope, and most of them use analog signal methods to obtain intermittent fault data. These methods are based on specific models and assumptions. Although they can provide certain fault information, they have the following limitations:

[0004] The analog signal method relies on pre-set models, which often simplify the complex situations in the actual working environment, so it is difficult to truly reflect the randomness of intermittent faults. In addition, the signal processing methods used, such as empirical mode decomposition (EMD), may have mode aliasing problems when processing nonlinear and non-stationary signals, which makes the signal decomposition effect less than ideal.

[0005] At the same time, due to the randomness and uncertainty of intermittent faults, existing fault diagnosis methods are difficult to accurately capture and identify the occurrence time and frequency of such faults, which limits their accuracy and reliability in practical applications.

[0006] Traditional one-dimensional convolutional neural networks (1D-CNN) often require a large number of training samples and a long training time to optimize model parameters, and a large number of parameters need to be adjusted when processing time series data, which not only increases the training time and computing resource requirements, but also may limit the generalization ability of the model.

[0007] Therefore, how to design a fiber optic gyroscope fault diagnosis method based on the MiniRocket network model, which can improve the accuracy and reliability of fiber optic gyroscope fault diagnosis under the premise of reducing model training time and computing resource consumption, is an urgent problem to be solved by technical personnel in this field. Summary of the invention

[0008] In view of this, the present invention provides a fiber optic gyroscope fault diagnosis method based on the MiniRocket network model, which can effectively extract intermittent fault features from complex signals; at the same time, with the help of the efficient time series classification capability of the MiniRocket network model, fast feature extraction and classification are achieved, which reduces the computing cost and model training time, thereby realizing accurate diagnosis of intermittent faults of fiber optic gyroscopes.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] A fiber optic gyroscope fault diagnosis method based on MiniRocket network model comprises the following steps:

[0011] S1. Calibrate multiple fiber optic gyroscopes of the same model using a standard angular velocity source to obtain an output signal and a corresponding temperature stress signal of each fiber optic gyroscope;

[0012] S2, label the output signal of each fiber optic gyroscope, perform noise reduction processing through the EEMD algorithm, and construct a data set;

[0013] S3. Use the data set to train the MiniRocket network model, and use the sparrow optimization algorithm to tune the model hyperparameters to obtain an optimized MiniRocket network model.

[0014] S4. Based on the optimized MiniRocket network model, perform fault diagnosis on the same type of fiber optic gyroscope and obtain the corresponding fault diagnosis results.

[0015] Preferably, in S2, labeling the output signal of each fiber optic gyroscope includes:

[0016] Calculate the zero drift value of each fiber optic gyroscope :

[0017]

[0018] Where K is a constant, n is the number of output signals, is the actual output value of the i-th sampling data point of the output signal, is the average value of all output signals;

[0019] Based on the zero drift value of each fiber optic gyroscope, the normal state, intermittent fault and permanent fault are marked in combination with the fiber optic gyroscope detection threshold;

[0020] The normal state is that the zero drift value of the fiber optic gyroscope is less than the detection threshold, the intermittent fault is that the zero drift value of the fiber optic gyroscope is greater than the detection threshold and then recovers to the normal state, and the permanent fault is that the zero drift value of the fiber optic gyroscope cannot recover to the normal state after it exceeds the detection threshold.

[0021] Preferably, in S2, performing noise reduction processing by using an EEMD algorithm includes:

[0022] Add white noise to the output signal of each fiber optic gyroscope after marking:

[0023]

[0024] in, is the output signal after adding white noise, is the output signal after annotation, is white noise;

[0025] For each output signal after adding white noise Perform multiple EMD decompositions to obtain multiple modal functions with different serial numbers. and the residual ; In each EMD decomposition process, a set of mode functions and a residual term are obtained;

[0026] For all mode functions with the same sequence number in each decomposition process and the residual Perform average processing to obtain decomposition results;

[0027] Based on the decomposition result, a reconstructed sequence synthesis is performed to obtain an output signal X(t) after noise reduction processing.

[0028] Preferably, the averaging process is expressed as:

[0029]

[0030]

[0031] in, is the IMF estimate, is the final residual term.

[0032] Preferably, in the EEMD algorithm, the difference between adjacent modal functions is calculated, and the EMD decomposition termination condition is generated in combination with a preset difference threshold; the EMD decomposition termination condition is: the difference between adjacent modal functions is less than the preset difference threshold.

[0033] Preferably, in S3, the MiniRocket network model includes:

[0034] Multi-channel random convolution layer: used to perform multi-channel random convolution operations on the data, perform initial feature fusion and standardization, and extract initial features;

[0035] Point convolution layer: used to perform point convolution feature fusion on the initial features to obtain fused features;

[0036] Normalization layer: used to normalize the fused features;

[0037] One-dimensional convolution layer: used to extract key features from the features after normalization operation;

[0038] Pooling layer: used to reduce the dimension of the key features by using maximum pooling or average pooling;

[0039] Multi-layer perceptron (MLP): used to predict fiber optic gyroscope fault types through low-dimensional key features.

[0040] Preferably, the multi-channel random convolution layer includes a set of preset number of random convolution kernels; the random convolution kernels randomly generate weights and biases for feature extraction;

[0041] In the multi-channel random convolution layer, the normalization process adopts z-score normalization.

[0042] Preferably, the multi-layer perceptron MLP is a single hidden layer structure, the number of neurons in the hidden layer is 100, the activation function of the hidden layer is a ReLu function, and the learning rate is adjusted in combination with the Adam optimization algorithm;

[0043] The multi-layer perceptron MLP output layer includes three neurons; the three neurons correspond to three types of fiber optic gyroscope states respectively, and the SoftMax function is used as the activation function.

[0044] Preferably, in S3, using the data set to train the MiniRocket network model includes: dividing the data set into a training set and a test set, which are applied to the MiniRocket network model training mode and evaluation mode respectively;

[0045] In the MiniRocket network model training mode and evaluation mode, the model training loss and classification accuracy are monitored in real time, and the training loss and classification accuracy curves are plotted to identify the model's fitting or underfitting.

[0046] Preferably, the S4 further includes:

[0047] Using a standard angular velocity source to calibrate the fiber optic gyroscope of the same model to obtain an output signal of the fiber optic gyroscope of the same model;

[0048] The output signal of the fiber optic gyroscope of the same model is subjected to noise reduction processing using the EEMD algorithm in S2.

[0049] It can be seen from the above technical solution that, compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0050] 1. Use a standard angular velocity source to calibrate the fiber optic gyroscope to ensure the accuracy of data collection. Calibrate multiple fiber optic gyroscopes of the same model and record the output signal and temperature stress signal at the same time, which helps to more comprehensively understand the working status of the fiber optic gyroscope and its relationship with environmental factors.

[0051] 2. Different working states are distinguished by calculating the zero drift value, and the EEMD (Ensemble Empirical Mode Decomposition) algorithm is used for noise reduction. By performing multiple EMD decompositions and averaging on the signal with white noise added, the noise interference can be effectively removed and the effective components of the signal can be retained.

[0052] 3. The multi-channel random convolution layer in the MiniRocket network model can quickly extract features from the signal, and the feature fusion is performed through the point convolution layer, further enhancing the feature expression ability. In addition, the hyperparameter tuning combined with the sparrow optimization algorithm helps to find the optimal model configuration and improve the diagnostic accuracy.

[0053] 4. Through continuous calibration with a standard angular velocity source and the optimized MiniRocket network model, different fault types of the fiber optic gyroscope in various scenarios can be efficiently and accurately identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0055] Figure 1 A flow chart of a fiber optic gyroscope fault diagnosis method based on a MiniRocket network model provided in an embodiment of the present invention;

[0056] Figure 2 MiniRocket network model structure framework diagram provided for an embodiment of the present invention;

[0057] Figure 3 A graph showing the loss curve of model training provided by an embodiment of the present invention;

[0058] Figure 4 A curve diagram of model classification accuracy provided by an embodiment of the present invention;

[0059] Figure 5A schematic diagram of a classification prediction confusion matrix provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Embodiment 1;

[0062] like Figure 1 As shown, this embodiment provides a fiber optic gyroscope fault diagnosis method based on the MiniRocket network model, comprising the following steps:

[0063] S1. Calibrate multiple fiber optic gyroscopes of the same model using a standard angular velocity source to obtain an output signal and a corresponding temperature stress signal of each fiber optic gyroscope;

[0064] S2, label the output signal of each fiber optic gyroscope, perform noise reduction processing through the EEMD algorithm, and construct a data set;

[0065] S3. Use the data set to train the MiniRocket network model, and use the sparrow optimization algorithm to tune the model hyperparameters to obtain an optimized MiniRocket network model.

[0066] S4. Based on the optimized MiniRocket network model, perform fault diagnosis on the same type of fiber optic gyroscope and obtain the corresponding fault diagnosis results.

[0067] This method can effectively improve the accuracy and reliability of fiber optic gyroscope fault diagnosis through precise data acquisition, effective preprocessing means, efficient feature extraction and model optimization, and real-time fault diagnosis. In particular, it can accurately distinguish between intermittent faults and permanent faults, thereby achieving better maintenance of the inertial navigation system.

[0068] The above steps and related technical features are further described in detail below:

[0069] In this embodiment S1, a standard angular velocity source is used to calibrate multiple fiber optic gyroscopes of the same model to obtain the output signal and the corresponding temperature stress signal of each fiber optic gyroscope;

[0070] In this step, each group of 8 fiber optic gyroscopes of the same model and produced in the same batch are placed under different temperature conditions, and variable temperature stress is applied during the measurement process. The sampling frequency is 1HZ (the fiber optic gyroscope output data and temperature stress data are collected once per second), each sampling lasts 2.5 hours, and the length of each sample is 8999. And the output signal of each fiber optic gyroscope is the output signal of the fiber optic gyroscope from the input angular velocity signal entering a stable state to the end of the experimental cycle.

[0071] In this embodiment S2, the output signal of each fiber optic gyroscope is labeled and subjected to noise reduction processing by the EEMD algorithm to construct a data set; 70% of the data is used as a training set and 30% of the data is used as a test set.

[0072] The output signal of each fiber optic gyroscope is marked, including:

[0073] Calculate the zero drift value of each fiber optic gyroscope :

[0074]

[0075] Where K is a constant, n is the number of output signals, is the actual output value of the i-th sampling data point of the output signal, is the average value of all output signals;

[0076] Based on the zero drift value of each fiber optic gyroscope, the normal state, intermittent fault and permanent fault are marked in combination with the fiber optic gyroscope detection threshold; the fiber optic gyroscope detection threshold is 0.03° / s, that is, when the fiber optic gyroscope zero drift value exceeds 0.03° / s, the fiber optic gyroscope is judged to have a fault. Among them, the labels corresponding to the normal state, intermittent fault and permanent fault are 0, 1, and 2 respectively.

[0077] The normal state is that the zero drift value of the fiber optic gyroscope is less than the detection threshold, the intermittent fault is that the zero drift value of the fiber optic gyroscope is greater than the detection threshold and then recovers to the normal state, and the permanent fault is that the zero drift value of the fiber optic gyroscope cannot recover to the normal state after it exceeds the detection threshold.

[0078] Furthermore, denoising is performed using the EEMD algorithm, which does not require the number of intrinsic mode functions (IMFs) to be pre-set. The algorithm introduces white noise into the data and performs multiple independent decompositions, effectively avoiding the problem of modal aliasing and significantly enhancing the stability and reliability of the decomposition results.

[0079] Specifically include: adding white noise to the output signal of each fiber optic gyroscope after marking:

[0080]

[0081] in, is the output signal after adding white noise, is the output signal after annotation, is white noise;

[0082] For each output signal after adding white noise Perform multiple EMD decompositions to obtain multiple modal functions with different serial numbers. and the residual ; In each EMD decomposition process, a set of mode functions and a residual term are obtained;

[0083] For all mode functions with the same sequence number in each decomposition process and the residual Perform average processing to obtain decomposition results;

[0084] Based on the decomposition result, a reconstructed sequence synthesis is performed to obtain an output signal X(t) after noise reduction processing.

[0085] The averaging process is expressed as:

[0086]

[0087]

[0088] in, is the IMF estimate, is the final residual term.

[0089] In the EEMD algorithm, the difference between adjacent modal functions is calculated, and the EMD decomposition termination condition is generated in combination with a preset difference threshold; it can adaptively determine the number of IMFs according to the characteristics of the data itself, thereby providing more accurate and reliable time series analysis results for various practical applications.

[0090] In this implementation, the EMD decomposition termination condition is: the difference between adjacent modal functions is less than a preset difference threshold.

[0091] In this step, the EEMD (Ensemble Empirical Mode Decomposition) algorithm can solve the mode aliasing problem that occurs in the EMD (Empirical Mode Decomposition) algorithm. By adding white noise to the original data, EMD decomposition is performed multiple times to reduce the interference between the original modes. Each added noise enables the algorithm to distinguish different intrinsic mode functions (IMFs) more clearly. The results of each decomposition are averaged to suppress the influence of random noise and enhance the stability of the results. The IMF generated by the EEMD algorithm is more stable and reliable than the traditional EMD algorithm, and can be used to process complex signals containing noise; the EEMD algorithm can effectively distinguish intermittent faults from observation noise.

[0092] In this embodiment S3, the input data of the MiniRocket network model are the output signal of the labeled and denoised fiber optic gyroscope and the corresponding temperature stress signal, that is, the input of the MiniRocket network model includes two features, both of which are time series signals.

[0093] like Figure 2 As shown, the MiniRocket network model includes:

[0094] Multi-channel random convolution layer: used to perform multi-channel random convolution operations on data, perform initial feature fusion and standardization processing, and extract initial features; the multi-channel random convolution layer includes a set of preset number of random convolution kernels, specifically Roc-1, Roc-2, Roc-3, Roc-(k-1), Roc-k; the random convolution kernel randomly generates weights and biases for feature extraction; in the multi-channel random convolution layer, z-score standardization is used for standardization processing.

[0095] In the design of the multi-channel random convolution layer, the idea of ​​multi-scale convolution is combined, and the convolution kernel in the traditional multi-scale convolution is replaced with a random convolution kernel. The diversity and extensiveness of the random convolution kernel are used to enhance the ability of feature extraction. The extracted features are initially fused and normalized to ensure the effectiveness and stability of the features in the classification model. This preliminary and diverse feature extraction provides a richer feature representation for subsequent layers, which helps capture patterns and details in the data that are not perceived by typical methods.

[0096] Point convolution layer: used to perform point convolution feature fusion on the initial features to obtain fused features; use the point convolution layer to fuse the features extracted by random convolution, and merge the features extracted by multiple convolution layers into a more compact and rich feature representation. Point convolution operates each feature point independently in the depth of the feature map, effectively merging information from different feature channels, and is suitable for processing time series data, which focuses on temporal dynamics rather than spatial or sequence relationships.

[0097] Normalization layer: used to normalize the fused features;

[0098] One-dimensional convolution layer: used to extract key features from the features after normalization operation; the one-dimensional convolution layer is designed to have two layers, with the number of convolution kernels being 64 and 128 respectively. The convolution kernel size is 3, and the step size is 1;

[0099] The one-dimensional convolution layer can be regarded as a "coarse to fine" feature extraction strategy. The random layer may capture rough and broad patterns, which can be analyzed and optimized more finely based on the multi-channel random convolution layer, thereby improving the overall feature expression ability.

[0100] Pooling layer: used to reduce the dimension of the key features by using maximum pooling or average pooling; the pooling window size is 2 and the step size is 2.

[0101] After extracting features through multi-channel random convolutional layers and one-dimensional convolutional layer operations, the spatial dimension of the feature map is further reduced through pooling operations, thereby reducing the number of parameters and computational complexity.

[0102] The pooling layer includes two forms: maximum pooling and average pooling. The maximum pooling layer maintains the significant features in the feature map by extracting the maximum value in the region, while the average pooling calculates the average value in the region, which helps to smooth the feature representation. In this way, the pooling layer can effectively reduce the risk of overfitting and improve the generalization ability of the model. In addition, the pooling operation can also help the model identify similar features at different locations and enhance the robustness of the model to position changes in the input data.

[0103] Multi-layer perceptron (MLP): used to predict fiber optic gyroscope fault types through low-dimensional key features.

[0104] The multilayer perceptron MLP is a single hidden layer structure, the number of neurons in the hidden layer is 100, the hidden layer activation function is the ReLu function, and the learning rate is adjusted in combination with the Adam optimization algorithm; the multilayer perceptron MLP output layer includes three neurons; the three neurons correspond to three types of fiber optic gyroscope states, respectively, and the SoftMax function is used as the activation function.

[0105] The MiniRocket network model receives dual-channel timing signal input, extracts initial features through a multi-channel random convolution layer and normalizes them, and then fuses these features using a point convolution layer. After key feature extraction by a one-dimensional convolution layer and dimensionality reduction by a pooling layer, a multi-layer perceptron (MLP) containing a single hidden layer is used to predict the type of fiber optic gyro fault.

[0106] Further, the data set is used to train the MiniRocket network model, including: dividing the data set into a training set and a test set, which are applied to the MiniRocket network model training mode and evaluation mode respectively;

[0107] like Figure 3 and Figure 4As shown in the figure, in the MiniRocket network model training mode and evaluation mode, the model training loss and recognition accuracy under different training rounds are monitored in real time, and the training loss and classification accuracy curves are plotted to identify the model's fitting or underfitting. By observing the changes in the training loss and classification accuracy curves with the training rounds (epochs), it can help understand the dynamic learning process of the model. The overfitting or underfitting problems of the model can be discovered early and handled in time to improve the generalization ability of the model.

[0108] To prevent overfitting, regularization techniques such as Dropout or L2 regularization are used. The entire training process goes through multiple training rounds to ensure that the model can fully learn the features in the training set.

[0109] like Figure 5 As shown in the figure, during the model verification process, the performance of the model is measured by calculating different evaluation indicators. This includes evaluating the model performance through the confusion matrix, measuring the gap between the model prediction results and the true labels, which helps to understand the classification effect of the model on different categories. For fiber optic gyroscope fault diagnosis, the confusion matrix can be used to see the prediction effect of the model on the three categories of normal state, intermittent fault and permanent fault, and then analyze which type of fault the model performs better or worse.

[0110] If you find that the performance of the model does not meet expectations, you need to return to the model adjustment stage to modify and optimize the network structure, parameters during training, or training strategies. For example, adjust the learning rate, increase the number of training rounds, change the optimization algorithm, or introduce a more complex model structure to improve the accuracy and generalization ability of the model. This iterative process is an indispensable part of model development to ensure that the final deployed model can achieve the best results in the actual environment.

[0111] In particular, the Sparrow Optimization Algorithm (SSA) is used to set hyperparameters such as learning rate, convolution kernel size, and training rounds as optimization targets to perform model hyperparameter tuning. Specifically, it includes:

[0112] Define the range of hyperparameters: determine the range of values ​​of hyperparameters such as learning rate, convolution kernel size, and training rounds; Initialize the population: randomly generate a certain number of solutions, each solution contains a set of hyperparameter configurations; Evaluate model performance: train the model for each solution (hyperparameter configuration) and evaluate the performance on the validation set; Execute the SSA algorithm: update the solutions in the population according to the rules of SSA, and find the optimal solution by simulating the foraging behavior of sparrows; Termination condition: stop the optimization process when the predetermined maximum number of iterations is reached or the performance is no longer significantly improved; Select the optimal hyperparameters: select the solution with the best performance from the final population as the optimal hyperparameter configuration.

[0113] Through the above steps, the sparrow search algorithm can be effectively used to optimize the model's hyperparameters and improve the model's accuracy and generalization ability. It is particularly suitable for tasks such as fiber optic gyroscope intermittent fault diagnosis, which can improve the model's diagnostic efficiency and accuracy.

[0114] In this embodiment S4, after completing the training evaluation and final optimization of the model, a final fault diagnosis model is obtained, and the model can be used to diagnose the intermittent fault of the fiber optic gyroscope of this model.

[0115] Embodiment 2;

[0116] In this embodiment, the fault diagnosis of the same type of fiber optic gyroscope based on the optimized MiniRocket network model is further described in detail;

[0117] S4.1. Calibration preparation:

[0118] Select a high-precision standard angular velocity source, such as a precision turntable. Identify the fiber optic gyro to be tested and make sure it is the same model as the one used to train the model.

[0119] Data acquisition: Connect the fiber optic gyro to a standard angular velocity source and record the output signal of the fiber optic gyro. If 5 minutes of data are collected, the sampling frequency is 100 Hz, which means there are 30,000 sampling points.

[0120] The fiber optic gyroscope is subjected to a series of known angular velocity changes, and the output signal is recorded. At the same time, the corresponding temperature stress signal is recorded for subsequent analysis.

[0121] S4.2, EEMD algorithm noise reduction processing:

[0122] An appropriate amount of white noise is added to the output signal of the fiber optic gyroscope to reduce the nonlinear effect in the signal. In this embodiment, the intensity of the white noise is set to 10% of the original signal intensity.

[0123] Perform empirical mode decomposition (EMD) on the signal with white noise added to obtain multiple mode functions (IMFs) and a residual term. Perform EMD decomposition five times, and average all mode functions and residual terms with the same sequence number in each decomposition process. Reconstruct the signal based on the processed IMFs and residual terms to obtain the output signal after noise reduction.

[0124] S43, fault diagnosis:

[0125] The output signal after noise reduction is converted into a data format suitable for MiniRocket network model input. The data format usually includes information such as signal length and sampling frequency.

[0126] The data is input into the trained and optimized MiniRocket network model: Inside the model, the data first passes through a multi-channel random convolution layer, which randomly generates multiple convolution kernels to extract initial features. Then, these features are fused through a point convolution layer and normalized through a normalization layer to reduce internal covariate shift.

[0127] Next, the data is passed through a one-dimensional convolutional layer to extract key features. These features are then reduced in dimension through a pooling layer to reduce the amount of computation and avoid overfitting.

[0128] Finally, the extracted features are fed into a multi-layer perceptron (MLP) for fault type prediction. The hidden layer of the MLP uses the ReLU activation function to increase nonlinearity, while the output layer uses the Softmax activation function to convert the output into a probability distribution. The Softmax function ensures that the sum of all output values ​​is 1, and each value represents the probability that the input data belongs to the corresponding fault type.

[0129] The model outputs a probability distribution with three elements, corresponding to the three fault types (normal state, intermittent fault, and permanent fault). The most likely fault type is determined by comparing these three probability values.

[0130] The probability distribution of the output is [0.1, 0.8, 0.1]. Since the probability of intermittent failure is the highest, it is diagnosed that the fiber optic gyroscope has an intermittent failure.

[0131] In the fiber optic gyro fault diagnosis method based on the MiniRocket network model in the above embodiment, a series of randomly generated convolution kernels and pooling operations are used to directly map the original time series data to a high-dimensional feature space, and then use these feature vectors for efficient classification. Compared with the traditional one-dimensional convolutional neural network (1D-CNN), it significantly reduces the dependence on a large number of training samples, shortens the training cycle, and simplifies the model tuning process, because its core convolution kernel parameters are random and fixed, and do not need to be optimized through a complex learning process.

[0132] This design not only reduces the computational burden, but also improves the efficiency and flexibility of model training. Given the high incidence of intermittent faults in fiber optic gyroscope systems and their serious impact on system stability, it can effectively improve the accuracy and reliability of fiber optic gyroscope fault diagnosis, especially accurately distinguishing intermittent faults from permanent faults, thereby achieving better maintenance of inertial navigation systems.

[0133] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0134] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fiber optic gyroscope fault diagnosis method based on MiniRocket network model, characterized in that: The following steps are involved: S1. Calibrate multiple fiber optic gyroscopes of the same model using a standard angular velocity source to obtain an output signal and a corresponding temperature stress signal of each fiber optic gyroscope; S2, label the output signal of each fiber optic gyroscope, perform noise reduction processing through the EEMD algorithm, and construct a data set; The step of marking the output signal of each fiber optic gyroscope comprises: Calculate the zero drift value B of each fiber optic gyroscope s : Where K is a constant, n is the number of output signals, and F i is the actual output value of the i-th sampling data point of the output signal, is the average value of all output signals; Based on the zero drift value of each fiber optic gyroscope, the normal state, intermittent fault and permanent fault are marked in combination with the fiber optic gyroscope detection threshold; The normal state is that the zero drift value of the fiber optic gyroscope is less than the detection threshold, the intermittent fault is that the zero drift value of the fiber optic gyroscope is greater than the detection threshold and then recovers to the normal state, and the permanent fault is that the zero drift value of the fiber optic gyroscope is greater than the detection threshold and cannot recover to the normal state; S3. Use the data set to train the MiniRocket network model, and use the sparrow optimization algorithm to tune the model hyperparameters to obtain an optimized MiniRocket network model. S4. Based on the optimized MiniRocket network model, perform fault diagnosis on the same type of fiber optic gyroscope and obtain the corresponding fault diagnosis results.

2. The fiber optic gyroscope fault diagnosis method based on the MiniRocket network model according to claim 1, characterized in that: In S2, noise reduction processing is performed by using the EEMD algorithm, including: Add white noise to the output signal of each fiber optic gyroscope after marking: X i (t))=X(t)+n i (t) Among them, x i (t) is the output signal after adding white noise, x(t) is the output signal after annotation, n i (t) is white noise; For each output signal x after adding white noise i (t) Perform multiple EMD decompositions to obtain multiple modal functions IMF under different serial numbers. ij (t) and the residual term r i (t); wherein, in each EMD decomposition process, a set of mode functions and a residual term are obtained; For all the modal functions IMF with the same sequence number in each decomposition process ij (t) and the residual term r i (t) Performing average processing to obtain decomposition results; Based on the decomposition result, a reconstructed sequence synthesis is performed to obtain an output signal X(t) after noise reduction processing.

3. The fiber optic gyroscope fault diagnosis method based on the MiniRocket network model according to claim 2, characterized in that: The averaging process is expressed as: Among them, IMF j (t) is the IMF estimate and r(t) is the final residual term.

4. The fiber optic gyroscope fault diagnosis method based on the MiniRocket network model according to claim 2, characterized in that: In the EEMD algorithm, the difference between adjacent modal functions is calculated, and the EMD decomposition termination condition is generated in combination with a preset difference threshold; the EMD decomposition termination condition is: the difference between adjacent modal functions is less than the preset difference threshold.

5. The fiber optic gyroscope fault diagnosis method based on the MiniRocket network model according to claim 1, characterized in that: The S3, MiniRocket network model, includes: Multi-channel random convolution layer: used to perform multi-channel random convolution operations on the data, perform initial feature fusion and standardization, and extract initial features; Point convolution layer: used to perform point convolution feature fusion on the initial features to obtain fused features; Normalization layer: used to normalize the fused features; One-dimensional convolution layer: used to extract key features from the features after normalization operation; Pooling layer: used to reduce the dimension of the key features by using maximum pooling or average pooling; Multi-layer perceptron (MLP): used to predict fiber optic gyroscope fault types through low-dimensional key features.

6. The method for diagnosing fiber optic gyroscope faults based on the MiniRocket network model according to claim 5, characterized in that: The multi-channel random convolution layer includes a set of preset number of random convolution kernels; the random convolution kernels randomly generate weights and biases for feature extraction; In the multi-channel random convolution layer, the normalization process adopts z-score normalization.

7. The method for diagnosing fiber optic gyroscope faults based on the MiniRocket network model according to claim 5, characterized in that: The multi-layer perceptron MLP is a single hidden layer structure, the number of neurons in the hidden layer is 100, the hidden layer activation function is the ReLu function, and the learning rate is adjusted in combination with the Adam optimization algorithm; The multi-layer perceptron MLP output layer includes three neurons; the three neurons correspond to three types of fiber optic gyroscope states respectively, and the SoftMax function is used as the activation function.

8. The method for diagnosing fiber optic gyroscope faults based on MiniRocket network model according to claim 1, characterized in that: In S3, the MiniRocket network model training is performed using the data set, including: dividing the data set into a training set and a test set, which are applied to the MiniRocket network model training mode and the evaluation mode respectively; In the MiniRocket network model training mode and evaluation mode, the model training loss and classification accuracy are monitored in real time, and the training loss and classification accuracy curves are plotted to identify the model's fitting or underfitting.

9. The method for diagnosing fiber optic gyroscope faults based on the MiniRocket network model according to claim 1, characterized in that: The S4 further includes: Using a standard angular velocity source to calibrate the fiber optic gyroscope of the same model to obtain an output signal of the fiber optic gyroscope of the same model; The output signal of the fiber optic gyroscope of the same model is subjected to noise reduction processing using the EEMD algorithm in S2.

Images