Intermittent fault feature identification method and system applied to fiber-optic gyroscope

Abstract

The invention discloses an intermittent fault feature identification method and system applied to a fiber-optic gyroscope, and relates to the technical field of fiber-optic gyroscope health management, and the method comprises the following specific steps: obtaining an FOG signal, and converting the FOG signal into null drift data; dividing the effective number into a normal section, an intermittent fault section and a permanent fault section according to a preset null drift threshold and a division rule, and extracting samples according to a non-overlapping principle to construct an FOG time sequence data set; generating a candidate Shapelets set based on the FOG time sequence data set through a sliding time window, and constructing a training data set by using the candidate Shapelets set; and constructing a learning Shapelets model, inputting the to-be-identified data into the learning Shapelets model, and outputting a classification result. According to the method, by determining the section segmentation points of the FOG signals, the original signals are segmented into the independent sections with the clear state modes, the feature extraction range is effectively narrowed, and the calculation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fiber optic gyroscope health management, and more specifically, to an intermittent fault feature recognition method and system applied to fiber optic gyroscopes. Background Art

[0002] The fiber optic gyroscope is a medium and high-precision angular velocity sensor commonly used in inertial navigation systems. It is one of the core components of inertial navigation systems and has become an indispensable technology in both civilian and military fields. Conducting research on intermittent faults of fiber optic gyroscopes and accurately identifying fault modes is of great significance for formulating reasonable maintenance strategies and ensuring the normal operation of equipment.

[0003] The entire life cycle of a fiber optic gyroscope includes three main states: normal state, intermittent fault state, and permanent fault state. Each state has different signal characteristics. By extracting the different fault state characteristics of the fiber optic gyroscope from the FOG monitoring data, the differences in the signal characteristics of each fault mode of the fiber optic gyroscope can be identified, and accurate control of the fault mode of the fiber optic gyroscope can be achieved. Using FOG data to identify the intermittent fault mode of a fiber optic gyroscope is a difficult point in current research. First, the working environment and monitoring equipment cause noise, data loss, abnormal data, and difficulties in dividing the signal segments of fault modes, resulting in difficulties in differentiating various fault characteristics of fiber optic gyroscopes; although existing data cleaning methods can handle general noise and errors, intermittent faults of fiber optic gyroscopes usually occur accidentally and discontinuously within a period of time, and the occurrence time and frequency are often not fixed, making it difficult to divide the intermittent fault signal sections; traditional time series classification methods usually lack sufficient interpretability and are not effective in local feature recognition problems. Since only permanent faults are currently concerned in fiber optic gyroscope fault recognition, no research has been carried out on intermittent fault problems.

[0004] The characteristics of intermittent fault problems are that the occurrence time and duration of each fault are different, and the mathematical model of fault generation is not clear. In this case, it is not suitable to use other various methods, and only the classification method based on local features is more in line with the requirements of intermittent fault problems. Among them, the Shapelet method is very suitable for analyzing intermittent fault problems. Shapelet is a feature extraction method for time series data classification. It represents a certain subsequence with discriminability in the time series, and this subsequence can best reflect the category information, can better identify the local features of faults, and has strong interpretability. However, the traditional Shapelet method needs to search the entire subsequence set, resulting in a huge amount of calculation and ultimately poor timeliness.

[0005] Therefore, developing a method that can effectively identify the intermittent fault characteristics of fiber optic gyroscopes is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides an intermittent fault feature recognition method and system for an optical fiber gyroscope, which overcomes the above defects.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] An intermittent fault feature recognition method for an optical fiber gyroscope, the specific steps are as follows:

[0009] Obtain the FOG signal and convert the FOG signal into zero drift data;

[0010] Extract valid data in the zero drift data that satisfies the initial point being lower than the threshold line and having more than two intersections with the threshold line, and divide the valid data into a normal section, an intermittent fault section, and a permanent fault section according to a preset zero drift threshold according to a division rule;

[0011] Extract samples in the normal section, the intermittent fault section, and the permanent fault section respectively according to the non-overlapping principle to construct a FOG time series data set; among them, when selecting intermittent fault samples, first determine the midpoint, and then intercept the FOG signals in equal-length sections before and after the midpoint as intermittent fault samples;

[0012] Generate a candidate Shapelets set based on the FOG time series data set through a sliding time window, calculate the information gain of each candidate Shapelets, and perform optimal Shapelets screening based on the information gain to construct an optimal Shapelets set; calculate distance data based on the optimal Shapelets set and the FOG time series data set, generate a plurality of feature vectors, and construct a training data set based on the plurality of feature vectors, and use a clustering algorithm to obtain clustering centroids according to the training data set;

[0013] Construct a preliminary learning Shapelets model, and use the training data set and the clustering centroids to iteratively train the preliminary learning Shapelets model to obtain a learning Shapelets model;

[0014] Obtain the data to be recognized, input the data to be recognized into the learning Shapelets model, and output a classification result.

[0015] Optionally, the conversion expression of the zero drift data is:

[0016]

[0017] In the formula, B s represents zero drift, K represents the scale factor, n represents the number of sampling times, F fRepresents the output value of the f-th point, Represents the output mean.

[0018] Optionally, the division rule is as follows: taking the starting point of the valid data as the starting point of the normal section, the intersection point where it first intersects with the threshold line as the ending point of the normal section, and the section between the starting point and the ending point of the normal section as the normal section; taking the intersection point where it last intersects with the threshold line as the starting point of the permanent fault section, the ending point of the valid data as the ending point of the permanent fault section, and the section between the starting point and the ending point of the permanent fault section as the permanent fault section; the section between the ending point of the normal section and the starting point of the permanent fault section is the intermittent fault section.

[0019] Optionally, the steps for obtaining the intermittent fault samples are as follows:

[0020] Obtain the degradation time series data of the intermittent fault section and segment the degradation time series data through the threshold line;

[0021] Identify all intersection points of the degradation time series data and the threshold line to form an intersection set;

[0022] Filter all even-numbered interaction events in the intersection set and extract the critical positions where the degradation time series data corresponding to each event drops back to the zero drift threshold;

[0023] Taking each of the critical positions as the midpoint, intercept equidistant data segments before and after in the FOG signal as the intermittent fault samples.

[0024] Optionally, the calculation formula for the information gain is:

[0025]

[0026] In the formula, Ent is entropy; D is the total number of samples; D1 is the number of normal samples; D2 is the number of intermittent fault samples; D3 is the number of permanent fault samples; Ent1 is the entropy of the normal sample subset; Ent2 is the entropy of the intermittent fault sample subset; Ent3 is the entropy of the permanent fault sample subset.

[0027] Optionally, the distance data calculation formula is:

[0028]

[0029] In the formula, D i,j Represents the minimum distance between the i-th time series and the j-th shapelet, T i,j+l-1 Represents the i-th time series T i The l-th point starting from position j in, S j，l Represents the l-th point of the j-th shapelet.

[0030] Optionally, the obtaining step of the learning Shapelets model is as follows:

[0031] Decompose the multi-classification problem into multiple binary-classification sub-problems, and construct sub-models for each binary-classification sub-problem respectively;

[0032] Use the clustering centroids as the initial parameters of the Shapelets of each sub-model, and iteratively train each sub-model in parallel using the training data set until the model converges to obtain the learning Shapelets model; during the training process, calculate the prediction loss of each sub-model through forward propagation, update the Shapelets and classification weights through backpropagation, and generate the final classification prediction result by combining the prediction results of all sub-models.

[0033] Optionally, the classification objective function expression is:

[0034]

[0035] In the formula, S represents the set of Shapelets; W represents the weight; F represents the output data; i represents the i-th time series sample; c represents the c-th category; L represents the logical loss between the true label and the predicted label; Y i d represents the true label; represents the predicted label; λ W represents the regularization coefficient.

[0036] Optionally, the expression of the prediction result of the sub-model is:

[0037]

[0038] In the formula, W c，0 represents the bias term in the sub-model; R represents the number of scales; K represents the number of shapes at each scale; M r，i，k is the similarity between the Shapelet and the time series, and W c，r，k is the importance of the Shapelet for classification.

[0039] An intermittent fault feature recognition system applied to a fiber optic gyroscope, comprising:

[0040] A data acquisition and conversion module, configured to acquire the FOG signal and convert the FOG signal into zero-drift data;

[0041] A data optimization and segmentation module, which is used to extract valid data that meets the conditions that the initial point is lower than the threshold line and there are more than two intersections with the threshold line from the zero-drift data, and divide the valid data into a normal section, an intermittent fault section, and a permanent fault section according to a preset zero-drift threshold and a division rule;

[0042] A sample extraction module, which is used to extract samples according to the non-overlapping principle in the normal section, the intermittent fault section, and the permanent fault section respectively to construct a FOG time series data set; when selecting intermittent fault samples, first determine the midpoint, and then intercept the FOG signals in equal-length sections before and after the midpoint as intermittent fault samples;

[0043] A training set construction module, which generates a candidate Shapelets set based on the FOG time series data set through a sliding time window, calculates the information gain of each candidate Shapelet, and performs optimal Shapelet screening based on the information gain to construct an optimal Shapelets set; calculates distance data based on the optimal Shapelets set and the FOG time series data set to generate a plurality of feature vectors, and constructs a training data set based on the plurality of feature vectors, and uses a clustering algorithm to obtain clustering centroids according to the training data set;

[0044] A model training module, which is used to construct a preliminary learning Shapelets model, and iteratively train the preliminary learning Shapelets model using the training data set and the clustering centroids to obtain a learning Shapelets model;

[0045] A classification and recognition module, which is used to obtain data to be recognized, input the data to be recognized into the learning Shapelets model, and output a classification result.

[0046] As can be seen from the above technical solutions, the present invention provides an intermittent fault feature recognition method and system for an optical fiber gyroscope. Compared with the prior art, it has the following beneficial effects:

[0047] 1. By determining the section segmentation points of the FOG signal, the original signal is segmented into independent intervals with clear state patterns, effectively reducing the feature extraction range. It not only avoids the computational redundancy of traditional global feature extraction, enables subsequent processing to focus on key state intervals, but also improves the computational efficiency while reducing the influence of noise interference on feature extraction.

[0048] 2. By converting the selected j Shapelets into vector inputs and constructing a Shapelets learning model, the effective classification of three types of FOG signal sections of the fiber optic gyroscope is realized, and the normal state, intermittent fault state, and permanent fault state of the fiber optic gyroscope are identified; it efficiently solves the problems of difficult extraction of intermittent fault characteristics of the fiber optic gyroscope and difficult distinction of three types of state signals, and improves the accuracy of identifying three types of state characteristics.

[0049] 3. The feature transformation method and K-Means initial clustering are adopted to not only reduce the computational complexity but also retain the signal features for identifying three types of states. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0051] Figure 1 It is a schematic flowchart of the method provided by the present invention;

[0052] Figure 2 It is a schematic diagram of the output Shapelets of the Shapelets learning model in the embodiment of the present invention;

[0053] Figure 3 It is a schematic diagram of the confusion matrix provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0055] One aspect of the embodiment of the present invention discloses an intermittent fault feature recognition method applied to a fiber optic gyroscope, as Figure 1 shown, and the specific steps are as follows:

[0056] Step 1: Obtain the Fiber Optic Gyroscope (FOG) signal and convert the FOG signal into zero-drift data;

[0057] Step 2: Extract valid data from the zero-drift data that meets the conditions that the initial point is below the threshold line and there are more than two intersections with the threshold line, and divide the valid data into a normal section, an intermittent fault section, and a permanent fault section according to the preset zero-drift threshold and the division rules;

[0058] Step 3: Extract samples according to the non-overlapping principle in the normal section, the intermittent fault section, and the permanent fault section respectively to construct the FOG time series dataset; when selecting intermittent fault samples, first determine the midpoint, and then intercept the FOG signals in the equal-length sections before and after the midpoint as the intermittent fault samples;

[0059] Step 4: Generate a candidate Shapelets set based on the FOG time series dataset through a sliding time window, calculate the information gain of each candidate Shapelet, and perform optimal Shapelets screening based on the information gain to construct an optimal Shapelets set; calculate the distance data based on the optimal Shapelets set and the FOG time series dataset to generate multiple feature vectors, and construct a training dataset based on the multiple feature vectors, and use a clustering algorithm to obtain the clustering centroids according to the training dataset;

[0060] Step 5: Construct a preliminary learning Shapelets model, and use the training dataset and the clustering centroids to perform iterative training on the preliminary learning Shapelets model to obtain a learning Shapelets model;

[0061] Step 6: Obtain the data to be recognized, input the data to be recognized into the learning Shapelets model, and output the classification result.

[0062] Further, collect the fiber optic gyro data. The diagnostic data for related faults is the FOG signal, and use the FOG signal as the diagnostic data type to reproduce the faults; collect the data, and use the formula in MATLAB to convert it into zero-drift data. In this embodiment, the zero-drift threshold of the fiber optic gyro is 0.1° / h; in the formula, B s represents the zero-drift, K represents the scale factor, n represents the number of sampling times, F f represents the output value of the f-th point, represents the output mean.

[0063] In one embodiment, the division rules are as follows: take the starting point of the valid data as the starting point of the normal section, the first intersection with the threshold line as the end point of the normal section, and the section between the starting point and the end point of the normal section is the normal section; take the last intersection with the threshold line as the starting point of the permanent fault section, the end point of the valid data as the end point of the permanent fault section, and the section between the starting point and the end point of the permanent fault section is the permanent fault section; the section between the end point of the normal section and the starting point of the permanent fault section is the intermittent fault section.

[0064] Furthermore, the starting point of each FOG signal serves as the starting point of the normal section. The first time the zero-drift threshold is reached, the corresponding FOG signal point is the end point of the normal section. The last time the zero-drift threshold is reached, the corresponding FOG signal point is the starting point of the permanent fault. The last point of this data serves as the end point of the permanent fault. The end point of the normal section serves as the starting point of the intermittent fault section, and the starting point of the permanent fault serves as the end point of the intermittent fault section.

[0065] Based on the starting and ending points of each section, it is determined that each piece of zero-drift data after conversion has more than two intersection points with the threshold line, and the initial point is below the threshold line. The data that meets the conditions is selected as valid data. For each piece of valid data, the valid data corresponding to the starting point and the first time the zero-drift threshold is reached is the normal section, and the valid data corresponding to the last time the zero-drift threshold is crossed to the last point of this piece of data is the permanent fault section. If each piece of data has a permanent fault section and a normal section, then the section between the end point of the normal section and the starting point of the permanent fault section is the intermittent fault section.

[0066] In one embodiment, the steps for obtaining intermittent fault samples are as follows:

[0067] Obtain the degradation time series data of the intermittent fault section and segment the degradation time series data through the threshold line;

[0068] Identify all the intersection points of the degradation time series data and the threshold line to form an intersection set;

[0069] Screen all the even-numbered interaction events in the intersection set and extract the critical positions where the degradation time series data corresponding to each event drops back to the zero-drift threshold;

[0070] Using each critical position as the midpoint, intercept equidistant data segments before and after in the FOG signal as intermittent fault samples.

[0071] Further, for normal data samples, their starting and ending points must be selected within the normal section, which is defined as the section where the original signal section is completely below the threshold line. In addition, there must be no overlap between the signal sections of normal data samples. For intermittent fault samples, all even interactions in a set of intersections between the degraded data and the threshold line are selected as the midpoints; the degraded data is obtained to fall back to the threshold point, and signal sections of equal length are taken before and after this point, and all intercepted data sections are used to construct intermittent fault samples. For permanent fault data samples, they must strictly belong to the permanent fault section. In addition, to ensure the independence between samples, the signal intervals of each data sample must be kept non-overlapping. After the samples are divided, conventional data cleaning is performed, including missing value compensation, outlier removal, sample smoothing, and label setting, to form the FOG time series dataset. In this embodiment, the FOG time series dataset is stored in a CSV file, and the complete dataset consists of 750 samples, including 300 normal samples, 300 intermittent fault samples, and 150 permanent fault samples.

[0072] In one embodiment, subsequences of the FOG time series dataset are obtained by means of a sliding time window to generate a Shapelets candidate set. According to the formula the information gain of all candidate Shapelets in the generated Shapelets candidate set is calculated and sorted, and the top 30% of the candidate Shapelets with the highest information gain are selected as the optimal Shapelets. The FOG time series data T i and the top 30% of the candidate Shapelets j adopt the formula to calculate the distance. In the formula, D i，j represents the minimum distance between the i-th time series and the j-th shapelet, L represents the length of the shapelet, T i，j+l-l represents the l-th point starting from position j in the i-th time series T i , and S j，l represents the l-th point of the j-th shapelet, to generate a feature vector and form the input feature set for learning Shapelets, and K-Means is used to obtain the clustering centroids of the training dataset.

[0073] In the formula, Ent is entropy; D is the total number of samples; D1 is the number of normal samples; D2 is the number of intermittent fault samples; D3 is the number of permanent fault samples; Ent1 is the entropy of the normal sample subset; Ent2 is the entropy of the intermittent fault sample subset; Ent3 is the entropy of the permanent fault sample subset.

[0074] In one embodiment, the steps for obtaining the learning Shapelets model are as follows:

[0075] Decompose the multi-classification problem into multiple binary-classification sub-problems, and construct sub-models for each binary-classification sub-problem respectively;

[0076] Use the clustering centroids as the initial parameters of the Shapelets of each sub-model, and iteratively train each sub-model in parallel using the training data set until the model converges to obtain the learned Shapelets model; during the training process, calculate the prediction loss of each sub-model through forward propagation, update the Shapelets and classification weights through backpropagation, and generate the final classification prediction result by combining the prediction results of all sub-models.

[0077] Furthermore, construct a learned Shapelets model, transform the three-classification problem into multiple binary-classification problems, classify each sub-problem, calculate the loss of each sub-problem, update the Shapelets and classification weights, learn a set of Shapelets with the most classification features from the time series data set, and use this classification as the final prediction.

[0078] Even further, in this embodiment, the state of the fiber optic gyro data involves three categories; transform this three-classification problem into three binary-classification problems; separate one category for each sub-problem and compare it with the other two categories. Among them, the binary label of class C is defined as in the formula In the formula, i represents the i-th time series sample, c represents the c-th category, d represents the decomposed label, and Y i ={1, 2, 3} represents the three-category labels, and the combined classification error and regularized multi-classification objective function are represented by the formula, In the formula, S represents the set of Shapelets, W represents the weight, and λ W represents the regularization coefficient, where 3 is the number of categories, L is the logistic loss between the true label and the predicted label, and for each category C, the predicted binary label of instance i is R is the number of scales, K is the number of shapes at each scale. In the formula, M r，i，k is used to measure the similarity between the Shapelets and the time series, and W c，r，k is used to represent the importance of the Shapelets in the classification process and is used to calculate the predicted target value. W c，0 is the bias term in the classification model. After the model is learned, the test instance indexed by t will be classified into the one-versus-one binary classifier that produces the maximum confidence. In Python, call the Tslearn toolkit to implement the learned Shapelets, set the regularization term to 0.01, the learning rate to 0.001, and the minimum length L minSet to 0.15, the scaling factor is set to 3, and the maximum step size is set to 1000; the training data set is imported into the classifier, and the 9 Shapelets results learned are as Figure 2 , these 9 Shapelets calculate the minimum distance with each time series to extract a set of discriminative feature vectors; these distance features are used as input and sent to the classifier in the algorithm. The model automatically learns the association relationship between Shapelets and different classes during training to achieve accurate discrimination of the three classes. For subsequent test time series, only the distance features need to be calculated with these 9 Shapelets and then input into the model to determine which category it belongs to, identifying the three states of the fiber optic gyro. The confusion matrix is used to quantify the classification results of the model for each category. The recognition results of normal state data, permanent fault state data, and intermittent fault state data are as Figure 3 described, and the classification accuracy is 82%.

[0079] In this embodiment, definitions of time series data set, Shapelets, entropy, information gain, sliding window, Shapelets transformation, etc. are introduced, where:

[0080] Time series data set: A set of data contains M signal samples, and the length of each signal sample is S. Therefore, the data set is represented as T M*S . The number of categories corresponding to the data set T is N. The digital labels from 1 to C are defined for each category. Therefore, the sequence target is a classification variable Y ∈ {1,..., C}, with a total of C categories.

[0081] Shapelet: This is a subsequence of the time series and also a part of the time series. Each Shapelet is an ordered time series data of length U. Moreover, U is smaller than S. It attempts to divide the time series data set into several different groups. The time series has multiple subsequences of different lengths, and among them, Shapelet is the most representative feature of the time series in the data.

[0082] Entropy: Assume that the data set D contains N time series, and the number of categories is C, where C i The proportion of the i-th class in the total N is p(C i i). Assuming there are three classes, 1, 2, and 3, the entropy of the data set D is Ent = -p(D1)log(p(D1)) - p(D2)log(p(D2)) - p(D3)log(p(D3)).

[0083] Information gain: Represents the classification effect of a certain Shapelet, denoted as

[0084] Sliding Window: A Shapelet is a part of a time series. To compare the Shapelet with the original sequence, a sliding window needs to be defined. The length of the sliding window segment is the same as that of the Shapelet. The sliding step size is set to 1. The segment corresponding to each sliding window segment is the time series to be compared. The time series dataset contains J segments, where J = S - U + 1.

[0085] Shapelet Transformation: The minimum Euclidean distance between a Shapelet and the best subsequence represents similarity. As a representative feature in the Shapelet feature space, by measuring the similarity between the time series and the Shapelet, a new feature space is constructed as the input for the subsequent classifier. By calculating the Shapelet and comparing it with the sliding window segment, the distance between the i-th sequence T i and the j-th Shapelet j is represented as

[0086] On the other hand, this embodiment discloses an intermittent fault feature recognition system applied to a fiber optic gyroscope, including:

[0087] Data acquisition and conversion module, used to acquire FOG signals and convert the FOG signals into zero drift data;

[0088] Data optimization and segmentation module, used to extract valid data in the zero drift data that satisfies the initial point being lower than the threshold line and having more than two intersections with the threshold line, and divide the valid data into a normal section, an intermittent fault section, and a permanent fault section according to the preset zero drift threshold and the division rule;

[0089] Sample extraction module, used to extract samples according to the non-overlapping principle in the normal section, the intermittent fault section, and the permanent fault section respectively to construct a FOG time series dataset; among them, when selecting intermittent fault samples, first determine the midpoint, and then intercept the FOG signals of equal length sections before and after the midpoint as intermittent fault samples;

[0090] Training set construction module, generates a candidate Shapelets set based on the FOG time series dataset through a sliding time window, calculates the information gain of each candidate Shapelet, and performs optimal Shapelet screening based on the information gain to construct an optimal Shapelets set; calculates distance data based on the optimal Shapelets set and the FOG time series dataset, generates multiple feature vectors, and constructs a training dataset based on the multiple feature vectors, and uses a clustering algorithm to obtain clustering centroids according to the feature vectors;

[0091] A model training module for constructing a preliminary learning Shapelets model and iteratively training the preliminary learning Shapelets model using feature vectors and clustering centroids to obtain a learning Shapelets model;

[0092] A classification and recognition module for obtaining data to be recognized, inputting the data to be recognized into the learning Shapelets model, and outputting a classification result.

[0093] In the present specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0094] 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 obvious to those skilled in the art. The general principles defined herein can 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 will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intermittent fault feature recognition method applied to a fiber optic gyroscope, characterized in that The specific steps are as follows: Obtain the FOG signal and convert the FOG signal into zero-drift data; Extract valid data in the zero-drift data that satisfies the condition that the initial point is lower than the threshold line and there are more than two intersection points with the threshold line, and divide the valid data into a normal section, an intermittent fault section, and a permanent fault section according to a preset zero-drift threshold and a division rule; Extract samples according to the non-overlapping principle in the normal section, the intermittent fault section, and the permanent fault section respectively to construct a FOG time series data set; among them, when selecting intermittent fault samples, first determine the midpoint, and then intercept the FOG signals in equal-length sections before and after the midpoint as intermittent fault samples; Generate a candidate Shapelets set based on the FOG time series data set through a sliding time window, calculate the information gain of each candidate Shapelets, and perform optimal Shapelets screening based on the information gain to construct an optimal Shapelets set; calculate distance data based on the optimal Shapelets set and the FOG time series data set, generate a plurality of feature vectors, and construct a training data set based on the plurality of feature vectors, and use a clustering algorithm to obtain clustering centroids according to the training data set; Construct a preliminary learning Shapelets model, and use the training data set and the clustering centroids to iteratively train the preliminary learning Shapelets model to obtain a learning Shapelets model; Obtain the data to be recognized, input the data to be recognized into the learning Shapelets model, and output a classification result.

2. The intermittent fault feature recognition method for an optical fiber gyroscope according to claim 1, characterized in that, The conversion expression of the zero-drift data is: Where B s represents zero drift, K represents the scale factor, n represents the number of sampling times, and F f represents the output value of the f-th point, represents the output mean value.

3. The intermittent fault feature recognition method for an optic fiber gyroscope according to claim 1, wherein The division rule is: taking the starting point of the valid data as the starting point of the normal section, the intersection point where it first intersects with the threshold line as the end point of the normal section, and the section between the starting point and the end point of the normal section as the normal section; taking the intersection point where it intersects with the threshold line for the last time as the starting point of the permanent fault section, the end point of the valid data as the end point of the permanent fault section, and the section between the starting point and the end point of the permanent fault section as the permanent fault section; the section between the end point of the normal section and the starting point of the permanent fault section is the intermittent fault section.

4. The intermittent fault feature recognition method applied to a fiber optic gyroscope according to claim 1, wherein The steps for obtaining the intermittent fault samples are as follows: Obtain the degradation time series data of the intermittent fault section and segment the degradation time series data through the threshold line; Identify all intersection points of the degradation time series data and the threshold line to form an intersection set; Screen all even-numbered interaction events in the intersection set and extract the critical positions where the degradation time series data corresponding to each event drops back to the zero-drift threshold; Taking each critical position as the midpoint, intercept equidistant data segments before and after in the FOG signal as the intermittent fault samples.

5. A method for identifying intermittent fault characteristics applied to a fiber optic gyroscope according to claim 1, characterized in that, The calculation formula for the information gain is: Where Ent is entropy; D is the total number of samples; D1 is the number of normal samples; D2 is the number of intermittent fault samples; D3 is the number of permanent fault samples; Ent1 is the entropy of the normal sample subset; Ent2 is the entropy of the intermittent fault sample subset; Ent3 is the entropy of the permanent fault sample subset.

6. The intermittent fault feature recognition method applied to an optical fiber gyroscope according to claim 1, wherein The distance data calculation formula is: Where D i，j represents the minimum distance between the i-th time series and the j-th shapelet, and T i，j+l-1 represents the l-th point starting from position j in the i-th time series T i and S j，l represents the l-th point of the j-th shapelet.

7. The intermittent fault feature recognition method for an optical fiber gyroscope according to claim 1, wherein The obtaining steps of the learning Shapelets model are: Decompose the multi-classification problem into multiple binary classification sub-problems, and construct sub-models for each binary classification sub-problem respectively; Use the clustering centroids as the initial parameters of the Shapelets of each sub-model, and use the training data set to perform parallel iterative training on each sub-model until the model converges to obtain the learning Shapelets model; during the training process, calculate the prediction loss of each sub-model through forward propagation, update the Shapelets and classification weights through backpropagation, and generate the final classification prediction result by combining the prediction results of all sub-models.

8. A method for identifying intermittent fault characteristics applied to a fiber optic gyroscope according to claim 7, characterized in that, The expression of the classification objective function is: Wherein, S represents the set of shapelets; W represents the weight; F represents the output data; i represents the i-th time series sample; c represents the c-th category; L represents the logical loss between the true label and the predicted label; represents the true label; represents the predicted label; λ W represents the regularization coefficient.

9. An intermittent fault feature recognition method applied to a fiber optic gyroscope according to claim 7, characterized in that, The expression of the prediction result of the sub-model is: where, W c，0 represents the bias term in the sub-model; R represents the number of scales; K represents the number of shapes at each scale; M r，i，k is the similarity between the Shapelet and the time series, and W c，r，k is the importance of the Shapelet for classification.

10. An intermittent fault feature recognition system applied to a fiber optic gyroscope, characterized in that, Including: A data acquisition and conversion module, which is used to acquire the FOG signal and convert the FOG signal into zero-drift data; A data optimization and segmentation module, which is used to extract valid data that meets the condition that the initial point is lower than the threshold line and has more than two intersections with the threshold line from the zero-drift data, and divide the valid data into a normal section, an intermittent fault section, and a permanent fault section according to the preset zero-drift threshold according to the division rule; A sample extraction module, which is used to extract samples according to the non-overlapping principle in the normal section, the intermittent fault section, and the permanent fault section respectively to construct a FOG time series data set; among them, when selecting intermittent fault samples, first determine the midpoint, and then intercept the FOG signals in the equal-length sections before and after the midpoint as intermittent fault samples; A training set construction module, which generates a candidate Shapelets set based on the FOG time series data set through a sliding time window, calculates the information gain of each candidate Shapelet, and performs optimal Shapelet screening based on the information gain to construct an optimal Shapelets set; calculates the distance data based on the optimal Shapelets set and the FOG time series data set, generates multiple feature vectors, and constructs a training data set based on the multiple feature vectors, and uses a clustering algorithm to obtain the clustering centroids according to the training data set; A model training module, which is used to construct a preliminary learning Shapelets model, and use the training data set and the clustering centroids to perform iterative training on the preliminary learning Shapelets model to obtain the learning Shapelets model; A classification and recognition module, which is used to obtain the data to be recognized, input the data to be recognized into the learning Shapelets model, and output the classification result.

Images