CN112668105B - Helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance - Google Patents
Helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance Download PDFInfo
- Publication number
- CN112668105B CN112668105B CN202110047086.3A CN202110047086A CN112668105B CN 112668105 B CN112668105 B CN 112668105B CN 202110047086 A CN202110047086 A CN 202110047086A CN 112668105 B CN112668105 B CN 112668105B
- Authority
- CN
- China
- Prior art keywords
- health
- baseline
- real
- transmission shaft
- vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Abstract
A helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance realizes the construction of a health baseline of a helicopter transmission shaft, the generation of a threshold value and abnormity judgment by utilizing the monitoring vibration data of the helicopter transmission shaft. Firstly, monitoring vibration data by using a helicopter transmission shaft, obtaining an input vector through FFT (fast Fourier transform) spectrum transformation and RMS (root mean square) compression transformation processing, and constructing a health baseline of the transmission shaft by using an SAE (adaptive sampling and analysis) model, wherein the health baseline comprises the SAE model after training and a health characteristic set output by the SAE model. Secondly, based on the overall distribution mean vector and covariance matrix of the health characterization set, the Mahalanobis distance between the health characterization set and each health vector in the health characterization set is calculated to generate a baseline statistical threshold of the helicopter transmission shaft, and the self-adaptive characterization of the normal and abnormal state quantitative distinguishing standard is realized. Thirdly, based on the health characterization set and the baseline statistical threshold, threshold abnormity judgment is carried out on the real-time test data sample, and real-time detection of the helicopter transmission shaft state is achieved.
Description
Technical Field
The invention belongs to the technical field of helicopter flight control, and particularly relates to a helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance.
Background
The transmission system is one of three key moving parts of the helicopter, and the transmission shaft is an important component of the transmission system. Once the transmission shaft is out of order, the helicopter is seriously affected and even fatal accidents can be caused. Therefore, the method has great significance for monitoring the development state of the helicopter transmission shaft. At present, most of the state monitoring methods for helicopter transmission shafts focus on threshold detection based on expert experience, namely, fault sensitive features are extracted by using a signal analysis means, and then expert thresholds are set to realize detection. The method not only needs abundant expert knowledge background, but also has more false alarm and false alarm missing problems in actual engineering. In addition, for a general shaft component, many intelligent fault diagnosis methods combining signal analysis and a machine learning algorithm have appeared in recent years, but such methods generally require a large amount of abnormal data, which is often difficult to obtain in practical engineering.
Aiming at the problems, the invention provides a healthy baseline construction and threshold generation method based on a Stacked auto-encoder (SAE for short) and Mahalanobis distance, which can realize the autonomous characterization of the healthy state of the helicopter transmission shaft and the generation of threshold parameters by only using normal data or combining a small amount of abnormal data, thereby completing the judgment of the abnormal state of the transmission shaft in practical engineering application.
Disclosure of Invention
The application aims to provide a helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance, so that the health state representation of the helicopter transmission shaft can be effectively constructed and an abnormity judgment threshold value can be generated under the condition of abnormal data shortage, and abnormity judgment of the helicopter transmission shaft is realized.
The invention provides a helicopter transmission shaft health baseline construction, threshold generation and abnormality judgment method based on SAE and Mahalanobis distance, which comprises the following steps:
the first step is as follows: constructing a health baseline by using the training data;
the second step is that: generating a baseline statistical threshold using the training data;
the third step: and carrying out real-time abnormity judgment aiming at the test data.
Preferably, the training data used for the healthy baseline construction and the threshold generation in the first step and the second step may be only normal data, or a small amount of abnormal data is introduced on the basis of the normal data, so as to further optimize a healthy baseline model and improve the accuracy of abnormal judgment.
Preferably, the method for constructing a healthy baseline in the first step includes performing FFT spectral transformation on training data, performing RMS compression transformation to construct an input vector, and finally completing construction of a healthy baseline by using the input vector of the training data.
Preferably, the health baseline includes an SAE model and a health status feature set, wherein the SAE model is trained by using training data, the trained SAE model can implement nonlinear and adaptive transformation of a high-dimensional input vector, and generate a health vector, and further, the health vectors obtained from normal data in all the training data form the health feature set, so as to form an adaptive quantitative representation of the health status.
Preferably, the second step of the baseline statistics threshold generation method includes adaptive generation of a distribution-sample mahalanobis distance metric and a threshold, where the distribution-sample mahalanobis distance metric calculates mahalanobis distances between the overall distribution of the health feature set and the health vectors of the training samples, and implements a quantitative metric of individual differences; and further, performing statistical distribution calculation on the generated Mahalanobis distance sequence, and combining a statistical principle to realize self-adaptive generation of the baseline threshold.
Preferably, the real-time anomaly determination method in the third step includes performing FFT spectral transformation on the test data, then performing RMS compressive transformation to construct an input vector, obtaining a real-time mahalanobis distance of the test data input vector based on the healthy baseline, and if the real-time mahalanobis distance exceeds a baseline statistical threshold, determining that the current state of the helicopter transmission shaft is abnormal, otherwise, determining that the current state of the helicopter transmission shaft is normal.
Preferably, in the method for obtaining the real-time mahalanobis distance of the test data based on the healthy baseline, the healthy baseline comprises an SAE model and a healthy feature set which are finished by training, and the test data is input into a vector and sent to the SAE model to obtain a real-time state vector; and further calculating the mahalanobis distance between the real-time state vector and the health characteristic set, thereby obtaining the real-time mahalanobis distance.
The invention has the advantages and positive effects that:
(1) by the SAE-based health baseline construction method, the health characterization vector of the helicopter transmission shaft can be extracted in a self-adaptive manner through deep nonlinear transformation under the condition of depending on prior knowledge as little as possible, and the method has strong universality and expansibility;
(2) aiming at the practical problems that the historical abnormal data of a helicopter transmission shaft is deficient and the abnormal judgment threshold is set too subjectively, a self-adaptive baseline statistical threshold generation method is provided, and a more objective and self-adaptive judgment standard is provided for judging the normal/abnormal state of the helicopter;
(3) the method fully considers the practical characteristic of the shortage of the abnormal samples of the helicopter transmission shaft, can finish the construction of the health baseline and the generation of the threshold value by only utilizing normal data, and further realizes effective abnormal judgment; on the basis of the normal data, if part of abnormal data is introduced, the construction effect of the health baseline and the accuracy of abnormal judgment can be further improved, and the current practical engineering problem can be effectively solved.
Drawings
FIG. 1 is a block flow diagram of the present invention;
FIG. 2 is a diagram of an original vibration signal according to an embodiment;
FIG. 3 is a diagram of FFT spectral transformation of training data according to an embodiment;
FIG. 4 is a diagram of an embodiment of a training data RMS compressional transformation process;
FIG. 5 is a graph illustrating the loss of training according to one embodiment;
FIG. 6 is a scatter plot visualization of a health characterization set, under an embodiment;
FIG. 7 is a test data spectrum transform of an embodiment;
FIG. 8 is a test data RMS compressional transformation diagram according to an embodiment.
FIG. 9 is a visualization of a real-time state vector scatter plot according to an embodiment;
FIG. 10 is a diagram illustrating an exception determination result of test data according to an embodiment;
FIG. 11 is a diagram of the original vibration signal of the second embodiment;
FIG. 12 is a diagram of a spectral transformation of training data according to the second embodiment;
FIG. 13 is a diagram of RMS packed transform processing for training data according to an embodiment two;
FIG. 14 is a graph of a loss chart of training according to the second embodiment;
FIG. 15 is a scattergram visualization of the health characterization of the second embodiment;
FIG. 16 is a graph of the spectrum transformation of test data according to the second embodiment;
FIG. 17 is a RMS compressional transform plot of test data from example two;
FIG. 18 is a scattergram visualization of real-time condition features according to an embodiment two;
FIG. 19 is a diagram illustrating an abnormal determination result of test data according to the second embodiment.
Detailed Description
The invention provides a helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance. The method comprises the following steps:
step one, healthy baseline construction S1: based on all normal data of the helicopter transmission shaft or a large amount of normal data and a small amount of abnormal data, a helicopter transmission shaft health baseline construction method is provided, and self-adaptive representation of the health state of the helicopter transmission shaft is realized. The method comprises the following specific steps:
101: the method comprises the steps of obtaining helicopter transmission shaft monitoring parameter samples (N) to form training data, wherein the training data can only comprise normal data or comprise a large amount of normal data and a small amount of abnormal data. The monitoring parameters are as follows: a time domain vibration signal;
102: and transforming the time domain vibration signal samples into frequency spectrum signals. Wherein, the frequency spectrum transformation method adopts Fourier transform (FFT);
103: and then, RMS compression transformation processing is carried out on the transformed high-dimensional frequency domain characteristic signals to obtain relatively low-dimensional input vectors, so that the overfitting risk of the healthy baseline model is reduced, and the construction quality of the model is improved. The key links comprise:
103-1: giving a frequency domain characteristic signal, and segmenting the frequency domain characteristic signal by using a sliding window with a certain length to obtain a local frequency spectrum signal;
103-2: respectively calculating the root mean square value (RMS) of each local frequency spectrum signal, and splicing to obtain an input vector after compression transformation;
104: the method comprises the steps of utilizing an input vector sample obtained by converting helicopter transmission shaft training data to input the input vector sample into a stacked self-encoder model (SAE), obtaining a trained SAE model and M health vectors (M is less than or equal to N) obtained by normal samples in the training data, and taking the health vectors as a health baseline for representing the helicopter transmission shaft. The key links comprise:
104-1: building an SAE model, wherein the model comprises a coding layer, a characteristic layer and a decoding layer;
104-2: for configuring the SAE model, the model configuration items include: the method comprises the following steps of (1) health vector dimension, model layer number, Dropout proportion, batch normalization dimension, model training round, learning rate, optimizer and loss function; the health vector dimension is the dimension of the vector after the SAE model coding layer and the feature layer are transformed.
104-3: and inputting training data into the constructed and configured SAE model, performing model training, and terminating the training when the training turns reach a preset configuration value.
104-4: adjusting the trained SAE model: and reserving the coding layer and the characteristic layer and removing the decoding layer. And sequentially inputting the N input vectors in the training data to the adjusted SAE model, and obtaining corresponding N output vectors after conversion of the coding layer and the characteristic layer. Selecting parts (M) obtained by training data normal samples from the N output vectors, wherein the M output vectors of the normal data are the representation of the health state of the transmission shaft of the helicopter, namely the health vectors; the M sets of health vectors derived from the training data are the health feature sets.
105-5: the adjusted SAE model and health feature set are collected as a health baseline of the helicopter health status.
Step two, generating a baseline statistical threshold value S2: based on the constructed healthy baseline, a method for generating a baseline statistical threshold of a transmission shaft of the helicopter is provided, and adaptive characterization of normal and abnormal state quantitative distinguishing standards is realized. The method comprises the following specific steps:
201: and (4) calculating the whole distribution mean vector and covariance matrix of the health characteristic set in the step one, and recording the vectors as U and C. The dimensionality of the vector U and the dimensionality of the square matrix C are consistent with the 'health vector dimensionality' configured in the step one;
202: sequentially calculating Mahalanobis distances between the distribution mean vector U and each health vector in the health characterization set by using the distribution mean vector U and the distribution covariance matrix C to obtain M baseline Mahalanobis distances, wherein the distance set describes the fluctuation condition of the distance result in a normal state;
203: and generating a helicopter transmission shaft baseline statistical threshold value by utilizing a statistical 6 sigma principle based on the M baseline Mahalanobis distances. The calculation method is as follows:
T＝μ+6σ
wherein mu is the mean value of the base line Mahalanobis distance, sigma is the standard deviation value of the base line Mahalanobis distance, and T is the base line statistical threshold. The baseline statistical threshold value T is a judgment standard for distinguishing the normal/abnormal states of the helicopter.
Step three, real-time anomaly determination S3: based on the constructed health baseline and given test data, the method for judging the real-time abnormity of the transmission shaft of the helicopter is provided, and the automatic real-time judgment of the state of the transmission shaft is realized. The method comprises the following specific steps:
301: and acquiring real-time test data. The monitoring parameter category is the same as that of the first step;
302: repeating the steps 102 and 103 on each vibration signal sample in the test data to obtain a real-time test data input vector;
303: and (3) sending the real-time test data input vector to the trained and adjusted SAE model obtained in the step one, and obtaining an output vector converted by a coding layer and a feature layer, namely a real-time state vector. The real-time state vector is a quantitative representation of the real-time state of the transmission shaft of the helicopter;
304: and (4) judging the threshold abnormity of the real-time state vector based on the health characteristic set obtained in the step one and the baseline statistical threshold obtained in the step two, so as to realize the real-time detection of the state of the transmission shaft of the helicopter. The key links comprise:
304-1: mahalanobis distance measurement, which is to calculate the distribution mean vector U and the distribution covariance matrix C of the health feature set (as described in step 201), and use U, C to calculate mahalanobis distance between the health feature set and the real-time state vector sample, thereby realizing quantitative difference measurement between the real-time state and the historical health state;
304-2: and D, judging the threshold value abnormity, namely judging whether the Mahalanobis distance of the real-time state vector exceeds the threshold value or not based on the baseline statistical threshold value obtained in the step two, if so, judging that the real-time state is abnormal, otherwise, judging that the real-time state is normal.
The method for judging the abnormity of the helicopter transmission shaft based on SAE and Mahalanobis distance is provided.
DETAILED DESCRIPTION OF EMBODIMENT (S) OF INVENTION
The helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance is shown in figure 1. FIG. 1 is a block flow diagram of the present invention.
Based on the 2-type typical data scenario (including normal data only, large amount of normal data and small amount of abnormal data) which is common in practice, the present invention provides two sets of embodiments, and other advantages and effects of the present invention can be easily understood by those skilled in the art from the description of the present specification. The first and second embodiments are described below.
Detailed description of the preferred embodiment
1 healthy Baseline construction
101: aiming at the common actual scene with only normal data, only helicopter transmission shaft monitoring parameter samples in a normal state are selected to form training data, and vibration signals are adopted as helicopter transmission shaft state representation signals in the invention. Specifically, the details of the training data of example one are shown in table 1.
Table 1: embodiment one training data detail table
| Training data composition | For scenes |
| 199 Normal vibration Signal | Only normal data exists, and abnormal data hardly exists |
As is clear from the visualization result shown in fig. 2, the original vibration signal exhibits high-frequency characteristics, and it is difficult to directly determine the actual state of the object, and therefore, it is necessary to further perform conversion processing. FIG. 2 illustrates a graph of a raw vibration signal.
102: the time domain vibration signal samples are transformed through FFT to obtain frequency spectrum signals, the visualization result is shown in figure 3, and it can be known from the figure that frequency domain characteristic information reflecting the object state is more prominent after the frequency spectrum transformation. Fig. 3 is a diagram of an FFT spectrum transform of training data.
103: and then, RMS compression transformation processing is carried out on the transformed high-dimensional frequency domain characteristic signals to obtain relatively lower-dimensional input vectors. The method comprises the following specific steps:
103-1: giving a frequency domain characteristic signal, and segmenting the frequency domain characteristic signal by using a sliding window to obtain a local frequency domain characteristic signal;
103-2: and for each local frequency domain characteristic signal, selecting a compression characteristic dimension of 512, and compressing the frequency spectrum characteristic below 12000Hz into a 512 multiplied by 1 RMS compression transformation vector. As shown in fig. 4, it can be known that, in the input vector after the feature compression, the signal characteristics are better preserved, the redundant components are compressed and removed, and an RMS compression transformation processing diagram of the training data in the embodiment of fig. 4 is shown.
104: the helicopter transmission shaft training data input vector obtained by the steps is input into an SAE model, the SAE model is trained to obtain a trained SAE model and a health vector, and the trained SAE model and the health vector are used as a health baseline for representing the helicopter transmission shaft, and the steps comprise:
104-1: an SAE model is built, the model comprises an encoding layer, a feature layer and a decoding layer, and the specific architecture details are shown in Table 2.
Table 2: SAE model Structure setup Table
104-2: the specific information of SAE model related configuration, health vector dimension, number of model layers, Dropout proportion, batch normalization dimension, model training round, learning rate, optimizer and loss function, is shown in Table 3:
table 3: SAE model related configuration table
104-3: inputting training data into the constructed and configured SAE model, performing model training, and when the training round is 350, the loss function reaches a preset configuration value, terminating the training, and visualizing the model training loss as shown in FIG. 5. As can be seen from the figure, as the training round increases, the model training loss continuously decreases, indicating that the model training is effective. Fig. 5 is a graph illustrating a training loss profile according to an embodiment.
104-4: and reserving the coding layer and the characteristic layer of the trained SAE model, and removing the decoding layer. And sequentially inputting the input vectors in the training data into the adjusted SAE model, and obtaining corresponding output vectors after conversion of the coding layer and the characteristic layer. The vector set of each output is a representation of the health state of the helicopter transmission shaft, namely a health vector set, the health vector dimension is set to be 2, and a scatter diagram of the health representation set is shown in fig. 6. FIG. 6 illustrates a health characterization set scatter plot visualization.
105-5: the adjusted SAE model and health feature set are collected as a health baseline of the helicopter health status. Since the training data in this embodiment is composed of all normal samples, the number of health vectors in the health characterization set is consistent with the total number of samples, i.e. 199.
2 Baseline threshold Generation
201: and (4) selecting the health characteristic set in the step one, and calculating the overall distribution mean vector and covariance matrix of the health characteristic set, and recording the vectors as U and C. The dimensionality of the vector U and the dimensionality of the square matrix C are consistent with the 'health vector dimensionality' configured in the step one;
202: sequentially calculating the Mahalanobis distance between the distribution mean vector U and each health vector in the health characterization set by using the distribution mean vector U and the distribution covariance matrix C as a baseline distance for describing the baseline fluctuation condition of the distance result in the normal state of the helicopter;
203: and generating a baseline statistical threshold value of the transmission shaft of the helicopter by utilizing a statistical 6 sigma principle based on the baseline Mahalanobis distance. The calculation method is as follows:
T＝μ+6σ
the mean value μ of the baseline mahalanobis distance is 1.2143, the standard deviation σ of the baseline mahalanobis distance is 0.7179, and the baseline statistical threshold T is 5.5214. And (4) taking the baseline statistical threshold T as a judgment standard for distinguishing the normal/abnormal states of the helicopter.
3 real-time anomaly determination
301: real-time test data were obtained, the details of which are shown in table 4:
table 4: test data detail table of embodiment one
| Number of normal samples | Number of |
| 200 | 135 |
302: the operations in steps 102 and 103 are repeated for each vibration signal sample in the test data to obtain a real-time input vector, and the visualization result after fourier transform and RMS compression is shown in fig. 7 and 8, which shows that the signal characteristics of the transformed input vector are more prominent, i.e., the test data spectrum transformation in the embodiment of fig. 7, and the test data RMS compression transformation in the embodiment of fig. 8.
303: and (3) inputting the test data in real time into the trained and adjusted SAE model obtained in the step one, and obtaining an output vector converted by a coding layer and a characteristic layer as a quantitative representation of the real-time state of the transmission shaft of the helicopter, wherein a real-time state representation set scatter diagram is shown in FIG. 9. FIG. 9 illustrates a real-time state vector scattergram visualization.
304: and (4) judging the abnormal threshold of the real-time test data sample based on the health characteristic set obtained in the step one and the baseline statistical threshold obtained in the step two, so as to realize the real-time detection of the state of the transmission shaft of the helicopter. The method comprises the following steps:
304-1: mahalanobis distance measurement, calculating the distribution mean vector U and the distribution covariance matrix C of the health feature set (as described in step 201), and using the U, C to calculate the mahalanobis distance between the health feature set and the test state vector samples, so as to realize the quantitative difference measurement between the real-time state vector and the historical health feature set, where the mahalanobis distance between the normal sample state vector and the abnormal sample state vector in the test data is shown in fig. 10: as can be seen from the graph, the distance values of the normal samples are all below the baseline statistical threshold, and the distance values of the abnormal samples are almost all above the baseline statistical threshold, so that the effectiveness of the method is proved. FIG. 10 shows an abnormal determination result of test data according to an embodiment.
304-2: and D, judging the threshold value abnormity, namely judging whether the Mahalanobis distance of the real-time test sample exceeds the threshold value or not based on the baseline statistical threshold value obtained in the step two, if so, judging that the real-time state is abnormal, otherwise, judging that the real-time state is normal. The anomaly determination is carried out based on the method, the determination result is displayed in a confusion matrix form, the result is shown in table 5, and the accuracy calculation formula is shown as follows:
table 5: example I quantitative results table for anomaly determination
Detailed description of the invention
1 healthy Baseline construction
101: aiming at the common scenes that the helicopter transmission shaft has more normal data volume and less abnormal data volume in practice, a large number of normal-state helicopter transmission shaft monitoring parameter samples and a small number of abnormal-state monitoring parameter samples are selected to form training data, and vibration signals are used as helicopter transmission shaft state representation signals. Specifically, the details of the training data in example two are shown in table 6.
Table 6: example two training data detail sheet
| Training data composition | Simulation scene |
| 199 Normal sample +14 abnormal sample | There is a large amount of normal data, only a small amount of abnormal data |
A certain abnormal sample is selected from the training data, and the visualization result is shown in fig. 11: as is clear from the figure, the original vibration signal exhibits high-frequency characteristics, and it is difficult to directly determine the actual state of the object, and therefore, further conversion processing is required. Fig. 11 illustrates the second example of the original vibration signal.
102: the time-domain vibration signal samples are transformed into frequency spectrum signals through FFT, and the visualization result is shown in fig. 12: as can be seen from the figure, the frequency domain characteristic information reflecting the state of the object is more prominent after the frequency spectrum conversion. Fig. 12 illustrates the spectral transformation of the training data according to the second embodiment.
103: and then, carrying out compression conversion processing on the converted high-dimensional frequency spectrum signal to obtain an input vector with relatively low dimensionality. The method comprises the following specific steps:
103-1: giving a frequency domain characteristic signal, and segmenting the frequency domain characteristic signal by using a sliding window to obtain a local frequency domain characteristic signal;
103-2: for each local frequency domain characteristic signal, selecting a compression characteristic dimension of 512, and compressing the frequency spectrum characteristic below 12000Hz to be an RMS compression transformation vector of 512 multiplied by 1, namely an input vector. As shown in fig. 13, it can be seen that the input vector after feature compression retains the signal characteristics better and the redundant components are compressed and removed. FIG. 13 illustrates an embodiment of a two training data RMS compressional transform process.
104: all input vector samples of the helicopter transmission shaft training data obtained by the steps are input into a stacked self-encoder model (SAE), an SAE model is trained to obtain a trained SAE model and health vectors which are used as a health baseline for representing the helicopter transmission shaft, and the steps comprise:
104-1: building an SAE model, wherein the information of a coding layer, a characteristic layer and a decoding layer of the model is the same as that of the first embodiment;
104-2: SAE model related configuration-health vector dimension, number of model layers, Dropout proportion, batch normalization dimension, model training round, learning rate, optimizer and loss function are the same as those of the first embodiment;
104-3: inputting training data into the constructed and configured SAE model, performing model training, terminating the training when the training round is 350 and the loss function reaches a preset configuration value, visualizing the model training loss as shown in FIG. 14, wherein the model training loss is known to continuously decrease along with the increase of the training round, which indicates that the model training is effective. FIG. 14 is a graph of loss from training for example two.
104-4: and reserving the coding layer and the characteristic layer of the trained SAE model, and removing the decoding layer. And sequentially inputting the input vectors in the training data into the adjusted SAE model, and obtaining corresponding output vectors after conversion of the coding layer and the characteristic layer. The vector set of the outputs is a representation of the health state of the transmission shaft of the helicopter, namely a health vector set, the health representation dimension is set to be 2, a scatter diagram of the health representation set is shown in fig. 15, and the two health representation scatter diagrams in the embodiment of fig. 15 are visualized.
105-5: and collecting the adjusted SAE model and health sign set as a health baseline of the health state of the helicopter. The training data in this embodiment includes 199 normal samples and 14 abnormal samples, and only the health vectors obtained by transforming the normal samples are selected as the health feature set, that is, the number of the health vectors in the health feature set in this embodiment is 14.
2 Baseline statistical threshold Generation
201: and (4) selecting the health characteristic set in the step one, and calculating the overall distribution mean vector and covariance matrix of the health characteristic set, and recording the vectors as U and C. The dimensionality of the vector U and the dimensionality of the square matrix C are consistent with the 'health vector dimensionality' configured in the step one;
202: and sequentially calculating the Mahalanobis distance between the distribution mean vector U and each health vector in the health characterization set by using the distribution mean vector U and the distribution covariance matrix C as a baseline distance for describing the baseline fluctuation condition of the distance result in the normal state of the helicopter.
203: and generating a baseline statistical threshold value of the transmission shaft of the helicopter by utilizing a statistical 6 sigma principle based on the baseline Mahalanobis distance. The calculation method is as follows:
T＝μ+6σ
the mean value μ of the baseline mahalanobis distance is 1.2552, the standard deviation σ of the baseline mahalanobis distance is 0.6438, and the baseline statistical threshold T is 5.1179. And (4) taking the baseline statistical threshold T as a judgment standard for distinguishing the normal/abnormal states of the helicopter.
3 real-time anomaly determination
301: real-time test data are obtained, and the test data in the second embodiment are consistent with the test data in the first embodiment;
302: the operations in steps 102 and 103 are repeated for each vibration signal sample in the test data to obtain a real-time compression transformation input vector, and the visualization after fourier transformation and RMS compression is as shown in fig. 16 and 17: as can be seen, the signal characteristics of the transformed input vector are more prominent. The second test data spectrum transformation of the embodiment of fig. 16, and the second test data RMS compression transformation of the embodiment of fig. 17.
303: and (3) sending the real-time input vector into the SAE model obtained in the step one after training and adjustment, and obtaining an output vector converted by a coding layer and a characteristic layer as a quantitative representation of the real-time state of the helicopter transmission shaft, namely a real-time state vector. The scatter plot is shown in fig. 18: as can be seen from the figure, the real-time state characterization scatter plot distinguishing degree in the normal state and the abnormal state is high, the overlapping degree is low, and the validity of the state characterization result is preliminarily proved. FIG. 18 is a two-embodiment real-time state feature scattergram visualization.
304: and (4) judging the abnormal threshold of the real-time test data based on the health characteristic set obtained in the step one and the baseline statistical threshold obtained in the step two, so as to realize the real-time detection of the state of the transmission shaft of the helicopter. The method comprises the following steps:
304-1: mahalanobis distance metric, which is a mean vector U and a covariance matrix C of the distribution of the health signature set (as described in step 201), is calculated, and the mahalanobis distance between the health signature set and the vector sample of the test state is calculated using U, C, thereby realizing a quantitative difference metric between the real-time state and the historical health state. The difference measurement results are shown in fig. 19: as can be seen from the figure, the distance values of the normal samples are all below the baseline statistical threshold, and the distance values of the abnormal samples are all above the baseline statistical threshold, which proves the effectiveness of the method, and the result of abnormal determination of the second test data in the embodiment of fig. 19 is shown.
304-2: and D, judging the threshold value abnormity, namely judging whether the Mahalanobis distance of the real-time state vector exceeds the threshold value or not based on the baseline statistical threshold value obtained in the step two, if so, judging that the real-time state is abnormal, otherwise, judging that the real-time state is normal. The abnormality judgment is carried out based on the method, the judgment result is displayed in a confusion matrix form, the result is shown in table 7, and the accuracy calculation formula is as follows:
table 7: example two quantitative results table for abnormality determination
As shown by the confusion matrix result, in this embodiment, after a small number of abnormal samples (14) are introduced to construct training data based on normal samples (199), the accuracy of the method provided by the present invention on the same test data reaches 100%. Compared with the first embodiment (training data is constructed by normal samples), the accuracy is further improved by 5.97%, and the effectiveness and superiority of the method under the actual condition of lack of abnormal samples are proved.
It is to be noted that the disclosed embodiments of the present application are intended to aid in further understanding of the present invention, but those skilled in the art will appreciate that: various substitutions and modifications are possible without departing from the spirit and scope of the invention and appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.
Claims (6)
1. A helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance comprises the following steps:
the first step is as follows: the method comprises the steps of establishing a health baseline by using training data, firstly carrying out FFT (fast Fourier transform) spectrum transformation on the training data, then carrying out RMS (root mean square) compression transformation to establish an input vector, and finally completing establishment of the health baseline by using the input vector of the training data;
the second step is that: generating a baseline statistical threshold by using training data, wherein the training data only comprise normal data or comprise a large amount of normal data and a small amount of abnormal data, and comprises distribution-sample mahalanobis distance measurement and threshold self-adaptive generation, wherein the distribution-sample mahalanobis distance measurement realizes quantitative measurement of individual difference by calculating mahalanobis distance between the overall distribution of the health state characteristic set and health vectors of all training samples; further, performing statistical distribution calculation on the generated Mahalanobis distance sequence, and combining a statistical principle to realize self-adaptive generation of a baseline threshold;
the third step: and carrying out real-time abnormity judgment on test data, wherein the test data comprises normal data and abnormal data, carrying out FFT (fast Fourier transform) spectrum transformation on the test data, then carrying out RMS (root mean square) compression transformation to construct an input vector, acquiring the real-time Mahalanobis distance of the input vector of the test data based on a healthy baseline, and judging that the current state of a helicopter transmission shaft is abnormal if the real-time Mahalanobis distance exceeds a baseline statistical threshold value, otherwise, judging that the current state of the helicopter transmission shaft is normal.
2. The helicopter transmission shaft abnormality determination method based on SAE and Mahalanobis distance according to claim 1, characterized in that:
the training data used for establishing the health baseline and the generating the baseline statistical threshold in the first step are only normal data, or a small amount of abnormal data is introduced on the basis of the normal data, so that a health baseline model is further optimized, and the accuracy of abnormal judgment is improved; and thirdly, the test data adopted by the real-time abnormity judgment comprise normal data and abnormal data.
3. The helicopter transmission shaft abnormality determination method based on SAE and Mahalanobis distance according to claim 1, characterized in that:
the health baseline comprises an SAE model and a health state characteristic set, wherein the SAE model is trained by utilizing training data, the trained SAE model can realize the nonlinear and adaptive transformation of a high-dimensional input vector and generate a health vector, and then the health vectors obtained from normal data in all the training data form the health state characteristic set to form the adaptive quantitative representation of the health state.
4. The helicopter transmission shaft abnormality determination method based on SAE and Mahalanobis distance according to claim 1, characterized in that:
the second step is a baseline statistical threshold generation method, which specifically comprises the following steps:
calculating the overall distribution mean vector and covariance matrix of the health state characteristic set, and recording the vectors as U and C; the dimensionality of the vector U and the dimensionality of the square matrix C are consistent with the dimensionality of the health vector;
sequentially calculating the Mahalanobis distance between the distribution mean vector U and each health vector in the health state characterization set by using the distribution covariance matrix C, obtaining the Mahalanobis distances of M baselines, and forming a distance set, wherein the distance set describes the fluctuation condition of the distance result in the normal state;
generating a helicopter transmission shaft baseline statistical threshold value by utilizing a statistical 6 sigma principle based on the M baseline Mahalanobis distances; the calculation method is as follows:
T＝μ+6σ
wherein mu is the mean value of the base line Mahalanobis distance, sigma is the standard deviation value of the base line Mahalanobis distance, and T is the base line statistical threshold; the baseline statistical threshold value T is a judgment standard for distinguishing the normal/abnormal states of the helicopter.
5. The helicopter transmission shaft abnormality determination method based on SAE and mahalanobis distance according to claim 1, wherein the third step is specifically:
the input vector obtained by aiming at the test data is sent to an SAE model after training is completed, and a real-time state vector transformed by the SAE model is obtained, wherein the real-time state vector is a quantitative representation of the real-time state of a helicopter transmission shaft;
based on the health state feature set and the baseline statistical threshold, performing threshold abnormity judgment on the real-time state vector to realize real-time detection of the state of the transmission shaft of the helicopter; the key links comprise:
the Mahalanobis distance measurement is carried out, the Mahalanobis distance between the health state characteristic set and the real-time state vector is calculated by utilizing the distribution mean vector U and the distribution covariance matrix C of the health state characteristic set, and the quantitative difference measurement between the real-time state and the historical health state is realized;
and (4) judging the abnormal threshold, namely judging whether the Mahalanobis distance of the real-time state vector exceeds the threshold or not based on the baseline statistical threshold, if so, judging that the real-time state is abnormal, otherwise, judging that the real-time state is normal.
6. The helicopter transmission shaft abnormality determination method based on SAE and Mahalanobis distance according to claim 5, characterized in that:
the method comprises the steps of obtaining a real-time Mahalanobis distance of a test data input vector based on a health baseline, wherein the health baseline comprises an SAE model and a health state characteristic set which are finished by training, and firstly, sending the test data input vector into the SAE model to obtain a real-time state vector; and further calculating the Mahalanobis distance between the real-time state vector and the health state syndrome, thereby obtaining the real-time Mahalanobis distance.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110047086.3A CN112668105B (en) | 2021-01-14 | 2021-01-14 | Helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110047086.3A CN112668105B (en) | 2021-01-14 | 2021-01-14 | Helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112668105A CN112668105A (en) | 2021-04-16 |
| CN112668105B true CN112668105B (en) | 2022-05-03 |
Family
ID=75415153
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110047086.3A Active CN112668105B (en) | 2021-01-14 | 2021-01-14 | Helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112668105B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113837591A (en) * | 2021-09-18 | 2021-12-24 | 北京航空航天大学 | Equipment health assessment method oriented to multi-working-condition operation conditions |
| CN114279555A (en) * | 2021-12-27 | 2022-04-05 | 深圳市双合电气股份有限公司 | Motor abnormal vibration monitoring method and system |
| CN114057053B (en) * | 2022-01-18 | 2022-04-26 | 杭州浅水数字技术有限公司 | Method for monitoring fatigue degree of component of special machine |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222765A (en) * | 2019-06-06 | 2019-09-10 | 中车株洲电力机车研究所有限公司 | A kind of permanent magnet synchronous motor health status monitoring method and system |
| CN110567722A (en) * | 2019-05-13 | 2019-12-13 | 上海飞机客户服务有限公司 | Civil engine starting system health monitoring method based on performance parameters |
| CN111652461A (en) * | 2020-04-13 | 2020-09-11 | 北京航空航天大学 | Aero-engine continuous health state evaluation method based on SAE-HMM |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102944435B (en) * | 2012-10-25 | 2015-02-11 | 北京航空航天大学 | Health assessment and fault diagnosis method for rotating machinery based on fisher discriminant analysis and mahalanobis distance |
| CN110309537B (en) * | 2019-05-24 | 2023-06-20 | 北京航天自动控制研究所 | Intelligent health prediction method and system for aircraft |
-
2021
- 2021-01-14 CN CN202110047086.3A patent/CN112668105B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110567722A (en) * | 2019-05-13 | 2019-12-13 | 上海飞机客户服务有限公司 | Civil engine starting system health monitoring method based on performance parameters |
| CN110222765A (en) * | 2019-06-06 | 2019-09-10 | 中车株洲电力机车研究所有限公司 | A kind of permanent magnet synchronous motor health status monitoring method and system |
| CN111652461A (en) * | 2020-04-13 | 2020-09-11 | 北京航空航天大学 | Aero-engine continuous health state evaluation method based on SAE-HMM |
Non-Patent Citations (2)
| Title |
|---|
| A Novel Bearing Health Indicator Construction Method Based on Ensemble Stacked Autoencoder;Pengfei Lin等;《2019 IEEE International Conference on Prognostics and Health Management (ICPHM)》;20190829;第1-9页 * |
| An Effective Fault Diagnosis Method for Aero Engines Based on GSA‑SAE;CUI Jianguo等;《Transactions of Nanjing University of Aeronautics and Astronautics》;20201031;第37卷(第5期);第750-757页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112668105A (en) | 2021-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112668105B (en) | Helicopter transmission shaft abnormity judgment method based on SAE and Mahalanobis distance | |
| CN108008332B (en) | New energy remote testing equipment fault diagnosis method based on data mining | |
| CN109766334B (en) | Processing method and system for online monitoring abnormal data of power equipment | |
| CN108763729B (en) | Process industry electromechanical system coupling state evaluation method based on network structure entropy | |
| CN109489977B (en) | KNN-AdaBoost-based bearing fault diagnosis method | |
| CN103346797B (en) | A kind of real-time compression method for gear distress signal | |
| CN111008363A (en) | Multivariable causal-driven complex electromechanical system service safety situation evaluation method | |
| Shi et al. | Online event detection in synchrophasor data with graph signal processing | |
| CN117421684B (en) | Abnormal data monitoring and analyzing method based on data mining and neural network | |
| CN110321947A (en) | A kind of multiplexing electric abnormality pre-judging method based on convolutional neural networks | |
| CN110737976A (en) | mechanical equipment health assessment method based on multi-dimensional information fusion | |
| CN116683648B (en) | Intelligent power distribution cabinet and control system thereof | |
| CN110766313A (en) | Cable tunnel comprehensive state evaluation method based on operation and maintenance system | |
| CN114416707A (en) | Method and device for automated feature engineering of industrial time series data | |
| CN115600088A (en) | Distribution transformer fault diagnosis method based on vibration signals | |
| CN117421687A (en) | Method for monitoring running state of digital power ring main unit | |
| CN110458071B (en) | DWT-DFPA-GBDT-based optical fiber vibration signal feature extraction and classification method | |
| CN107463872A (en) | A kind of rotating machinery Crack Fault Diagnosis in Shaft method | |
| CN117349657A (en) | Distributed data acquisition module and monitoring system for hydraulic engineering environment monitoring | |
| CN112464554A (en) | Operating parameter optimization method of gasoline refining equipment | |
| CN109522657B (en) | Gas turbine anomaly detection method based on correlation network and SVDD | |
| CN115987692A (en) | Safety protection system and method based on flow backtracking analysis | |
| CN114925476A (en) | Method for predicting quaternion of rolling bearing degradation track augmentation and storage medium | |
| CN112784744B (en) | Mechanical component vibration signal preprocessing method with missing value | |
| Luo et al. | A novel method for remaining useful life prediction of roller bearings involving the discrepancy and similarity of degradation trajectories |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |