CN117609836A - Electromagnetic sensitivity prediction and health management method for integrated module - Google Patents
Electromagnetic sensitivity prediction and health management method for integrated module Download PDFInfo
- Publication number
- CN117609836A CN117609836A CN202311617209.8A CN202311617209A CN117609836A CN 117609836 A CN117609836 A CN 117609836A CN 202311617209 A CN202311617209 A CN 202311617209A CN 117609836 A CN117609836 A CN 117609836A
- Authority
- CN
- China
- Prior art keywords
- data
- fault
- electromagnetic
- monitoring
- time
- 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.)
- Pending
Links
- 230000035945 sensitivity Effects 0.000 title claims abstract description 57
- 230000036541 health Effects 0.000 title claims abstract description 50
- 238000007726 management method Methods 0.000 title claims abstract description 26
- 238000012544 monitoring process Methods 0.000 claims abstract description 100
- 238000000034 method Methods 0.000 claims abstract description 67
- 238000012360 testing method Methods 0.000 claims abstract description 59
- 238000012545 processing Methods 0.000 claims abstract description 45
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 36
- 230000006870 function Effects 0.000 claims abstract description 20
- 230000001681 protective effect Effects 0.000 claims abstract description 9
- 238000000605 extraction Methods 0.000 claims abstract description 6
- 238000003745 diagnosis Methods 0.000 claims description 69
- 238000013527 convolutional neural network Methods 0.000 claims description 39
- 238000009826 distribution Methods 0.000 claims description 22
- 238000004458 analytical method Methods 0.000 claims description 20
- 238000010586 diagram Methods 0.000 claims description 20
- 238000005516 engineering process Methods 0.000 claims description 20
- 230000007246 mechanism Effects 0.000 claims description 20
- 238000002347 injection Methods 0.000 claims description 17
- 239000007924 injection Substances 0.000 claims description 17
- 230000009466 transformation Effects 0.000 claims description 17
- 238000011156 evaluation Methods 0.000 claims description 16
- 238000011160 research Methods 0.000 claims description 14
- 238000012549 training Methods 0.000 claims description 14
- 238000013461 design Methods 0.000 claims description 12
- 230000002159 abnormal effect Effects 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 10
- 238000009472 formulation Methods 0.000 claims description 10
- 239000000203 mixture Substances 0.000 claims description 10
- 238000007405 data analysis Methods 0.000 claims description 9
- 230000000737 periodic effect Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 8
- 238000012795 verification Methods 0.000 claims description 8
- 238000012952 Resampling Methods 0.000 claims description 7
- 230000002829 reductive effect Effects 0.000 claims description 7
- 238000010187 selection method Methods 0.000 claims description 7
- 238000013145 classification model Methods 0.000 claims description 6
- 238000009499 grossing Methods 0.000 claims description 6
- 238000012417 linear regression Methods 0.000 claims description 6
- 238000003062 neural network model Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 230000003449 preventive effect Effects 0.000 claims description 5
- 238000003672 processing method Methods 0.000 claims description 5
- 230000008859 change Effects 0.000 claims description 4
- 230000011218 segmentation Effects 0.000 claims description 4
- 238000004088 simulation Methods 0.000 claims description 4
- 230000000007 visual effect Effects 0.000 claims description 4
- 230000008901 benefit Effects 0.000 claims description 3
- 230000002457 bidirectional effect Effects 0.000 claims description 3
- 239000003086 colorant Substances 0.000 claims description 3
- 238000003066 decision tree Methods 0.000 claims description 3
- 239000006185 dispersion Substances 0.000 claims description 3
- 230000005670 electromagnetic radiation Effects 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 238000004519 manufacturing process Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000005259 measurement Methods 0.000 claims description 3
- 238000002156 mixing Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000007637 random forest analysis Methods 0.000 claims description 3
- 230000004044 response Effects 0.000 claims description 3
- 230000002441 reversible effect Effects 0.000 claims description 3
- 238000007619 statistical method Methods 0.000 claims description 3
- 230000002194 synthesizing effect Effects 0.000 claims description 3
- 238000012163 sequencing technique Methods 0.000 claims description 2
- 230000005855 radiation Effects 0.000 abstract description 6
- 238000005457 optimization Methods 0.000 abstract description 2
- 230000000875 corresponding effect Effects 0.000 description 17
- 239000000306 component Substances 0.000 description 8
- 238000013135 deep learning Methods 0.000 description 5
- 238000004891 communication Methods 0.000 description 4
- 238000002790 cross-validation Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000012423 maintenance Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 238000009877 rendering Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 239000000295 fuel oil Substances 0.000 description 1
- 230000002427 irreversible effect Effects 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007257 malfunction Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000008054 signal transmission Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/096—Transfer learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to an electromagnetic sensitivity prediction and health management method of a comprehensive module, which comprises the following steps: and (3) data acquisition: constructing a matched test platform based on a conduction time domain sensitivity test; and (3) data processing: extracting useful information; and (3) state monitoring: the state monitoring layer receives the parameter data which is provided by the data processing layer and passes through each data processing algorithm; health assessment: constructing a fault prediction model according to the output of the state monitoring layer by adopting a fault prediction method, and comprehensively judging the health state of the comprehensive module; dynamic adjustment: the dynamic adjustment layer is unfolded around the incremental learning algorithm; decision support: is responsible for providing feasible protective measures and decision bases for related electromagnetic sensitivity phenomena and fault degrees thereof. The problems caused by the influence of electromagnetic interference and radiation on various devices are solved, the device has stronger data processing and feature extraction capacity, the health management and early warning of emc fault problems are realized, and the electromagnetic random optimization and active protection functions of the comprehensive processing module are realized.
Description
Technical Field
The invention relates to an electromagnetic compatibility and protection technology, in particular to an electromagnetic sensitivity prediction and health management method of a comprehensive module.
Background
The complex electromagnetic environment is formed by overlapping a plurality of electromagnetic signals which are densely distributed, multiple in quantity, complex in pattern and dynamic random in a time domain, a frequency domain, an energy domain and a space domain, and has influence on equipment, fuel oil, personnel and the like, and the normal operation of an information system and electronic equipment is seriously hindered.
It is important to protect electronic devices from electromagnetic interference and radiation. The significance of researching electromagnetic compatibility and protection is not only embodied in protecting various electronic equipment from interference and radiation, but also in ensuring the reliability and effectiveness of future technologies. Various types of equipment are subject to electromagnetic interference and radiation, which may lead to unpredictable results. Therefore, research on electromagnetic compatibility and protection technology not only has scientific and engineering significance, but also is a strategic requirement.
Disclosure of Invention
Aiming at the problem that various devices are affected by electromagnetic interference and radiation and can possibly lead to unpredictable results, the electromagnetic sensitivity prediction and health management method of the comprehensive module is provided. The corresponding relation between the electromagnetic compatibility fault problem and the module board card operation parameter is found, potential association is excavated by utilizing a front-edge artificial intelligence technology, an intelligent algorithm capable of realizing state monitoring, fault prediction and health assessment is provided, health management and early warning of emc fault problem are realized, and electromagnetic random optimization and active protection functions of the comprehensive module are processed.
The technical scheme of the invention is as follows:
the electromagnetic sensitivity prediction and health management method of the comprehensive module comprises a data acquisition layer, a data processing layer, a state monitoring layer, a health evaluation layer, a dynamic adjustment layer and a decision support layer, and comprises the following steps of:
step 1: and (3) data acquisition: based on the conduction time domain sensitivity test, a matched test platform is constructed, comprising: interference injection test, electromagnetic parameter selection and real-time monitoring tool design;
step 2: and (3) data processing: through comprehensive data processing steps to extract useful information and ensure data quality, analyzing and processing input signals from a data acquisition layer to generate meaningful state descriptions;
step 3: and (3) state monitoring: the state monitoring layer receives the parameter data provided by the data processing layer and passing through each data processing algorithm, and inputs electromagnetic characteristic parameters which are input in accordance with the neural network to a data-driven fault diagnosis algorithm to determine electromagnetic sensitivity and fault degree so as to monitor the state of the comprehensive module; dividing the fault type and degree of the obtained key electromagnetic parameters through a classifier; when a fault occurs, the type and the fault degree of an electromagnetic sensitivity phenomenon are rapidly diagnosed, and then the electromagnetic sensitivity boundary is accurately estimated; based on fault diagnosis, the running state of the comprehensive module can be acquired and evaluated in real time, and real-time information support is provided for a later health evaluation layer;
Step 4: health assessment: adopting a fault prediction method, combining the real-time performance state and the historical data information according to the output of the state monitoring layer, constructing a fault prediction model, and comprehensively judging the health state of the comprehensive module; the fault prediction technology is key feature information obtained through a data processing and feature extraction technology, and is used for predicting the running state of the board card by combining the real-time running state and historical monitoring data, and can be combined with a fault diagnosis technology to provide more fault prediction information;
step 5: dynamic adjustment: the dynamic adjustment layer is unfolded around the incremental learning algorithm, so that the model in state monitoring and health evaluation can have the capability of autonomous learning, new knowledge is learned from real-time data which are continuously acquired, and the fault diagnosis model and the fault prediction model are updated according to whether a fault diagnosis result output in fault diagnosis and a fault prediction result output in fault prediction reach expectations or not, thereby improving the accuracy of diagnosis and prediction, and simultaneously enabling the model to have the capability of autonomous learning;
step 6: decision support: the decision support layer is responsible for providing feasible protective measures for related electromagnetic sensitivity phenomena and fault degrees thereof, optimizing the safety performance of the system and guiding the task formulation and the protection scheme; and various data of other layers are collected, processed and analyzed to provide decision basis for system operation.
Further, the interference injection test, electromagnetic parameter selection and real-time monitoring tool in the step 1 specifically comprises the following steps: setting up a whole set of test platform for the interference injection test, designing a whole set of test flow, and exciting the sensitive phenomenon of the tested equipment; the electromagnetic compatibility characteristic parameter selection method is to study which working parameters of the tested modules have stronger correlation with the sensitive phenomenon according to the touch of the sensitive characteristic of the board card in the early test, and provide design basis for the follow-up real-time monitoring of the board card; the real-time monitoring tool integrates the sensor, the monitoring circuit and the data interface, and improves the effectiveness of monitoring parameters by designing and optimizing monitoring points, so that support is provided for follow-up intelligent processing.
Further, the step 1 specifically includes:
step 1.1: the injection test needs to determine the monitoring point position through hardware measurement or a design drawing of a hardware drawing generation module tool;
step 1.2: manufacturing a simulation motherboard to cooperate with a test tool to develop a sensitivity threshold test of the module, and establishing a corresponding relation among the type of an interference signal, the intensity of the interference signal, the injection position, the tested module, the working state, the sensitivity phenomenon and the electromagnetic signal intensity of a monitoring point;
step 1.3: applying the test fixture to the assembly, installing a tested module to perform equipment-level test work, establishing a mathematical relationship from interference to monitoring points, and positioning accumulated data for the problem;
Step 1.4: aiming at the packaged sensitive test data, electromagnetic characteristic data acquisition and collection are carried out;
the electromagnetic characteristics of the modules under different states are respectively inspected by enabling the tested object to be in a certain working state;
step 1.5: the stable operation of the case is kept, and faults are eliminated;
the data sample is mainly composed of a power port, a network port and a video port, wherein after the 3 ports are respectively injected with interference specified by a test, the recorded frequency domain information is recorded; the characteristic parameters are frequency, voltage and monitoring signal intensity respectively, and a sample is generated according to the mode; obtaining test data from different dimensions requires selection and setting according to the functional characteristics of the board.
Further, the specific steps in step 1.5 are:
step 1.5.1: the chassis is electrified and preheated to enable the chassis to reach a stable working state;
step 1.5.2: monitoring the state of the board card, and ensuring the normal work of oscilloscope data acquisition;
step 1.5.3: acquiring electromagnetic characteristic data at the moment as a reference and simultaneously observing the working performance of the board card in real time;
step 1.5.4: connecting the signal generator and the power amplifier together to a power port on the tested module through a current injection clamp;
step 1.5.5: various electromagnetic interference signals specified by the test are injected into the power port according to a certain amount until the board card has abnormal conditions in the sensitive criterion;
Step 1.5.6: comparing and calibrating the electromagnetic characteristic data obtained at the moment with electromagnetic characteristic data in normal operation, and storing voltage data of each test point by using an oscilloscope;
step 1.5.7: reducing the interference field intensity until the work of the board card is recovered to be normal;
step 1.5.8: steps 1.5.4-1.5.6 are repeated until no new failure of the chassis occurs.
Further, the step 2 specifically includes:
step 2.1: and (3) data preprocessing design: firstly, filling missing values, smoothing noise data, smoothing or deleting outliers to solve the problem of data inconsistency, and secondly, aiming at time series data related to electromagnetic compatibility obtained from a board card, obtaining periodic characteristics by extracting average values, variances, maximum and minimum values of each parameter and applying trigonometric function fitting; finally, normalizing the features to ensure that the features have the same scale;
step 2.2: the data analysis method comprises the following steps: the key steps in the data analysis and feature selection process are used for improving the selection of monitoring parameters, the monitoring efficiency and the accuracy; key features related to the sensitive phenomenon are screened out through a visual feature map, a feature selection function and a tree-based related feature selection method, the effectiveness of each parameter information corresponding to the sensitive phenomenon is analyzed, the selection of monitoring points and monitoring parameters is continuously optimized, and support is provided for monitoring and understanding the sensitive phenomenon;
Step 2.3: data balancing problem solving strategy: in order to improve the classification accuracy of few samples, an unbalanced data processing strategy is used, and an undersampling method based on Euclidean distance and an oversampling method based on K-Means SMOTE of cluster distribution are combined to reconstruct sample data;
step 2.4: other signal processing methods: in order to fully utilize the time-frequency domain information, the limitation of the time domain or the frequency domain can be effectively overcome by a time-frequency combination mode, and a two-dimensional time-frequency diagram is generated by using continuous wavelet transformation.
Further, step 2.1 specifically includes:
step 2.1.1: selecting a fixed time window length for dividing the data segments in the time series data;
step 2.1.2: the following steps are performed for each parameter: calculating the average value, variance, maximum value and minimum value of the parameters in a given time window; these statistics provide the basic information of the data distribution, reflecting the trend of the variation and the amplitude fluctuation of the parameters within the time window; the average value is the central trend of the data distribution, representing the overall level of the data over a time window; the variance measures the degree of dispersion of the data and reflects the fluctuation of the data; the maximum and minimum values represent the highest and lowest points of the data within the time window, respectively;
Step 2.1.3: performing discrete Fourier transform on the time series data of the parameters by using discrete Fourier transform, and converting the signals from a time domain to a frequency domain; the result of Fourier transformation is analyzed, main frequency components can be identified, the periodic characteristics of representative parameters are subjected to trigonometric function fitting on the selected frequency components, frequency information in fitting parameters is obtained, and the periodic characteristics are calculated;
step 2.1.4: normalizing the eigenvalues obtained from each parameter to ensure that they are consistent in scale;
step 2.2 is specifically:
step 2.2.1: making a clustering scatter diagram, clustering according to characteristic values, displaying data points of different clusters on the diagram by using different colors or marks, and observing the distribution of the data points on the diagram to see whether natural data clustering phenomenon exists or not, so as to judge whether a certain association relationship exists between the characteristics or not;
step 2.2.2: making a thermodynamic diagram, displaying a correlation coefficient matrix between the features, and filtering high-correlation features, wherein the closer the value of the correlation coefficient is 1 or-1, the stronger the correlation between the two features is represented; deleting one of the features with the correlation of 1 or-1 in the thermodynamic diagram;
Step 2.2.3: based on the related feature selection of the feature function, calculating the relation degree between each feature and the result by carrying out statistical analysis, sequencing from high score to low score, and selecting the features with the highest scores for modeling; the feature number is selected by means of cross verification, verification results are carried out on the initially selected feature parameters, higher accuracy can be obtained when the feature parameters with higher use scores are displayed, and at the moment, the remaining features with the lowest scores can be discarded;
step 2.2.4, based on relevant feature selection of the tree model, evaluating feature importance by utilizing a decision tree or random forest algorithm, drawing an importance ranking chart, and selecting the first few main features for training verification according to importance ranking;
the step 2.3 is specifically as follows:
step 2.3.1: classifying the state data sample set D according to the sample attribute of each module board card; a minority class sample set is defined as S, and a majority class sample set is defined as M;
step 2.3.2: resampling is carried out under the classification according to the unbalanced data set, and the resampling proportion is set;
step 2.3.3: oversampling a minority class of samples by using a K-Means SMOTE method for generating samples based on cluster distribution; firstly, based on a K-means algorithm, K groups are clustered by using a K-means, clusters used for oversampling are filtered and selected, and clusters with a high proportion of minority samples are reserved; then, the number of the synthesized samples is distributed, and most samples are distributed to clusters with sparse distribution of few samples; only the data of the key parameters are oversampled, other index samples are kept as they are, and the indexes with high information value and high meaning related to electromagnetic sensitivity phenomena are emphasized; finally, a SMOTE algorithm is applied to each selected cluster to synthesize a new sample, and the new sample is added into a minority class set S;
Step 2.3.4: undersampling a majority of samples in a non-boundary domain by adopting an undersampling method based on Euclidean distance; the first step is to determine the center point of the majority sample set M; secondly, calculating the values of the samples in the plurality of types of sample sets from the center point, and arranging the calculated values in sequence; thirdly, deleting most samples according to the rule from far to near from the center point according to the undersampling multiplying power until the most samples are balanced with the data of few samples; according to the method, a boundary sample is found, most types of samples far away from a classification boundary and small in influence on classification are deleted, and safety samples with large classification influence and close to the classification boundary are stored in a training set;
step 2.3.5: synthesizing the updated minority class samples and majority class samples into a new balanced sample set D new The method comprises the steps of carrying out a first treatment on the surface of the Novel equalized sample set D new The equalization of sample types is realized, and the data distribution characteristics of the original data set are maintained;
step 2.4 is specifically: other signal processing methods:
step 2.4.1: let a be the scale factor, f s For sampling frequency f c For the wavelet center frequency, the actual frequency f corresponding to the scale factor a a The method comprises the following steps:
f a ＝f c ×f s /a (1)
step 2.4.2: in order for the converted frequency sequence to be an equal sequence, the form of the scale sequence must be as follows:
c/totalscal,…,c/(totalscal-1),c/4,c/2,c (2)
In the above description, total is the length of the selected scale sequence when the signal is subjected to continuous wavelet transformation, and is generally selected to be 256, and c is a constant;
step 2.4.3: from equation (1), the actual frequency corresponding to the scale c/total should be f s 2, then it is possible to obtain:
2×F c ×totalscal (3)
a scale sequence meeting the above criteria can be obtained;
step 2.4.4: after determining wavelet basis functions and scales, the wavelet coefficients W are obtained according to the wavelet transformation principle f (a, b), then the corresponding relation between the scale factors and the frequencies is used for solving an actual frequency sequence f, and then a two-dimensional time-frequency transformation chart can be drawn according to a time sequence t.
7. The electromagnetic sensitivity prediction and health management method of the integrated module according to claim 1, wherein the step 3 is specifically:
setting different fault diagnosis methods for a one-dimensional signal and a two-dimensional time-frequency data subjected to signal processing, fully utilizing data information, wherein the one-dimensional signal adopts the following step 3.2, and the two-dimensional signal adopts the following step 3.1;
step 3.1: a fault diagnosis method based on a multichannel pyramid attention mechanism comprises the following steps:
the multi-dimensional time-frequency domain analysis is carried out on various signals, and the selected signals are accurately classified by means of the advantage that characteristics are extracted together by the neural network models with various different structures and combining with the characteristic fusion of a multi-channel pyramid attention mechanism;
The specific flow is as follows:
step 3.1.1: collecting data and analyzing signals; respectively carrying out time domain analysis, frequency domain analysis and time-frequency domain analysis on the original signals, and reserving data modes of related signals;
step 3.1.2: building a neural network model; respectively constructing residual neural network and cavity convolutional neural network models of all channels, and determining parameters of all modules;
step 3.1.3: constructing a pyramid attention mechanism; after the SE component is inserted, the weight of the feature map channel can be adjusted, so that the element information expression is more complete; to a certain extent, the channel characteristics of inactivity and unobvious are further inhibited; an SPC segmentation module is applied on the basis of an SE attention mechanism, and a complete pyramid attention mechanism model is built;
step 3.1.4: data splicing and network connection; the data obtained in the steps are subjected to dimension transformation, orderly spliced, input into a pyramid attention mechanism and finally obtained through a full-connection layer and a classification layer;
step 3.2: fault diagnosis technology based on one-dimensional convolutional neural network:
the fault diagnosis method based on the one-dimensional convolutional neural network utilizes a CNN structure with end-to-end characteristics, so that the problem of combination of a feature extractor and a classifier is avoided, and the method can directly act on a time domain signal to avoid losing important time domain features; meanwhile, a CNN model with a hierarchical structure is constructed to further improve the diagnosis precision; the first level of the model is to divide fault types, and the sampling data are divided into different fault types through a CNN classification model; in the second hierarchy, different CNN classification models are respectively constructed for each fault type, and a certain fault type is further divided into a plurality of classes with different fault degrees so as to more accurately classify the faults;
The fault diagnosis steps are as follows:
step 3.2.1: fault type classification: in this step, emphasis is placed on constructing a CNN model to determine fault type labels for each sample; taking the whole data set as the input of a hierarchical CNN model; constructing a new CNN, training a network to obtain network parameters, and dividing fault types by using the trained network;
step 3.2.2: fault degree division: for a certain fault type divided in the previous step, the fault degree can be further divided; the procedure described in step 3.2.2 is similar to the previous step; constructing a second layer of the model according to the fault type dividing result, respectively constructing a CNN for each fault type, training a network to obtain network parameters, and determining the fault degree;
step 3.2.3: calculating the network classification precision: the classification precision of the model is measured through the error classification times; the final accuracy is calculated by the ratio of the total number of classifications to the total number of samples;
the state monitoring layer is divided into two types, one type is used for carrying out multi-channel neural network fault diagnosis algorithm research on signal data after time-frequency mixing, the other type is used for directly carrying out one-dimensional convolutional neural network fault diagnosis algorithm research on the basis of one-dimensional signal data, and meanwhile, fault degrees are subdivided according to different electromagnetic sensitivity phenomena, electromagnetic sensitivity boundaries are better evaluated, so that corresponding electromagnetic protection measures are set according to different fault degrees, and the influence caused by key module loss is reduced.
Further, the step 4 specifically includes:
firstly, extracting key features from sequentially input data by using a feature extractor; then, establishing a plurality of layers of bidirectional LSTM at the top of the feature extractor to realize the encoding of time information; then constructing a stacked full-connection layer and a linear regression layer on the basis of a time encoder for predicting target data; the predicted data is then input into a fault diagnostor, and analyzed to provide further reference information;
SAE can extract data features, while SAE-based depth bi-directional LSTM can encode the input sequence information and learn its representation; comprises four major parts: the device comprises a feature extractor, a time encoder, a full connection layer, a linear regression layer and a fault diagnosis device; applying a layer of SAE on the original input sequence to extract the data features, after extracting the features, constructing a multi-layer depth bi-directional LSTM on top of the previous SAE for encoding time information; then stacking two fully connected layers together for processing the output of the deep bi-directional LSTM; finally, a fault diagnosis device is added at the top of the system to diagnose the predicted information so as to provide more predicted information.
Further, the step 5 specifically includes:
based on the dynamic adjustment of incremental learning, comparing the predicted result of the predicted model with actual data in the running process of the system, comparing the manually input fault type with the diagnosis result of the diagnosis model, and defaulting to a normal state if no input exists; if the model parameters are different, replacing the predicted data with the actual data to carry out reverse error propagation, and updating the model parameters; by the dynamic adjustment method based on incremental learning, the established diagnosis model and the prediction model do not need to be re-established when facing the newly added data, but only update the change caused by the newly added sample on the basis of the original model.
Further, the step 6 specifically includes:
providing safeguard information: the decision support layer provides feasible protective measure information through analysis and evaluation of related electromagnetic sensitivity phenomena; corresponding response schemes are recommended aiming at different fault types and degrees, so that the influence of electromagnetic radiation on a system and a comprehensive module is reduced, and the safety of equipment and personnel is ensured;
optimizing the safety performance of the system: the decision support layer optimizes the safety performance of the system by monitoring the state of the system and the fault output condition in real time; analyzing abnormal parameters and fault output conditions, rapidly identifying potential problems, taking preventive measures, avoiding faults and improving the stability and reliability of the system;
And (3) guiding task formulation and protection scheme: the data analysis and decision result of the decision support layer provide important guidance for the task planning and the formulation of the protection scheme; through the evaluation of the fault type and degree, the system can help users reasonably arrange tasks and formulate corresponding protection strategies so as to ensure the smooth progress of the tasks and the safe operation of equipment.
The invention has the beneficial effects that:
stable operation of the electronic device is critical to successful task execution. The research result can be applied to electromagnetic compatibility fault diagnosis and analysis of equipment. By analyzing the electromagnetic characteristic data set, the real-time monitoring and diagnosis of the electromagnetic compatibility problem of the electronic equipment can be realized, so that the reliability and efficiency of the system are improved. Command and control systems are often faced with complex electromagnetic environments that are susceptible to electromagnetic interference. The electromagnetic sensitivity prediction and health management method of the comprehensive module based on deep learning can be used for monitoring the working state of the electronic equipment in real time, predicting potential faults and taking automatic measures to protect the system from electromagnetic interference.
Emphasis is placed on the application of deep learning algorithms in state monitoring and health assessment. Compared with the traditional method, the method has stronger data processing and feature extraction capability, and can be better suitable for complex electromagnetic compatibility environments.
Drawings
FIG. 1 is a diagram of an electromagnetic compatibility PHM architecture of the present invention;
FIG. 2 is a clustered scatter plot of the present invention;
FIG. 3 is a thermodynamic diagram of the present invention;
FIG. 4 is a graph of analysis of the cross-validation results of the present invention;
FIG. 5 is a tree-based feature importance ranking graph of the present invention;
FIG. 6 is a flow chart of non-equalized data resampling in accordance with the present invention;
FIG. 7 is a flowchart of an algorithm diagnostic of the present invention;
FIG. 8 is a logic diagram of a fault diagnosis model implementation based on a one-dimensional convolutional neural network in accordance with the present invention;
FIG. 9 is a diagram of a fault prediction model of the present invention based on key features and DBLSTM;
FIG. 10 is a graph of a predictive model based on incremental learning in accordance with the present invention.
Detailed Description
The invention will now be described in detail with reference to the drawings and specific examples. The present embodiment is implemented on the premise of the technical scheme of the present invention, and a detailed implementation manner and a specific operation process are given, but the protection scope of the present invention is not limited to the following examples.
Technical terms:
and (3) a comprehensive module: an integrated module generally refers to a portion of a complex electronic device or system that may include multiple components, sensors, circuitry, etc. for performing a particular function. The integrated module may be a subsystem or modular component in the electronic product.
Electromagnetic sensitivity: electromagnetic sensitivity generally refers to the degree to which an integrated module or electronic device is sensitive to electromagnetic interference or radiation. This may lead to performance degradation or malfunction of the electronic device, and thus predictive and regulatory measures need to be taken to reduce the effects of electromagnetic interference.
Health management: health management refers to the process of monitoring, diagnosing and maintaining integrated modules or electronic devices to ensure their proper operation. This includes activities such as fault diagnosis, performance monitoring, preventive maintenance, etc.
PHM (Prognostics Health Management fault prediction and health management) technology can accurately sense and predict the system state development trend, evaluate the reliability of the system, predict faults in advance and prevent irreversible faults. The PHM technology is not applied to the field of electromagnetic compatibility at present, an electromagnetic compatibility protection technology based on PHM is established, three types of typical hardware modules are taken as research objects, state monitoring, anomaly prediction and health management are realized, and the electromagnetic compatibility problem is solved through PHM technology theory and method.
Six main functional modules are defined based on PHM technology, and are respectively data acquisition, data processing, state detection, health evaluation, dynamic adjustment and decision support. The architecture is shown in fig. 1, and the functional modules are formed according to the hierarchical relationship of information flows.
The study framework of the present invention differs from the traditional framework mainly in the following aspects, emphasizes dynamic adjustment and decision support, and achieves the mutual combination of state monitoring and health assessment.
While conventional frameworks often lack dynamics, new research frameworks introduce dynamic tuning modules that introduce the idea of incremental learning, enabling diagnostic and predictive models to be continually learned and tuned from new data, rather than relying solely on fixed predefined rules. This improves the adaptability and long-term stability of the algorithm, thereby rendering the diagnostic model and the predictive model adaptive. A decision support layer is introduced, and the main task of the layer is to provide information of feasible protective measures for related electromagnetic sensitive phenomena, optimize the safety performance of a system, guide the formulation of tasks, a protective scheme and the like. This makes the research framework more practical and comprehensive. The state monitoring and health assessment are combined with each other, and are no longer isolated functional modules. The characteristic parameters provided by the state monitoring layer are not only used for fault diagnosis, but also used for health assessment, so that deeper information sharing and utilization are realized, and the comprehensive performance of the whole algorithm is improved. Emphasis is placed on the application of deep learning algorithms in state monitoring and health assessment. Compared with the traditional method, the method has stronger data processing and feature extraction capability, and can be better suitable for complex electromagnetic compatibility environments.
In general, the invention emphasizes comprehensive treatment and decision support for the electromagnetic compatibility problem in the dynamic environment, realizes higher-level state monitoring and health assessment through modern technologies such as deep learning, incremental learning and the like, and brings new working and research directions for the field of electromagnetic compatibility PHM.
An electromagnetic sensitivity prediction and health management method for a comprehensive module comprises the following steps:
step 1: data acquisition layer
Based on the conduction time domain sensitivity test, a matched test platform is constructed, and the test platform mainly comprises three parts of interference injection test, electromagnetic parameter selection and real-time monitoring tool design. The method comprises the steps of constructing a whole set of test platform for the interference injection test, designing a whole set of test flow, and exciting the sensitive phenomenon of the tested equipment; the electromagnetic compatibility characteristic parameter selection method is to study which working parameters of the tested modules have stronger correlation with the sensitive phenomenon according to the touch of the sensitive characteristic of the board card in the early test, and provide design basis for the follow-up real-time monitoring of the board card; the real-time monitoring tool integrates the sensor, the monitoring circuit and the data interface, and improves the effectiveness of monitoring parameters by designing and optimizing monitoring points, so that support is provided for follow-up intelligent processing.
The specific data set construction steps are as follows:
step 1.1: the injection test needs to determine the monitoring point position through hardware measurement or a design drawing of a hardware drawing generation module tool;
step 1.2: manufacturing an analog motherboard (only providing a power supply) and carrying out sensitivity threshold test of a module by matching with a test tool, and establishing corresponding relations among the type of an interference signal, the intensity of the interference signal, the injection position, the tested module, the working state, the sensitivity phenomenon and the electromagnetic signal intensity of a monitoring point;
step 1.3: applying the test fixture to the assembly, installing a tested module to develop equipment-level test work, establishing mathematical relations such as transfer functions from interference to monitoring points, and accumulating data for problem positioning;
step 1.4: and (5) aiming at the packaged sensitive test data, carrying out electromagnetic characteristic data acquisition and collection.
The electromagnetic characteristics of the modules under different states are respectively inspected by enabling the tested object to be in a certain working state. The main working states are as follows: (1) system idle: the method is characterized in that no user application or task is run except the operation system and basic system service; (2) slight load: running a log system, serial port communication, network sending monitoring and other lightweight application programs; (3) workload: running typical run-time tasks such as signal processing, display rendering, data distribution, etc., (4) baking machine load: a specific program is executed so that the core component (CPU, GPU, DSP, RAM, etc.) runs fully.
When the sensitivity phenomenon test is carried out, the judging method for judging the working performance reduction of the tested product comprises the following steps: whether the high-performance display control module works normally or not comprises whether abnormal phenomena such as water ripple, black screen, ethernet packet loss, interruption and the like occur or not; whether the high-performance data processing module works normally or not comprises the phenomena of Ethernet packet loss, interruption, system crash and the like; whether the high-density signal processing module works normally or not includes whether the phenomena of Ethernet packet loss, interruption, serial port interruption and the like occur or not.
The following experimental step 1.5 was then performed:
step 1.5.1: the chassis is electrified and preheated to enable the chassis to reach a stable working state.
Step 1.5.2: and monitoring the state of the board card to ensure the normal work of the data acquisition of the oscilloscope.
Step 1.5.3: and acquiring electromagnetic characteristic data at the moment as a reference and simultaneously observing the working performance of the board card in real time.
Step 1.5.4: the signal generator and the power amplifier are connected to a power port on the tested module through a current injection clamp.
Step 1.5.5: various electromagnetic interference signals specified by the test are injected into the power port according to a certain amount until the board card has abnormal conditions in the sensitive criteria.
Step 1.5.6: and comparing and calibrating the electromagnetic characteristic data obtained at the moment with electromagnetic characteristic data in normal operation, and storing voltage data of each test point by using an oscilloscope.
Step 1.5.7: the interference field intensity is reduced until the operation of the board card is recovered to be normal.
Step 1.5.8: repeat steps 1.5.4-1.5.6: until the chassis no longer fails.
The data sample is mainly the frequency domain information recorded after the power port, the network port and the video port and the 3 ports are respectively injected with the interference specified by the test. The characteristic parameters are frequency, voltage and monitoring signal strength, respectively, in such a way that a sample is generated. Obtaining test data from different dimensions requires selection and setting according to the functional characteristics of the board. In the project, according to the requirements of fault diagnosis and analysis special diagnosis tests, data of two dimensions of time domain information and frequency domain information are selected, and data processing and analysis are carried out according to test results so as to obtain comprehensive evaluation. By learning and extracting the correlation between sample data and the sensitive phenomenon, a mapping relation is established, a simulation prediction model is constructed based on signal transmission coefficients, and the prediction and simulation work of the electromagnetic sensitive threshold value is carried out.
Step 2: data processing layer
When the electromagnetic compatibility related parameter data is obtained from the board card to be tested, the data obtained directly from the monitoring equipment generally has various problems, such as disorder, deficiency, overhigh dimensionality, unbalanced class, insufficient characteristic information and the like, and the comprehensive data processing steps are required to extract useful information and ensure the data quality. The input signals from the data acquisition layer are analyzed and processed to generate meaningful state descriptions.
Step 2.1: data preprocessing design
Firstly, filling missing values, smoothing noise data, smoothing or deleting outliers to solve the problem of data inconsistency, and secondly, aiming at time series data related to EMC (electromagnetic compatibility) obtained from a board card, obtaining periodic characteristics by extracting average values, variances, maximum and minimum values of each parameter and applying trigonometric function fitting. Finally, the features are normalized to ensure that the features have the same scale.
The main processing steps are as follows:
step 2.1.1: a fixed time window length is selected for dividing the data segments in the time series data.
Step 2.1.2: the following steps are performed for each parameter: the mean, variance, maximum and minimum of the parameters over a given time window are calculated. These statistics provide the basic information of the data distribution, reflecting the trend of the parameter variation and the amplitude fluctuations within the time window. The average is the central trend of the data distribution, representing the overall level of data over a time window. The variance measures the degree of dispersion of the data and reflects the volatility of the data. The maximum and minimum values represent the highest and lowest points of the data within the time window, respectively.
Step 2.1.3: discrete fourier transform (FFT) is used to perform a discrete fourier transform on the time-series data of the parameters, converting the signal from the time domain to the frequency domain. By analyzing the result of Fourier transformation, the main frequency components can be identified, the periodic characteristics of the representative parameters are subjected to trigonometric function fitting on the selected frequency components, the frequency information in the fitting parameters is obtained, and the periodic characteristics are calculated.
Step 2.1.4: the eigenvalues obtained from each parameter are normalized to ensure that they are consistent in scale.
Step 2.2: data analysis method
The key steps in the data analysis and feature selection process are used for improving the selection of monitoring parameters, the monitoring efficiency and the accuracy. The key features related to the sensitive phenomenon are screened out through a visual feature map, a feature selection function and a tree-based related feature selection method, the effectiveness of each parameter information corresponding to the sensitive phenomenon is analyzed, the selection of monitoring points and monitoring parameters is continuously optimized, and support is provided for monitoring and understanding the sensitive phenomenon.
Visual feature map:
step 2.2.1: clustering the clustered scatter diagram according to the characteristic values, and displaying data points of different clusters on the diagram by using different colors or marks, wherein the abscissa is the characteristic value extracted from the data obtained by monitoring the point, as shown in fig. 2; the ordinate is the result of classifying the state of the integrated module using the feature value. ZC is normal state, GZ is network port fault, SJ is dead state, DGZ is network port and serial port fault, CKGZ is serial port fault. Fig. 2 can observe whether individual fault states can be clearly and completely divided according to individual characteristic values. By observing the distribution of the data points on the graph, we can see whether a natural data clustering phenomenon exists, so as to judge whether a certain association relationship exists between the features.
Step 2.2.2: as shown in fig. 3, the abscissa and the ordinate are characteristic values extracted from data obtained by monitoring the point positions; the numbers in the figure represent the correlation of two eigenvalues, 1 being the two features highly correlated, -1 being highly uncorrelated. And displaying a correlation coefficient matrix between the features, filtering the high-correlation features, wherein the closer the value of the correlation coefficient is 1 or-1, the stronger the correlation between the two features is indicated. One of the features of the thermodynamic diagram with a correlation of 1 or-1 is deleted.
The characteristic selection method comprises the following steps:
step 2.2.3: based on the related feature selection of the feature function, the relation degree between each feature and the result is calculated through statistical analysis, and the top k features with highest scores are selected for modeling according to the sequence from high score to low score. And selecting the feature number by using a cross-validation mode, as shown in fig. 4, wherein the abscissa is the feature number selected during training validation; the ordinate is the prediction accuracy obtained by cross-validation, and it can be seen that the accuracy obtained is highest when 16 features are selected in this embodiment. The verification result performed on the first selected 20 feature parameters shows that when 16 feature parameters are used, higher accuracy can be obtained, and at the moment, the 4 features with the lowest scores can be discarded.
Step 2.2.4: based on the related feature selection of the tree model, evaluating feature importance by utilizing algorithms such as decision trees or random forests, drawing an importance ranking chart, and as shown in fig. 5, the abscissa is a feature value extracted from data obtained by monitoring the point positions; the ordinate is the importance score of a feature, the higher the score, the higher the importance representing the classification of the feature to the model. And selecting the first eight main features for training verification according to the importance degree sequence.
Step 2.3: data balancing problem solving strategy
The fault samples which usually generate electromagnetic sensitivity phenomena are far less than normal state samples, the value of accurately identifying the fault samples in a few classes is higher than that of the normal state samples in a plurality of classes, meanwhile, the time from generating faults to causing harm of a key module is not obviously regular, and the minimum time is only a few minutes. In order to improve the classification accuracy of minority class samples, an unbalanced data processing strategy is used, and an undersampling method based on Euclidean distance and an oversampling method based on clustering distribution K-Means (K-Means clustering K average algorithm) SMOTE (Synthetic Minority Oversampling Technique synthesized minority class oversampling technology) are combined with reconstructed sample data, and the flow is shown in fig. 6, and the steps are as follows:
Step 2.3.1: and classifying the state data sample set D according to the sample attribute monitoring state data of each module board card. The minority class sample set is defined as S and the majority class sample set is defined as M.
Step 2.3.2: resampling ratio is set according to the sub-class resampling of the unbalanced data set.
Step 2.3.3: a few classes of samples are oversampled using the K-Means SMOTE method that generates samples based on a cluster distribution. First, K groups are clustered using K-means based on the K-means algorithm, clusters are filtered to select for oversampling, clusters with a high proportion of minority class samples are retained. The number of composite samples is then allocated, with most of the samples being allocated to clusters with sparse distribution of few samples. Considering the practical significance of each parameter, avoiding generating meaningless noise, only oversampling the data of the key parameter, keeping other index samples as they are, and emphasizing the index with high information value and significant significance highly related to electromagnetic sensitivity phenomenon. And finally, a new sample is synthesized by applying an SMOTE algorithm in each selected cluster, and the new sample is added into the minority class set S.
Step 2.3.4: and undersampling a plurality of types of samples in the non-boundary domain by adopting an undersampling method based on Euclidean distance. The first step is to determine the center point of the majority sample set M; secondly, calculating the values of the samples in the plurality of types of sample sets from the center point, and arranging the calculated values in sequence; and thirdly, deleting the majority sample according to the rule from far to near from the center point according to the undersampling multiplying power until the majority sample is balanced with the minority sample data. According to the method, the boundary samples are found out, most class samples which are far away from the classification boundary and have small influence on classification are deleted, and safety samples which have large classification influence and are close to the classification boundary are stored in a training set.
Step 2.3.5: synthesizing the updated minority class samples and majority class samples into a new balanced sample set D new . Novel equalized sample set D new The equalization of sample types is realized, and the data distribution characteristics of the original data set are maintained.
Step 2.4: other signal processing methods
In order to fully utilize the time-frequency domain information, the limitation of the time domain or the frequency domain can be effectively overcome by a time-frequency combination mode, and a two-dimensional time-frequency diagram is generated by using continuous wavelet transformation, wherein the flow is as follows:
step 2.4.1: let a be the scale factor, f s For sampling frequency f c For the wavelet center frequency, the actual frequency f corresponding to the scale factor a a The method comprises the following steps:
f a ＝f c ×f s /a (1)
step 2.4.2: in order for the converted frequency sequence to be an equal sequence, the form of the scale sequence must be as follows:
c/totalscal,…,c/(totalscal-1),c/4,c/2,c (2)
in the above equation, total is the length of the selected scale sequence when CWT (Continuous Wavelet Transform continuous wavelet transform) is performed on the signal, and is generally chosen to be 256, and c is a constant.
Step 2.4.3: from equation (1), the actual frequency corresponding to the scale c/total should be f s 2, then it is possible to obtain:
2×F c ×totalscal (3)
a sequence of scales meeting the above criteria can be obtained.
Step 2.4.4: after determining wavelet basis functions and scales, the wavelet coefficients W are obtained according to the wavelet transformation principle f (a, b), then the corresponding relation between the scale factors and the frequencies is used for solving an actual frequency sequence f, and then a two-dimensional time-frequency transformation chart can be drawn according to a time sequence t.
Step 3: state monitoring layer
The state monitoring layer receives the parameter data provided by the data processing layer and passing through each data processing algorithm, and inputs electromagnetic characteristic parameters which are input in accordance with the neural network into the data-driven fault diagnosis algorithm to determine electromagnetic sensitivity and fault degree so as to monitor the state of the comprehensive module. Dividing the fault type and degree of the obtained key electromagnetic parameters through a classifier; when a fault occurs, the type and the fault degree of the electromagnetic sensitivity phenomenon are rapidly diagnosed, and then the electromagnetic sensitivity boundary is accurately estimated. Based on fault diagnosis, the operation state of the comprehensive module can be acquired and evaluated in real time, and real-time information support is provided for the following health evaluation layer.
Different fault diagnosis methods are set for the one-dimensional signal and the two-dimensional time-frequency data subjected to signal processing, data information is fully utilized, the following step 3.2 is adopted in one dimension, and the following step 3.1 is adopted in two dimensions.
Step 3.1: fault diagnosis method based on multichannel pyramid attention mechanism
And carrying out multidimensional time-frequency domain analysis on multiple signals, and accurately classifying the selected signals by combining the characteristic fusion of a multichannel pyramid attention mechanism by virtue of the advantage of jointly extracting the characteristics of multiple neural network models with different structures. The algorithm diagnostic flow is shown in fig. 7.
The specific flow is as follows:
step 3.1.1: and (5) collecting data and analyzing signals. Respectively carrying out time domain analysis, frequency domain analysis and time-frequency domain analysis on the original signals, and reserving data modes of related signals;
step 3.1.2: and building a neural network model. Respectively constructing residual neural network and cavity convolutional neural network models of all channels, and determining parameters of all modules;
step 3.1.3: a pyramid attention mechanism is constructed. Squeeze-and-Excitation (SE) is a typical attention module assembly. In the traditional neural network, a feature map is obtained after convolution operation and pooling operation, and the weights of all channels are consistent; after the SE component is inserted, the weight of the feature map channel can be adjusted, so that the element information expression is more complete; to some extent, the inactive and unobvious channel features are further suppressed. The PSA attention mechanism (Pyramid Split Attention, PSA) is based on SE attention mechanism, by virtue of its compression and self-learning weight properties, a better hierarchical segmentation can be achieved. Multiple switching between channels is achieved by means of Split and connect modules (SPC). An SPC segmentation module is applied on the basis of an SE attention mechanism, and a complete pyramid attention mechanism model is built;
Step 3.1.4: and (5) data splicing and network connection. And carrying out dimension transformation on the data obtained in the steps, carrying out orderly splicing, inputting the data into a pyramid attention mechanism, and obtaining the final accuracy through a full connection layer and a classification layer.
Step 3.2: fault diagnosis technology based on one-dimensional convolutional neural network
The invention discloses a fault diagnosis method based on a one-dimensional convolutional neural network, which utilizes a CNN (Convolutional Neural Networks convolutional neural network) structure with end-to-end characteristics, avoids the problem of combination of a feature extractor and a classifier, and can directly act on a time domain signal to avoid losing important time domain features. Meanwhile, a CNN model with a hierarchical structure is constructed to further improve the diagnosis precision.
As shown in fig. 8, the first hierarchy of the model is to divide fault types, and the sampling data is divided into different fault types by a CNN classification model; in the second hierarchy, different CNN classification models are constructed for each fault type, further classifying a fault class into several classes of different fault levels, so as to more accurately classify the fault.
The fault diagnosis steps are as follows:
step 3.2.1: (failure type division): in this step, emphasis is placed on constructing a CNN model to determine fault type signatures for each sample. The entire dataset is taken as input to the hierarchical CNN model. Constructing a new CNN, training a network to obtain network parameters, and dividing fault types by using the trained network.
Step 3.2.2: (failure degree division): for a certain fault type divided in the step one, the fault degree may be further divided. The procedure described in step two is similar to step one described above. And constructing a second layer of the model according to the fault type dividing result, respectively constructing a CNN for each fault type, training a network to obtain network parameters, and determining the fault degree.
Step 3.2.3: (network classification accuracy calculation) the classification accuracy of the model is measured by the number of erroneous classification. The final accuracy is calculated by the ratio of the total number of classifications to the total number of samples.
In summary, the state monitoring layer is developed around two aspects of fault diagnosis algorithms, one class of fault diagnosis algorithm research of the neural network is developed aiming at the signal data after time-frequency mixing, the other class of fault diagnosis algorithm research of the one-dimensional convolutional neural network is directly developed based on the one-dimensional signal data, meanwhile, the fault degree is subdivided aiming at different electromagnetic sensitivity phenomena (fault types), and the electromagnetic sensitivity boundary is better evaluated, so that corresponding electromagnetic protection measures are provided according to different fault degrees, and the influence caused by the loss of a key module is reduced.
Step 4: health assessment layer
And (3) expanding the model around a fault prediction method, combining the real-time performance state and the historical data information according to the output of the state monitoring layer, constructing a fault prediction model, and comprehensively judging the health state of the comprehensive module. The fault prediction technology is key feature information obtained through a data processing and feature extraction technology, and is used for predicting the running state of the board card by combining the real-time running state and historical monitoring data, and can be combined with the fault diagnosis technology to provide more fault prediction information.
The invention is based on the fault prediction method of the key feature and the two-way LSTM (Long Short Term Memory long-short-term memory), as shown in figure 9, firstly, the key feature is extracted from the data input in sequence by using a feature extractor; then, establishing a plurality of layers of bidirectional LSTM at the top of the feature extractor to realize the encoding of time information; then constructing a stacked full-connection layer and a linear regression layer on the basis of a time encoder for predicting target data; the predicted data is then input into a fault diagnostor, which analyzes the predicted data to provide further reference information.
SAE (Stacked Auto-Encoder) can extract data features, while SAE-based deep bi-directional LSTM can encode the input sequence information and learn its representation. Mainly consists of four parts: a feature extractor, a time encoder, a full connection layer, a linear regression layer and a fault diagnosis device. A layer of SAE is applied on top of the original input sequence to extract the data features, and after extracting the features, a layer of depth bi-directional LSTM is built on top of the previous SAE for encoding the time information. The two fully connected layers are then stacked together for processing the output of the deep bi-directional LSTM. Finally, a fault diagnosis device is added at the top of the system to diagnose the predicted information so as to provide more predicted information.
Step 5: dynamic adjustment
The dynamic adjustment layer expands the research around the incremental learning algorithm, so that the model in the state monitoring and health evaluation can have the capability of autonomous learning, new knowledge is learned from the continuously acquired real-time data, and the fault diagnosis model and the fault prediction model are updated according to whether the fault diagnosis result output in the fault diagnosis and the fault prediction result output in the fault prediction reach the expectations, thereby improving the accuracy of diagnosis and prediction, and simultaneously enabling the model to have the capability of autonomous learning. Most diagnostic models and predictive models belong to static models, have no autonomous learning capability, and usually after the models are obtained through one-time modeling, model parameters remain unchanged and the influence of newly added data on the models is not considered. However, in actual monitoring, new data are continuously acquired, and incremental learning uses the new data to continuously optimize the model.
The dynamic adjustment flow based on incremental learning is shown in fig. 10, in the running process of the system, the prediction result of the prediction model is compared with actual data, the manually input fault type (default normal state if no input is performed) is compared with the diagnosis result of the diagnosis model, if the fault type is different, the actual data is used for replacing the prediction data to perform reverse error propagation, and the model parameters are updated. By means of the dynamic adjustment method based on incremental learning, the established diagnosis model and the prediction model do not need to be re-established when facing newly-increased data, but only change caused by newly-increased samples is updated on the basis of an original model, so that the trained model is changed to learn hidden knowledge in the newly-increased samples, and meanwhile time cost caused by re-training the model can be well avoided.
Step 6: decision support
The decision support layer is responsible for providing feasible protective measures for related electromagnetic sensitivity phenomena and fault degrees thereof, optimizing the safety performance of the system and guiding the task formulation and the protection scheme. And monitoring other layers in real time, and providing decision basis for system operation by collecting, processing and analyzing various data. Its main functions include:
step 6.1: providing safeguard information
The decision support layer provides feasible protective measure information through analysis and evaluation of related electromagnetic sensitivity phenomena. Corresponding response schemes are recommended for different fault types and degrees, so that the influence of electromagnetic radiation on a system and a comprehensive module is reduced, and the safety of equipment and personnel is ensured.
Step 6.2: optimizing system security performance
The decision support layer optimizes the safety performance of the system by monitoring the state of the system and the fault output condition in real time. And analyzing abnormal parameters and fault output conditions, rapidly identifying potential problems, taking preventive measures, avoiding faults and improving the stability and reliability of the system.
Step 6.3: instruction task formulation and protection scheme
And the data analysis and decision result of the decision support layer provide important guidance for the task planning and the formulation of the protection scheme. Through the evaluation of the fault type and degree, the system can help users reasonably arrange tasks and formulate corresponding protection strategies so as to ensure the smooth progress of the tasks and the safe operation of equipment.
The research result has a wide application direction in the aspect of electromagnetic compatibility (EMC), and is helpful for improving the reliability and stability of equipment and ensuring the normal operation of the equipment in a complex electromagnetic environment.
The main focus is on how to identify and solve electromagnetic compatibility faults, and ensure reliable operation of equipment under electromagnetic interference:
monitoring and maintaining electronic equipment: and the real-time monitoring of the equipment is realized by utilizing an electromagnetic characteristic data set and an electromagnetic compatibility fault diagnosis method. If the equipment has electromagnetic compatibility problems, the equipment can be rapidly diagnosed and maintenance measures can be taken, so that the reliability of the equipment is ensured.
Abnormal fault feature analysis: by exploring the correlation between the abnormal fault in the electromagnetic property data and the key electromagnetic property parameters, the fault source can be rapidly located. This helps take corrective action quickly.
And (3) fault prediction: based on historical data, future electromagnetic compatibility failures can be predicted using deep learning or machine learning algorithms. This helps to take preventive maintenance measures to reduce the impact of equipment failure on task execution.
And (3) state monitoring: electromagnetic characteristic parameters of the electronic equipment can be monitored in real time. If an abnormal change in the parameter occurs, the system may issue an alarm to take action in time.
These achievements may be applied to a variety of devices and systems, including but not limited to:
communication apparatus: reliable operation of the communication equipment under electromagnetic interference is ensured, and connectivity of the command and control system is maintained.
Radar system: the radar system is ensured to accurately operate in an electromagnetic interference environment, and accurate target tracking and information are provided.
Navigation system: the stability and the accuracy of the navigation equipment are maintained, and the equipment can be accurately navigated.
And (3) collecting information: communication and information transmission modes of hostile force are identified by monitoring electromagnetic activity, and support is provided for information collection and information war.
The above examples represent only 1 embodiment of the present invention, which is described in more detail and detail, but are not to be construed as limiting the scope of the invention. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the invention, which are all within the scope of the invention. Accordingly, the scope of protection of the present invention is to be determined by the appended claims.
Claims (10)
1. The electromagnetic sensitivity prediction and health management method of the comprehensive module is characterized by comprising a data acquisition layer, a data processing layer, a state monitoring layer, a health evaluation layer, a dynamic adjustment layer and a decision support layer, and comprises the following steps of:
Step 1: and (3) data acquisition: based on the conduction time domain sensitivity test, a matched test platform is constructed, comprising: interference injection test, electromagnetic parameter selection and real-time monitoring tool design;
step 2: and (3) data processing: through comprehensive data processing steps to extract useful information and ensure data quality, analyzing and processing input signals from a data acquisition layer to generate meaningful state descriptions;
step 3: and (3) state monitoring: the state monitoring layer receives the parameter data provided by the data processing layer and passing through each data processing algorithm, and inputs electromagnetic characteristic parameters which are input in accordance with the neural network to a data-driven fault diagnosis algorithm to determine electromagnetic sensitivity and fault degree so as to monitor the state of the comprehensive module; dividing the fault type and degree of the obtained key electromagnetic parameters through a classifier; when a fault occurs, the type and the fault degree of an electromagnetic sensitivity phenomenon are rapidly diagnosed, and then the electromagnetic sensitivity boundary is accurately estimated; based on fault diagnosis, the running state of the comprehensive module can be acquired and evaluated in real time, and real-time information support is provided for a later health evaluation layer;
step 4: health assessment: adopting a fault prediction method, combining the real-time performance state and the historical data information according to the output of the state monitoring layer, constructing a fault prediction model, and comprehensively judging the health state of the comprehensive module; the fault prediction technology is key feature information obtained through a data processing and feature extraction technology, and is used for predicting the running state of the board card by combining the real-time running state and historical monitoring data, and can be combined with a fault diagnosis technology to provide more fault prediction information;
Step 5: dynamic adjustment: the dynamic adjustment layer is unfolded around the incremental learning algorithm, so that the model in state monitoring and health evaluation can have the capability of autonomous learning, new knowledge is learned from real-time data which are continuously acquired, and the fault diagnosis model and the fault prediction model are updated according to whether a fault diagnosis result output in fault diagnosis and a fault prediction result output in fault prediction reach expectations or not, thereby improving the accuracy of diagnosis and prediction, and simultaneously enabling the model to have the capability of autonomous learning;
step 6: decision support: the decision support layer is responsible for providing feasible protective measures for related electromagnetic sensitivity phenomena and fault degrees thereof, optimizing the safety performance of the system and guiding the task formulation and the protection scheme; and various data of other layers are collected, processed and analyzed to provide decision basis for system operation.
2. The electromagnetic sensitivity prediction and health management method of the integrated module according to claim 1, wherein the interference injection test, electromagnetic parameter selection and real-time monitoring tool in step 1 is specifically: setting up a whole set of test platform for the interference injection test, designing a whole set of test flow, and exciting the sensitive phenomenon of the tested equipment; the electromagnetic compatibility characteristic parameter selection method is to study which working parameters of the tested modules have stronger correlation with the sensitive phenomenon according to the touch of the sensitive characteristic of the board card in the early test, and provide design basis for the follow-up real-time monitoring of the board card; the real-time monitoring tool integrates the sensor, the monitoring circuit and the data interface, and improves the effectiveness of monitoring parameters by designing and optimizing monitoring points, so that support is provided for follow-up intelligent processing.
3. The method for electromagnetic sensitivity prediction and health management of an integrated module according to claim 1, wherein step 1 specifically comprises:
step 1.1: the injection test needs to determine the monitoring point position through hardware measurement or a design drawing of a hardware drawing generation module tool;
step 1.2: manufacturing a simulation motherboard to cooperate with a test tool to develop a sensitivity threshold test of the module, and establishing a corresponding relation among the type of an interference signal, the intensity of the interference signal, the injection position, the tested module, the working state, the sensitivity phenomenon and the electromagnetic signal intensity of a monitoring point;
step 1.3: applying the test fixture to the assembly, installing a tested module to perform equipment-level test work, establishing a mathematical relationship from interference to monitoring points, and positioning accumulated data for the problem;
step 1.4: aiming at the packaged sensitive test data, electromagnetic characteristic data acquisition and collection are carried out;
the electromagnetic characteristics of the modules under different states are respectively inspected by enabling the tested object to be in a certain working state;
step 1.5: the stable operation of the case is kept, and faults are eliminated;
the data sample is mainly composed of a power port, a network port and a video port, wherein after the 3 ports are respectively injected with interference specified by a test, the recorded frequency domain information is recorded; the characteristic parameters are frequency, voltage and monitoring signal intensity respectively, and a sample is generated according to the mode; obtaining test data from different dimensions requires selection and setting according to the functional characteristics of the board.
4. The electromagnetic sensitivity prediction and health management method of an integrated module according to claim 3, wherein the specific steps in step 1.5 are as follows:
step 1.5.1: the chassis is electrified and preheated to enable the chassis to reach a stable working state;
step 1.5.2: monitoring the state of the board card, and ensuring the normal work of oscilloscope data acquisition;
step 1.5.3: acquiring electromagnetic characteristic data at the moment as a reference and simultaneously observing the working performance of the board card in real time;
step 1.5.4: connecting the signal generator and the power amplifier together to a power port on the tested module through a current injection clamp;
step 1.5.5: various electromagnetic interference signals specified by the test are injected into the power port according to a certain amount until the board card has abnormal conditions in the sensitive criterion;
step 1.5.6: comparing and calibrating the electromagnetic characteristic data obtained at the moment with electromagnetic characteristic data in normal operation, and storing voltage data of each test point by using an oscilloscope;
step 1.5.7: reducing the interference field intensity until the work of the board card is recovered to be normal;
step 1.5.8: steps 1.5.4-1.5.6 are repeated until no new failure of the chassis occurs.
5. The electromagnetic sensitivity prediction and health management method of the integrated module according to claim 1, wherein step 2 specifically comprises:
Step 2.1: and (3) data preprocessing design: firstly, filling missing values, smoothing noise data, smoothing or deleting outliers to solve the problem of data inconsistency, and secondly, aiming at time series data related to electromagnetic compatibility obtained from a board card, obtaining periodic characteristics by extracting average values, variances, maximum and minimum values of each parameter and applying trigonometric function fitting; finally, normalizing the features to ensure that the features have the same scale;
step 2.2: the data analysis method comprises the following steps: the key steps in the data analysis and feature selection process are used for improving the selection of monitoring parameters, the monitoring efficiency and the accuracy; key features related to the sensitive phenomenon are screened out through a visual feature map, a feature selection function and a tree-based related feature selection method, the effectiveness of each parameter information corresponding to the sensitive phenomenon is analyzed, the selection of monitoring points and monitoring parameters is continuously optimized, and support is provided for monitoring and understanding the sensitive phenomenon;
step 2.3: data balancing problem solving strategy: in order to improve the classification accuracy of few samples, an unbalanced data processing strategy is used, and an undersampling method based on Euclidean distance and an oversampling method based on K-Means SMOTE of cluster distribution are combined to reconstruct sample data;
Step 2.4: other signal processing methods: in order to fully utilize the time-frequency domain information, the limitation of the time domain or the frequency domain can be effectively overcome by a time-frequency combination mode, and a two-dimensional time-frequency diagram is generated by using continuous wavelet transformation.
6. The method for electromagnetic susceptibility prediction and health management of an integrated module of claim 5,
the step 2.1 specifically comprises the following steps:
step 2.1.1: selecting a fixed time window length for dividing the data segments in the time series data;
step 2.1.2: the following steps are performed for each parameter: calculating the average value, variance, maximum value and minimum value of the parameters in a given time window; these statistics provide the basic information of the data distribution, reflecting the trend of the variation and the amplitude fluctuation of the parameters within the time window; the average value is the central trend of the data distribution, representing the overall level of the data over a time window; the variance measures the degree of dispersion of the data and reflects the fluctuation of the data; the maximum and minimum values represent the highest and lowest points of the data within the time window, respectively;
step 2.1.3: performing discrete Fourier transform on the time series data of the parameters by using discrete Fourier transform, and converting the signals from a time domain to a frequency domain; the result of Fourier transformation is analyzed, main frequency components can be identified, the periodic characteristics of representative parameters are subjected to trigonometric function fitting on the selected frequency components, frequency information in fitting parameters is obtained, and the periodic characteristics are calculated;
Step 2.1.4: normalizing the eigenvalues obtained from each parameter to ensure that they are consistent in scale;
step 2.2 is specifically:
step 2.2.1: making a clustering scatter diagram, clustering according to characteristic values, displaying data points of different clusters on the diagram by using different colors or marks, and observing the distribution of the data points on the diagram to see whether natural data clustering phenomenon exists or not, so as to judge whether a certain association relationship exists between the characteristics or not;
step 2.2.2: making a thermodynamic diagram, displaying a correlation coefficient matrix between the features, and filtering high-correlation features, wherein the closer the value of the correlation coefficient is 1 or-1, the stronger the correlation between the two features is represented; deleting one of the features with the correlation of 1 or-1 in the thermodynamic diagram;
step 2.2.3: based on the related feature selection of the feature function, calculating the relation degree between each feature and the result by carrying out statistical analysis, sequencing from high score to low score, and selecting the features with the highest scores for modeling; the feature number is selected by means of cross verification, verification results are carried out on the initially selected feature parameters, higher accuracy can be obtained when the feature parameters with higher use scores are displayed, and at the moment, the remaining features with the lowest scores can be discarded;
Step 2.2.4, based on relevant feature selection of the tree model, evaluating feature importance by utilizing a decision tree or random forest algorithm, drawing an importance ranking chart, and selecting the first few main features for training verification according to importance ranking;
the step 2.3 is specifically as follows:
step 2.3.1: classifying the state data sample set D according to the sample attribute of each module board card; a minority class sample set is defined as S, and a majority class sample set is defined as M;
step 2.3.2: resampling is carried out under the classification according to the unbalanced data set, and the resampling proportion is set;
step 2.3.3: oversampling a minority class of samples by using a K-Means SMOTE method for generating samples based on cluster distribution; firstly, based on a K-means algorithm, K groups are clustered by using a K-means, clusters used for oversampling are filtered and selected, and clusters with a high proportion of minority samples are reserved; then, the number of the synthesized samples is distributed, and most samples are distributed to clusters with sparse distribution of few samples; only the data of the key parameters are oversampled, other index samples are kept as they are, and the indexes with high information value and high meaning related to electromagnetic sensitivity phenomena are emphasized; finally, a SMOTE algorithm is applied to each selected cluster to synthesize a new sample, and the new sample is added into a minority class set S;
Step 2.3.4: undersampling a majority of samples in a non-boundary domain by adopting an undersampling method based on Euclidean distance; the first step is to determine the center point of the majority sample set M; secondly, calculating the values of the samples in the plurality of types of sample sets from the center point, and arranging the calculated values in sequence; thirdly, deleting most samples according to the rule from far to near from the center point according to the undersampling multiplying power until the most samples are balanced with the data of few samples; according to the method, a boundary sample is found, most types of samples far away from a classification boundary and small in influence on classification are deleted, and safety samples with large classification influence and close to the classification boundary are stored in a training set;
step 2.3.5: synthesizing the updated minority class samples and majority class samples into a new balanced sample set D new The method comprises the steps of carrying out a first treatment on the surface of the Novel equalized sample set D new Not only realize sample classOther equalization, the data distribution characteristics of the original data set are maintained;
step 2.4 is specifically: other signal processing methods:
step 2.4.1: let a be the scale factor, f s For sampling frequency f c For the wavelet center frequency, the actual frequency f corresponding to the scale factor a a The method comprises the following steps:
f a ＝f c ×f s /a (1)
step 2.4.2: in order for the converted frequency sequence to be an equal sequence, the form of the scale sequence must be as follows:
c/totalscal，...，c/(totalscal-1)，c/4，c/2，c (2)
In the above description, total is the length of the selected scale sequence when the signal is subjected to continuous wavelet transformation, and is generally selected to be 256, and c is a constant;
step 2.4.3: from equation (1), the actual frequency corresponding to the scale c/total should be f s 2, then it is possible to obtain:
2×F c ×totalscal (3)
a scale sequence meeting the above criteria can be obtained;
step 2.4.4: after determining wavelet basis functions and scales, the wavelet coefficients W are obtained according to the wavelet transformation principle f (a, b), then the corresponding relation between the scale factors and the frequencies is used for solving an actual frequency sequence f, and then a two-dimensional time-frequency transformation chart can be drawn according to a time sequence t.
7. The electromagnetic sensitivity prediction and health management method of the integrated module according to claim 1, wherein the step 3 is specifically:
setting different fault diagnosis methods for a one-dimensional signal and a two-dimensional time-frequency data subjected to signal processing, fully utilizing data information, wherein the one-dimensional signal adopts the following step 3.2, and the two-dimensional signal adopts the following step 3.1;
step 3.1: a fault diagnosis method based on a multichannel pyramid attention mechanism comprises the following steps:
the multi-dimensional time-frequency domain analysis is carried out on various signals, and the selected signals are accurately classified by means of the advantage that characteristics are extracted together by the neural network models with various different structures and combining with the characteristic fusion of a multi-channel pyramid attention mechanism;
The specific flow is as follows:
step 3.1.1: collecting data and analyzing signals; respectively carrying out time domain analysis, frequency domain analysis and time-frequency domain analysis on the original signals, and reserving data modes of related signals;
step 3.1.2: building a neural network model; respectively constructing residual neural network and cavity convolutional neural network models of all channels, and determining parameters of all modules;
step 3.1.3: constructing a pyramid attention mechanism; after the SE component is inserted, the weight of the feature map channel can be adjusted, so that the element information expression is more complete; to a certain extent, the channel characteristics of inactivity and unobvious are further inhibited; an SPC segmentation module is applied on the basis of an SE attention mechanism, and a complete pyramid attention mechanism model is built;
step 3.1.4: data splicing and network connection; the data obtained in the steps are subjected to dimension transformation, orderly spliced, input into a pyramid attention mechanism and finally obtained through a full-connection layer and a classification layer;
step 3.2: fault diagnosis technology based on one-dimensional convolutional neural network:
the fault diagnosis method based on the one-dimensional convolutional neural network utilizes a CNN structure with end-to-end characteristics, so that the problem of combination of a feature extractor and a classifier is avoided, and the method can directly act on a time domain signal to avoid losing important time domain features; meanwhile, a CNN model with a hierarchical structure is constructed to further improve the diagnosis precision; the first level of the model is to divide fault types, and the sampling data are divided into different fault types through a CNN classification model; in the second hierarchy, different CNN classification models are respectively constructed for each fault type, and a certain fault type is further divided into a plurality of classes with different fault degrees so as to more accurately classify the faults;
The fault diagnosis steps are as follows:
step 3.2.1: fault type classification: in this step, emphasis is placed on constructing a CNN model to determine fault type labels for each sample; taking the whole data set as the input of a hierarchical CNN model; constructing a new CNN, training a network to obtain network parameters, and dividing fault types by using the trained network;
step 3.2.2: fault degree division: for a certain fault type divided in the previous step, the fault degree can be further divided; the procedure described in step 3.2.2 is similar to the previous step; constructing a second layer of the model according to the fault type dividing result, respectively constructing a CNN for each fault type, training a network to obtain network parameters, and determining the fault degree;
step 3.2.3: calculating the network classification precision: the classification precision of the model is measured through the error classification times; the final accuracy is calculated by the ratio of the total number of classifications to the total number of samples;
the state monitoring layer is divided into two types, one type is used for carrying out multi-channel neural network fault diagnosis algorithm research on signal data after time-frequency mixing, the other type is used for directly carrying out one-dimensional convolutional neural network fault diagnosis algorithm research on the basis of one-dimensional signal data, and meanwhile, fault degrees are subdivided according to different electromagnetic sensitivity phenomena, electromagnetic sensitivity boundaries are better evaluated, so that corresponding electromagnetic protection measures are set according to different fault degrees, and the influence caused by key module loss is reduced.
8. The method for electromagnetic sensitivity prediction and health management of an integrated module according to claim 1, wherein step 4 specifically comprises:
firstly, extracting key features from sequentially input data by using a feature extractor; then, establishing a plurality of layers of bidirectional LSTM at the top of the feature extractor to realize the encoding of time information; then constructing a stacked full-connection layer and a linear regression layer on the basis of a time encoder for predicting target data; the predicted data is then input into a fault diagnostor, and analyzed to provide further reference information;
SAE can extract data features, while SAE-based depth bi-directional LSTM can encode the input sequence information and learn its representation; comprises four major parts: the device comprises a feature extractor, a time encoder, a full connection layer, a linear regression layer and a fault diagnosis device; applying a layer of SAE on the original input sequence to extract the data features, after extracting the features, constructing a multi-layer depth bi-directional LSTM on top of the previous SAE for encoding time information; then stacking two fully connected layers together for processing the output of the deep bi-directional LSTM; finally, a fault diagnosis device is added at the top of the system to diagnose the predicted information so as to provide more predicted information.
9. The method for electromagnetic sensitivity prediction and health management of an integrated module according to claim 1, wherein step 5 specifically comprises:
based on the dynamic adjustment of incremental learning, comparing the predicted result of the predicted model with actual data in the running process of the system, comparing the manually input fault type with the diagnosis result of the diagnosis model, and defaulting to a normal state if no input exists; if the model parameters are different, replacing the predicted data with the actual data to carry out reverse error propagation, and updating the model parameters; by the dynamic adjustment method based on incremental learning, the established diagnosis model and the prediction model do not need to be re-established when facing the newly added data, but only update the change caused by the newly added sample on the basis of the original model.
10. The electromagnetic sensitivity prediction and health management method of the integrated module according to claim 1, wherein step 6 specifically comprises:
providing safeguard information: the decision support layer provides feasible protective measure information through analysis and evaluation of related electromagnetic sensitivity phenomena; corresponding response schemes are recommended aiming at different fault types and degrees, so that the influence of electromagnetic radiation on a system and a comprehensive module is reduced, and the safety of equipment and personnel is ensured;
Optimizing the safety performance of the system: the decision support layer optimizes the safety performance of the system by monitoring the state of the system and the fault output condition in real time; analyzing abnormal parameters and fault output conditions, rapidly identifying potential problems, taking preventive measures, avoiding faults and improving the stability and reliability of the system;
and (3) guiding task formulation and protection scheme: the data analysis and decision result of the decision support layer provide important guidance for the task planning and the formulation of the protection scheme; through the evaluation of the fault type and degree, the system can help users reasonably arrange tasks and formulate corresponding protection strategies so as to ensure the smooth progress of the tasks and the safe operation of equipment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311617209.8A CN117609836A (en) | 2023-11-29 | 2023-11-29 | Electromagnetic sensitivity prediction and health management method for integrated module |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311617209.8A CN117609836A (en) | 2023-11-29 | 2023-11-29 | Electromagnetic sensitivity prediction and health management method for integrated module |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117609836A true CN117609836A (en) | 2024-02-27 |
Family
ID=89951236
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311617209.8A Pending CN117609836A (en) | 2023-11-29 | 2023-11-29 | Electromagnetic sensitivity prediction and health management method for integrated module |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117609836A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117806914A (en) * | 2024-02-29 | 2024-04-02 | 潍坊鼎好信息科技有限公司 | Computer fault monitoring and alarming system |
-
2023
- 2023-11-29 CN CN202311617209.8A patent/CN117609836A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117806914A (en) * | 2024-02-29 | 2024-04-02 | 潍坊鼎好信息科技有限公司 | Computer fault monitoring and alarming system |
| CN117806914B (en) * | 2024-02-29 | 2024-05-07 | 潍坊鼎好信息科技有限公司 | Computer fault monitoring and alarming system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101634605B (en) | Intelligent gearbox fault diagnosis method based on mixed inference and neural network | |
| CN108038300B (en) | Optical fiber state evaluation method based on improved membership function combined with neural network | |
| Byington et al. | Data-driven neural network methodology to remaining life predictions for aircraft actuator components | |
| CN113469060A (en) | Multi-sensor fusion convolution neural network aeroengine bearing fault diagnosis method | |
| CN110702418A (en) | Aircraft engine fault prediction method | |
| US20220308568A1 (en) | System and method for monitoring soil gas and performing responsive processing on basis of result of monitoring | |
| CN104035431B (en) | The acquisition methods of kernel functional parameter and system for non-linear process monitoring | |
| CN112036042B (en) | Power equipment abnormality detection method and system based on variational modal decomposition | |
| CN112785091A (en) | Method for performing fault prediction and health management on oil field electric submersible pump | |
| CN117609836A (en) | Electromagnetic sensitivity prediction and health management method for integrated module | |
| CN116380445B (en) | Equipment state diagnosis method and related device based on vibration waveform | |
| CN114692758A (en) | Power communication fault analysis method and device, terminal equipment and medium | |
| CN117235524A (en) | Learning training platform of automatic valuation model | |
| CN115392782A (en) | Method and system for monitoring and diagnosing health state of process system of nuclear power plant | |
| CN117371207A (en) | Extra-high voltage converter valve state evaluation method, medium and system | |
| CN117191147A (en) | Flood discharge dam water level monitoring and early warning method and system | |
| Tolas et al. | Periodicity detection algorithm and applications on IoT data | |
| Dang et al. | seq2graph: Discovering dynamic non-linear dependencies from multivariate time series | |
| CN114046816B (en) | Sensor signal fault diagnosis method based on lightweight gradient lifting decision tree | |
| Techane et al. | Rotating machinery prognostics and application of machine learning algorithms: Use of deep learning with similarity index measure for health status prediction | |
| CN116910574A (en) | Mechanical equipment diagnosis method and device based on meta-learning and time convolution network | |
| CN115659271A (en) | Sensor abnormality detection method, model training method, system, device, and medium | |
| CN114266013A (en) | Deep learning virtual perception network-based transmission system vibration decoupling method | |
| RU2699685C1 (en) | Method of analyzing and monitoring the state of a technical installation comprising a plurality of dynamic systems | |
| CN112834194B (en) | Fault intelligent detection method based on soft target measurement under fault-free sample |
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 |