CN111751714A - Radio frequency analog circuit fault diagnosis method based on SVM and HMM - Google Patents
Radio frequency analog circuit fault diagnosis method based on SVM and HMM Download PDFInfo
- Publication number
- CN111751714A CN111751714A CN202010528645.8A CN202010528645A CN111751714A CN 111751714 A CN111751714 A CN 111751714A CN 202010528645 A CN202010528645 A CN 202010528645A CN 111751714 A CN111751714 A CN 111751714A
- Authority
- CN
- China
- Prior art keywords
- data
- state
- radio frequency
- model
- training
- 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
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/28—Testing of electronic circuits, e.g. by signal tracer
- G01R31/316—Testing of analog circuits
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The invention relates to a radio frequency analog circuit fault diagnosis technology, in particular to a radio frequency analog circuit fault diagnosis method based on SVM and HMM, which extracts effective fault feature information by using a low noise amplifier circuit aiming at ATF54143 as a core and processes data features, wherein the method comprises the combination of a K-Means algorithm and a support vector machine, thereby reducing the training time of the HMM, solving the problem that the HMM is suitable for processing and classifying the small sample data and the support vector machine is likely to have misjudgment risks due to the fact that the maximum likelihood probability is too close or equal when determining the maximum likelihood probability of the most similar model comparison in the training process of the HMM. The introduction of K-Means aims to cluster experimental data to form data with label categories, and then put the data into a support vector machine model to perform more accurate classification.
Description
Technical Field
The invention relates to a radio frequency analog circuit fault diagnosis technology, in particular to a radio frequency analog circuit fault diagnosis method based on an SVM (support vector machine) and an HMM (hidden Markov model).
Background
With the widespread application of high-power circuits, it is very important and urgent to grasp the health status information of the power circuit and predict the life of the core semiconductor device in real time. However, the conventional fault diagnosis methods such as "timing detection" and "post-repair" have failed to meet the actual requirements, and it is very important to perform fault diagnosis and fault prediction on the power circuit and the key power devices therein in order to improve the reliability, maintainability, supportability, and safety of the equipment and the system. The practical application of fault Prediction and Health Management (PHM) in electronic systems has become a future development trend, and PHM refers to the real-time monitoring and collection of various state parameters of the system by using advanced sensors, the evaluation of system states of the collected parameters by establishing models and algorithms, and the prediction of system faults, and provides a reasonable decision for the maintenance and guarantee of the system to realize the state maintenance of the system. Therefore, the essence of the failure prediction of the radio frequency analog circuit is a problem of pattern classification and identification, which mainly lies in two aspects: firstly, extracting and preprocessing characteristic parameters of a research object, and secondly, diagnosing faults by using a mode identification method. With the rapid development of artificial intelligence, the core technology of machine learning is also continuously applied to aspects such as computer image recognition, route planning and navigation, voice recognition, medical diagnosis and the like.
The method is applied to the field of fault prediction and health management, a circuit fault model is established by combining HMM and K-Means-SVM, fault recognition is carried out on the circuit state, and the service life of the used semiconductor chip is further predicted.
Disclosure of Invention
The invention provides a radio frequency analog circuit fault diagnosis method based on an SVM and an HMM, so that the reliability, maintainability, supportability and safety of a radio frequency analog circuit related to equipment or a system are improved, the category difference of the radio frequency analog circuit related to the equipment or the system can be reflected to the maximum extent, and more accurate classification is carried out.
The invention aims to realize the method for diagnosing the faults of the radio frequency analog circuit based on the SVM and the HMM, which is characterized in that: at least comprises the following steps:
100. the method comprises the steps of determining a radio frequency analog circuit, and carrying out fault prediction research by adopting a low-noise amplification circuit which is widely applied in the radio frequency circuit, wherein the noise coefficient of the radio frequency analog circuit is lower than that of a common amplifier, and the radio frequency analog circuit is generally used for high-frequency or intermediate-frequency preamplifiers of various radio receiving devices such as communication, radars, electronic countermeasure and the like.
101. The device in the radio frequency analog circuit is determined, and through consideration of working frequency, noise and the like, the ATF54143 is taken as an E-PHEMT with high gain, wide dynamic range and low noise, and is relatively suitable to be taken as a research object for feature extraction of radio frequency circuit fault diagnosis and prediction.
102. Obtaining the fault prediction characteristics of devices in the radio frequency analog circuit by adopting ADS simulation software;
103. carrying out quantitative processing and storage on the determined technical parameters of the radio frequency analog circuit device to obtain the fault characteristics expressed by the change of the technical parameters;
104. the ADS is adopted to complete circuit design, temperature is used as an environment change factor in a simulation mode, multi-dimensional parameter attributes are extracted, the actual working state of the low-noise amplifier circuit is represented, and matching of minimum noise and maximum gain is included;
105. clustering the extracted multi-dimensional parameter attribute data, and representing the multi-dimensional parameter attribute as a cluster with a label category through a K-Means algorithm;
106. randomly selecting a part of data of each category from the results after the K-Means clustering as a test sample set;
107. selecting the remaining data as a training sample set after the test sample set is selected;
108. selecting a left and right HSMM type without jumping, determining initial parameters, and performing model training through an EM algorithm;
109. obtaining an HSMM model of the state I and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state I;
110. obtaining an HSMM model of the state II and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state II;
111. obtaining an HSMM model of the state III and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state III;
112. obtaining an HSMM model of the state four and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state four;
113. taking the relation curves obtained in the steps 109, 110, 111 and 112 in a one-to-one correspondence manner as the input of the support vector machine classifier;
114. training a support vector machine classifier, and performing class prediction by using a trained support vector machine algorithm;
115. training and classifying by a support vector machine algorithm to obtain the classification accuracy of the state data;
116. and training a full-life state model, and representing the residual life of each state by using a time-dwell function in the HSMM model.
The steps 100, 102, 103, and 104 are specifically given by the following processes:
200. obtaining the minimum noise coefficient, gain, third-order intermodulation point and 1dB compression point contained in the device in the radio frequency analog circuit by taking the datasheet of the determined change curve of the device in the radio frequency analog circuit under the conditions of the maximum power, voltage, current and temperature which can be borne by the device and the key performance of the transistor under different frequencies and different biases as the basis, and obtaining data which can represent the state characteristics of the device by taking the internal structure of the device as a model;
201. DC analysis to determine the DC operating point of the device, combined with datasheet and bias circuit to determine the DC operating point of the device is (3V, 28.6mA)
202. Designing a direct-current working point biasing circuit;
203. and (3) analyzing stability, carrying out S parameter simulation, and enabling a Stabfact control in the ADS to represent the stability to be more than 1. And adding small inductors to two sources of the transistor to be used as negative feedback, and repeatedly adjusting feedback inductance values to ensure that the transistor is stable in the working frequency range.
ADS provides DA smithhartmatch matching, and after no-reflection matching a conjugate of the minimum noise figure impedance is obtained, Z ═ 24.9-j6.2) Ω.
205. The smith chart is used to obtain the output impedance that maximizes gain.
206. All matching is completed and a schematic diagram of the overall amplifier circuit can be obtained.
The steps 105 and 106 further include:
300. inputting a data set D without labels for K-Means;
the K-Means algorithm divides original data into K data clusters by taking the distance between data as a standard, so that the intra-cluster similarity is high and the inter-cluster similarity is low;
302. and outputting cluster division results after the K-Means cluster analysis.
The steps 106 and 107 specifically include:
400. and determining a k value, namely the number of categories obtained by clustering the expected initial data, and selecting k data points as centroids in the data set according to experience or at random.
401. Carrying out initialization;
402. calculating the distances from the points except the clustering center in the data set to the k centroids, and dividing all sample points into clusters of each centroid according to the shortest clustering among the clusters;
403. forming k data clusters and updating cluster division;
404. recalculating centroids for the k sets;
405. and repeating the steps, and stopping the algorithm when the criterion function is converged or the cluster between the newly calculated centroid and the original centroid is in the set threshold range.
The steps 106 and 114 specifically include:
600. after K-Means cluster analysis, taking data with label category information as input of a support vector machine algorithm, and training and testing a classifier;
601. the normalization of data is realized by applying a mapminmax function in an MATLAB toolbox;
602. in order to improve the classification accuracy, parameters c and g of the kernel function are optimized by adopting a grid search method;
603. establishing a model by using a libsvmtrain function and a libsvmpredict function in an LIBSVM tool box, and training and predicting;
604. running a program to obtain the classification accuracy;
605. and under the condition that the accuracy is not high enough, changing the optimization mode or certain parameters and optimizing again.
The steps 107, 108, 109, 110, 111 and 112 comprise:
700. gradually extracting fault parameters representing the health state of the low-noise discharging circuit by using the temperature as a variation factor;
701. determining HSMM model category, and determining initial parameters A, B, Π and D. Inputting data after the K-Means clustering, and performing model training by using an EM algorithm;
702. training the full-life data through an EM algorithm to obtain an HSMM model with all states;
703. after the model is trained, a reestimated parameter value can be obtained, wherein the reestimated parameter value comprises the mean value and the variance of the residence time distribution function of each state, and the residence time of each state can be further obtained;
704. inputting the test set as an observation sequence into the trained HSMM model;
705. through model training, the relationship between the iteration step number and the maximum likelihood function of each observation state can be obtained;
706. putting the maximum likelihood function and the state labels into a trained support vector machine classifier in a one-to-one correspondence manner, and checking the classification accuracy;
707. and determining the confidence interval of the remaining service life of the current state according to the mean value and the variance of the state residence time.
The steps specifically include:
800. taking the collected circuit performance parameters as characteristic data, selecting a proper HSMM model, and determining initial parameters and iteration steps;
801. initializing the B and the time residence function;
802. training iteration is carried out by using an EM algorithm;
803. setting an iteration variable, and starting iteration from 1;
804. calculating the maximum likelihood probability by using a Viterbi algorithm;
805. judging whether a convergence condition is reached;
806. continuing iteration if the convergence condition is not met;
807. the maximum number of iteration steps is reached;
808. if the maximum iteration step number is exceeded, the model does not reach convergence, and quitting is performed;
809. and obtaining a training model when a convergence condition is reached.
The invention has the advantages that: the method extracts effective fault feature information by using a low noise amplifier circuit aiming at ATF54143 as a core, processes data features, wherein the fault feature information comprises the combination of a K-Means algorithm and a support vector machine, reduces the training time of the HMM, solves the problem that when the HMM determines the maximum likelihood probability of a most similar model comparison on small sample data in the training process, the maximum likelihood probability is too close or equal, the risk of misjudgment is possibly caused, and the support vector machine is suitable for processing and classifying and can reflect the category difference to the maximum extent. The introduction of K-Means aims to cluster experimental data to form data with label categories, and then put the data into a support vector machine model to perform more accurate classification.
Drawings
The invention is further illustrated with reference to the accompanying drawings of embodiments:
FIG. 1 is a flow chart of an embodiment of the present invention;
FIG. 2 is a design drawing thereof by ADS;
FIG. 3 is a graph of the effect of K-Means cluster analysis;
FIG. 4 is a schematic flow chart of a clustering algorithm;
FIG. 5 is a diagram of the effect of clustering algorithm in MATLAB on data processing;
FIG. 6 is a schematic diagram of an implementation of a support vector machine;
FIG. 7 is a diagram of a radio frequency circuit fault prediction scheme based on a combination of a support vector machine and a hidden Markov model;
FIG. 8 is a flow diagram of an algorithmic implementation of parameter estimation in a hidden Markov model;
fig. 9 is a diagram of the test results of maximum likelihood probability by an SVM classifier.
Detailed Description
As shown in fig. 1, a method for diagnosing faults of a radio frequency analog circuit based on SVM and HMM is characterized in that: at least comprises the following steps:
100. the method comprises the steps of determining a radio frequency analog circuit, and carrying out fault prediction research by adopting a low-noise amplification circuit which is widely applied in the radio frequency circuit, wherein the noise coefficient of the radio frequency analog circuit is lower than that of a common amplifier, and the radio frequency analog circuit is generally used for high-frequency or intermediate-frequency preamplifiers of various radio receiving devices such as communication, radars, electronic countermeasure and the like.
101. The device in the radio frequency analog circuit is determined, and through consideration of working frequency, noise and the like, the ATF54143 is taken as an E-PHEMT (enhanced mode pseudomorphic high electron mobility transistor) with high gain, wide dynamic range and low noise, and is relatively suitable to be used as a research object for feature extraction of fault diagnosis and prediction of the radio frequency circuit.
102. Obtaining the fault prediction characteristics of devices in the radio frequency analog circuit by adopting ADS simulation software;
103. carrying out quantization processing and storage on the determined technical parameters of the radio frequency analog circuit device; obtaining the fault characteristics of the technical parameter change expression;
104. the ADS is adopted to complete circuit design, temperature is used as an environment change factor in a simulation mode, multi-dimensional parameter attributes are extracted, the actual working state of the low-noise amplifier circuit is represented, and matching of minimum noise and maximum gain is included;
105. clustering the extracted multi-dimensional parameter attribute data, and representing the multi-dimensional parameter attribute as a cluster with a label category through a K-Means algorithm;
106. randomly selecting a part of data of each category from the results after the K-Means clustering as a test sample set;
107. selecting the remaining data as a training sample set after the test sample set is selected;
108. selecting a left and right HSMM type without jumping, determining initial parameters, and performing model training through an EM algorithm;
109. obtaining an HSMM model of the state I and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state I;
110. obtaining an HSMM model of the state II and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state II;
111. obtaining an HSMM model of the state III and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state III;
112. obtaining an HSMM model of the state four and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state four;
113. taking the relation curves obtained in the steps 109, 110, 111 and 112 in a one-to-one correspondence manner as the input of the support vector machine classifier;
114. training a support vector machine classifier, and performing class prediction by using a trained support vector machine algorithm;
115. training and classifying by a support vector machine algorithm to obtain the classification accuracy of the state data;
116. and training a full-life state model, and representing the residual life of each state by using a time-dwell function in the HSMM model.
As shown in fig. 2, the steps 100, 102, 103, and 104 specifically include the following processes:
200. obtaining the minimum noise coefficient, gain, third-order intermodulation point and 1dB compression point contained in a device (ATF53143) in the radio frequency analog circuit according to the datasheet of the determined variation curve of the maximum power, voltage, current and temperature which can be borne by the device in the radio frequency analog circuit and the critical performance of the transistor under different frequencies (2GHz, 900MHz) and different biases, and obtaining data which can represent the state characteristics of the device by taking the internal structure of the device (the tube is known through a library file of the transistor) as a model;
201. dc analysis determines the dc operating point of the device (transistor), and in combination with datasheet and bias circuitry, determines that the dc operating point of the device (ATF54143) is (3V, 28.3 mA).
202. Designing a direct-current working point biasing circuit;
203. and (3) analyzing stability, carrying out S parameter simulation, and enabling a Stabfact control in the ADS to represent the stability to be more than 1. And adding small inductors to two sources of the transistor to be used as negative feedback, and repeatedly adjusting feedback inductance values to ensure that the transistor is stable in the working frequency range.
ADS provides DA smithhartmatch matching, and after no-reflection matching a conjugate of the minimum noise figure impedance is obtained, Z ═ 24.9-j6.2) Ω.
205. The smith chart is used to obtain the output impedance that maximizes gain.
206. All matching is completed and a schematic diagram of the overall amplifier circuit can be obtained.
The K-Means clustering algorithm flow is shown in FIG. 3:
compared with the previous research, the characteristic data set collected by the low-noise amplifier along with the temperature change is improved, the noise coefficient and the stability coefficient are increased to comprehensively represent the performance characteristics of the low-noise amplifier circuit in consideration of the singleness of the S parameter, and the significance of the S parameter is as follows: the ratio of the incident wave to the reflected wave of the radio frequency circuit can be obtained, so that the problem of measuring the voltage and the current of the radio frequency circuit as characteristics is solved; for the radio frequency amplifying circuit, the gain is the most important parameter, and the variation of the gain can be observed remarkably through the S parameter. However, the noise coefficient NF is also a main technical index of low noise amplifier, which indicates that after a signal passes through an amplifier, the signal-to-noise ratio is deteriorated due to noise generated by the amplifier, and the multiple of the reduction of the signal-to-noise ratio is the noise coefficient, so that the performance of the amplifier can be represented to a certain extent. The stability coefficient can also represent the performance change of the low-noise amplifier circuit in the using process, the performance degradation factor of the low-noise amplifier circuit in the actual working process is represented by the change of simulation temperature, and the six parameters are found to be changed regularly along with the rise of the temperature.
The steps 105 and 106 further include:
300. inputting a data set D without labels for K-Means;
the K-Means algorithm divides original data into K data clusters by taking the distance between data as a standard, so that the intra-cluster similarity is high and the inter-cluster similarity is low;
302. and outputting cluster division results after the K-Means cluster analysis.
As shown in fig. 4, the steps 106 and 107 specifically include:
400. and determining a k value, namely the number of categories obtained by clustering the expected initial data, and selecting k data points as centroids in the data set according to experience or at random.
401. Carrying out initialization;
402. calculating the distances from the points except the clustering center in the data set to the k centroids, and dividing all sample points into clusters of each centroid according to the shortest clustering among the clusters;
403. forming k data clusters and updating cluster division;
404. recalculating centroids for the k sets;
405. and repeating the steps, and stopping the algorithm when the criterion function is converged or the cluster between the newly calculated centroid and the original centroid is in the set threshold range.
As shown in fig. 5, the data set is put into a K-Means algorithm model in MATLAB, and the clustering effect shown in fig. 5 can be obtained by running, and the mapping result shows that the regular data set is clustered into 4 clusters, and four clustering centers are obtained at the same time.
As shown in fig. 6, fig. 6 is an implementation schematic diagram of a support vector machine:
given a training sample value D { (x)1,y1),(x2,y2)...,(xm,ym)},yi∈ { -1, +1}, wherein yiAs a function of class, xiIt may be a vector constructed directly from some eigenvalues in the object sample, or it may be a mapping vector that maps the original vector to a higher dimensional space by some sum function.
Constructing a hyperplane in the sample space:
ωTx+b＝0
wherein ω ═ ω (ω ═ ω)1；ω2；...；ωd) The direction of the hyperplane is determined as the normal vector, and the distance between the hyperplane and the origin is determined as the displacementAnd (5) separating.
The hypothesis hyperplane (m, b) can correctly classify the training samples, i.e. for (x)i,yi) ∈ D if y i1, then ωTxi+ b > 0, if yiWhen 1 is not substituted, then ωTxi+ b < 0. Order:
ω can be described as:
defining a discriminant function:
f(x)＝ωTx+b
the classification function for the test set can be described as: label (x) sign (f (x)) sign (ω)Tx + b) the maximum classification boundary can be obtained by minimizing the formula, introducing a relaxation variableiAnd a penalty factor c, namely:
the satisfied constraint conditions are as follows:
yi(ωTxi+b)...1-i,i＝1,2,...,n
i...0,i＝1,2,...,n
solving the solution of the above function may be equivalent to maximizing the following function, introducing the Lagrangian multiplier αi：
And the constraint conditions are satisfied:
0 embroidering αic，i＝1，2，...，n
In the formula, k (x)iyj) Is a kernel function, and the classification function finally obtained is:
after the optimal hyperplane is obtained through training data, for a given unknown sample, the value of the indicating function is obtained only by substituting the formula, and the classification of the sample can be judged according to the value.
The selection of the kernel function is determined by a t parameter in the program;
for optimization of a punishment parameter c and a kernel function parameter g in the support vector machine model, a Grid Search method is adopted, and the effect of each pair of parameters is sequentially tested in a two-dimensional parameter matrix formed by c and g, so that the overall optimum effect can be obtained.
Preliminary results: accuracy 94.1176% (32/34) (classification), further improvements are needed. The flow when diagnosis and evaluation were performed is as in FIG. 7:
the application of the support vector machine model aims at solving the problem of multi-classification, and adopts a one-to-one multi-classification method, namely, the clustering result of K-Means is put into the method and randomly divided into a training set and a test set, and the optimization method comprises the following steps:
and obtaining a comparison result of the prediction classification accuracy of the final test set by adopting different normalization modes without normalization pretreatment:
adopting different kernel functions for comparison, and selecting the optimal kernel function; common kernel functions are linear, polymodal, radial basis functions, sigmoid, and the like,
as shown in fig. 6, the steps 106 and 114 specifically include:
600. and after the K-Means cluster analysis, taking the data with the label class information as the input of a support vector machine algorithm to train and test the classifier.
601. And (4) normalizing the data by using a mapminmax function in a MATLAB tool box.
602. In order to improve the classification accuracy, parameters c and g of the kernel function are optimized by adopting a grid search method.
603. And establishing a model by using a libsvmtrain function and a libsvmpredict function in the LIBSVM tool box for training and prediction.
604. And operating the program to obtain the classification accuracy.
605. And under the condition that the accuracy is not high enough, changing the optimization mode or certain parameters and optimizing again.
606. The steps realize the multi-classification effect of the support vector machine.
As shown in fig. 7, the steps 107, 108, 109, 110, 111, and 112 include:
700. and gradually extracting fault parameters representing the health state of the low-noise discharging circuit by using the temperature as a variation factor.
701. Determining HSMM model category, and determining initial parameters A, B, Π and D. Inputting the data after the K-Means clustering, and performing model training by using an EM algorithm.
702. The life-cycle data was trained by the EM algorithm to the HSMM model with all states.
703. After the model training, the re-estimated parameter values can be obtained, wherein the mean and the variance of the residence time distribution function of each state are included, and the residence time of each state can be further obtained.
704. The test set was input as an observation sequence to the trained HSMM model.
705. Through model training, the relationship between the iteration step number and the maximum likelihood function of each observation state can be obtained.
706. And putting the maximum likelihood function and the state labels into a trained support vector machine classifier in a one-to-one correspondence manner, and checking the classification accuracy.
707. And determining the confidence interval of the remaining service life of the current state according to the mean value and the variance of the state residence time.
As shown in fig. 8, fig. 8 is an algorithm implementation flow of parameter estimation in the HSMM model:
and performing clustering analysis on the extracted characteristic data in MATLAB by a K-Means multidimensional clustering algorithm to form four degradation states of the circuit, randomly selecting a training set from data of each group of states, and putting the training set into an HSMM model for simulation to obtain four trained state models. After the test set is preprocessed in the same way as the training set, the Viterbi algorithm is used for calculating the maximum likelihood value of the test set under the HSMM model of each state through 4 trained CHSMMs, wherein the maximum likelihood value represents the matching degree of a sample and the state model, so that about 20 groups of observed values of each state can obtain 4 probability values, then the sample and the maximum likelihood value are normalized together and then added into a fault label to be input into a trained SVM classifier for judgment, the correct rate of the test sample can be obtained, and then the combined result of the two methods is compared with the result of a single algorithm, as shown in the following table:
the comparison results show that: the SVM-HMM model has a better recognition effect on the test sample than a single model, is improved by 4.87% and 3.31% on average relative to the SVM and HMM single models, and has higher accuracy. Therefore, the SVM-HMM model has good application value and reference significance when being applied to the PHM of the radio frequency analog circuit.
The steps specifically include:
800. and (3) taking the collected circuit performance parameters as characteristic data, selecting a proper HSMM model, and determining initial parameters and iteration steps.
801. B and the time mainstream function are initialized.
802. And training iteration is carried out by using an EM algorithm.
803. An iteration variable is set, and iteration is started from 1.
804. The maximum likelihood probability is calculated using the Viterbi algorithm.
805. And judging whether a convergence condition is reached.
806. And continuing iteration if the convergence condition is not met.
807. The maximum number of iteration steps is reached.
808. And if the maximum iteration step number is exceeded, the model does not reach convergence, and the operation is exited.
809. And obtaining a training model when a convergence condition is reached.
As can be seen from fig. 9, the training of the support vector machine is still relatively ideal, and only individual sample points on the boundary are misjudged by the real class comparison. The SVM classifier trained by the method can be used in subsequent combination with the HSMM, and can effectively perform multi-classification on the maximum likelihood function trained by the HSMM model, so that the effect of state recognition is achieved.
The most important task of machine learning is to estimate and guess the unknown variables of interest based on some evidence that has been observed. The probabilistic model provides a descriptive framework that attributes the learning task to the probability distribution of the computational variables. Hidden Markov Model (HMM) is a dynamic Bayesian network with the simplest structure, is a famous directed graph Model, can be used for time series data modeling, and has wide application in the fields of speech recognition, natural language processing and the like. Therefore, attempts were made to use HMM models for fault prediction of the rf circuit studied this time.
The HMM model is developed on the basis of a Markov chain as a dynamic signal time series statistical model, consists of two random processes of an observation value sequence and a state value sequence, and is very suitable for processing continuous or discrete dynamic signals. According to a pattern matching principle, searching a pattern most similar to an unknown signal as an identification result, and emphasizing the similarity degree in an expression signal category; the SVM is suitable for processing classification problems, samples in different modes can be separated by a distance as large as possible according to limited sample information, the best popularization capability is obtained, and the difference among the classes is well expressed.
A probability distribution function representing the state dwell time of the hidden state is added to the HMM to represent the dwell time of each state. Therefore, the application field of the HSMM is wider, and the HSMM can be applied to the life prediction of each system. On the extraction of fault features: the attribute characteristics of the low-noise amplifier circuit are expanded, the single S parameter is expanded into the multi-dimensional data characteristics,
the increase of the noise coefficient and the stability coefficient can improve the accuracy of fault identification to a certain extent.
On the method of data processing: the effective combination of HMM and SVM improves the misjudgment risk of HMM on the classification problem, the HMM is suitable for processing dynamic continuous change and can well represent the similarity in the category, but when determining the maximum likelihood probability of the most similar model, the maximum likelihood probability of a plurality of HMM models is too close or equal, the SVM is suitable for processing the multi-classification problem and can reflect the difference between the categories to the maximum extent, the SVM belongs to a supervised classification method, the SVM needs to adopt a sample of artificial identification for training, the artificial identification of the sample is a more complicated process, the K-Means belongs to an unsupervised classification method, according to the respective advantages and disadvantages of the two, the radio frequency circuit state classification method based on the combination of SVM and K-Means is used for data clustering firstly, then each type of data with identification is randomly divided into a training set and a testing set, training an SVM classifier, and finally comparing with a real label, the method avoids the defect of poor classification effect of an unsupervised method, simultaneously omits the tedious process of manually identifying a sample in the SVM method, obtains a preliminary result through the experiment, proves the feasibility and the superiority of the method, and needs further research in order to enable the new method to be better applied to practice.
As shown in fig. 9, after the test set is similarly preprocessed, the Viterbi algorithm is used to calculate the maximum likelihood value of the test set under the HSMM model of each state through the trained 4 CHSMMs, where the maximum likelihood value represents the matching degree of the sample and the state model, so that about 20 groups of observed values of each state can obtain 4 probability values, and then the sample and the maximum likelihood value are normalized together and then added to the fault label to be input into the trained SVM classifier for judgment, so as to obtain the classification accuracy of 97.0588%.
Claims (8)
1. A radio frequency analog circuit fault diagnosis method based on SVM and HMM is characterized in that: at least comprises the following steps:
100. determining a radio frequency analog circuit;
101. determining an ATF54143 in the radio frequency analog circuit;
102. obtaining the fault prediction characteristics of devices in the radio frequency analog circuit by adopting ADS simulation software;
103. carrying out quantitative processing and storage on the determined technical parameters of the radio frequency analog circuit device to obtain the fault characteristics expressed by the change of the technical parameters;
104. the ADS is adopted to complete circuit design, temperature is used as an environment change factor in a simulation mode, multi-dimensional parameter attributes are extracted, the actual working state of the low-noise amplifier circuit is represented, and matching of minimum noise and maximum gain is included;
105. clustering the extracted multi-dimensional parameter attribute data, and representing the multi-dimensional parameter attribute as a cluster with a label category through a K-Means algorithm;
106. randomly selecting a part of data of each category from the results after the K-Means clustering as a test sample set;
107. selecting the remaining data as a training sample set after the test sample set is selected;
108. selecting a left and right HSMM type without jumping, determining initial parameters, and performing model training through an EM algorithm;
109. obtaining an HSMM model of the state I and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state I;
110. obtaining an HSMM model of the state II and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state II;
111. obtaining an HSMM model of the state III and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state III;
112. obtaining an HSMM model of the state four and a relation curve of the iteration step number and the maximum likelihood probability of the observation sequence of the state four;
113. taking the relation curves obtained in the steps 109, 110, 111 and 112 in a one-to-one correspondence manner as the input of the support vector machine classifier;
114. training a support vector machine classifier, and performing class prediction by using a trained support vector machine algorithm;
115. training and classifying by a support vector machine algorithm to obtain the classification accuracy of the state data;
116. and training a full-life state model, and representing the residual life of each state by using a time-dwell function in the HSMM model.
2. The method for diagnosing faults of a radio frequency analog circuit based on the SVM and HMM as claimed in claim 1, wherein: the steps 100, 102, 103, and 104 are specifically given by the following processes:
200. obtaining the minimum noise coefficient, gain, third-order intermodulation point and 1dB compression point contained in the device in the radio frequency analog circuit by taking the datasheet of the determined change curve of the device in the radio frequency analog circuit under the conditions of the maximum power, voltage, current and temperature which can be borne by the device and the key performance of the transistor under different frequencies and different biases as the basis, and obtaining data which can represent the state characteristics of the device by taking the internal structure of the device as a model;
201. the direct current analysis is carried out, the direct current operating point of the device is determined, and the direct current operating point of the device can be determined by combining the datasheet and the bias circuit; (3V, 28.3mA)
202. Designing a direct-current working point biasing circuit;
203. and (3) analyzing stability, carrying out S parameter simulation, and enabling a Stabfact control in the ADS to represent the stability to be more than 1. And adding small inductors to two sources of the transistor to be used as negative feedback, and repeatedly adjusting feedback inductance values to ensure that the transistor is stable in the working frequency range.
ADS provides DA smithhartmatch matching, and after no-reflection matching a conjugate of the minimum noise figure impedance is obtained, Z ═ 24.9-j6.2) Ω.
205. The smith chart is used to obtain the output impedance that maximizes gain.
206. All matching is completed and a schematic diagram of the overall amplifier circuit can be obtained.
3. The method for diagnosing faults of a radio frequency analog circuit based on the SVM and HMM as claimed in claim 1, wherein: the steps 105 and 106 further include:
300. inputting a data set D without labels for K-Means;
the K-Means algorithm divides original data into K data clusters by taking the distance between data as a standard, so that the intra-cluster similarity is high and the inter-cluster similarity is low;
302. and outputting cluster division results after the K-Means cluster analysis.
4. The method for diagnosing faults of a radio frequency analog circuit based on the SVM and HMM as claimed in claim 1, wherein: the steps 106 and 107 specifically include:
400. determining a k value, namely the number of categories obtained by clustering the expected initial data, and selecting k data points as centroids in the data set according to experience or at random;
401. carrying out initialization;
402. calculating the distances from the points except the clustering center in the data set to the k centroids, and dividing all sample points into clusters of each centroid according to the shortest clustering among the clusters;
403. forming k data clusters and updating cluster division;
404. recalculating centroids for the k sets;
405. and repeating the steps, and stopping the algorithm when the criterion function is converged or the cluster between the newly calculated centroid and the original centroid is in the set threshold range.
5. The method for diagnosing faults of a radio frequency analog circuit based on the SVM and HMM as claimed in claim 1, wherein: the steps 106 and 114 specifically include:
600. after K-Means cluster analysis, taking data with label category information as input of a support vector machine algorithm, and training and testing a classifier;
601. the normalization of data is realized by applying a mapminmax function in an MATLAB toolbox;
602. in order to improve the classification accuracy, parameters c and g of the kernel function are optimized by adopting a grid search method;
603. establishing a model by using a libsvmtrain function and a libsvmpredict function in an LIBSVM tool box, and training and predicting;
604. running a program to obtain the classification accuracy;
605. and under the condition that the accuracy is not high enough, changing the optimization mode or certain parameters and optimizing again.
6. The method for diagnosing faults of a radio frequency analog circuit based on the SVM and HMM as claimed in claim 1, wherein: the steps 107, 108, 109, 110, 111 and 112 comprise:
700. gradually extracting fault parameters representing the health state of the low-noise discharging circuit by using the temperature as a variation factor;
701. determining HSMM model category, and determining initial parameters A, B, Π and D. Inputting data after the K-Means clustering, and performing model training by using an EM algorithm;
702. training the full-life data through an EM algorithm to obtain an HSMM model with all states;
703. after the model is trained, a reestimated parameter value can be obtained, wherein the reestimated parameter value comprises the mean value and the variance of the residence time distribution function of each state, and the residence time of each state can be further obtained;
704. inputting the test set as an observation sequence into the trained HSMM model;
705. through model training, the relationship between the iteration step number and the maximum likelihood function of each observation state can be obtained;
706. putting the maximum likelihood function and the state labels into a trained support vector machine classifier in a one-to-one correspondence manner, and checking the classification accuracy;
707. and determining the confidence interval of the remaining service life of the current state according to the mean value and the variance of the state residence time.
7. The method for diagnosing faults of a radio frequency analog circuit based on an SVM and an HMM as claimed in claim 1, wherein said step 108 specifically comprises:
800. taking the collected circuit performance parameters as characteristic data, selecting a proper HSMM model, and determining initial parameters and iteration steps;
801. initializing the B and the time residence function;
802. training iteration is carried out by using an EM algorithm;
803. setting an iteration variable, and starting iteration from 1;
804. calculating the maximum likelihood probability by using a Viterbi algorithm;
805. judging whether a convergence condition is reached;
806. continuing iteration if the convergence condition is not met;
807. the maximum number of iteration steps is reached;
808. if the maximum iteration step number is exceeded, the model does not reach convergence, and quitting is performed;
809. and obtaining a training model when a convergence condition is reached.
8. The method for diagnosing faults of a radio frequency analog circuit based on an SVM and an HMM as claimed in claim 1, wherein said step 114 comprises:
given a training sample value D { (x)1,y1),(x2,y2)...,(xm,ym)},yi∈ { -1, +1}, wherein yiAs a function of class, xiPossibly a vector constructed directly from some eigenvalues in the object sample;
constructing a hyperplane in the sample space:
ωTx+b＝0
wherein ω ═ ω (ω ═ ω)1；ω2；...；ωd) Determining the direction of the hyperplane as a normal vector, and determining the distance between the hyperplane and the origin as a displacement;
the hypothesis hyperplane (m, b) can correctly classify the training samples, i.e. for (x)i,yi) ∈ D if yi1, then ωTxi+ b > 0, if yiWhen 1 is not substituted, then ωTxi+ b is less than 0; order:
ω can be described as:
defining a discriminant function:
f(x)＝ωTx+b
the classification function for the test set can be described as: label (x) sign (f (x)) sign (ω)Tx + b) the maximum classification boundary can be obtained by minimizing the formula, introducing a relaxation variableiAnd a penalty factor c, namely:
the satisfied constraint conditions are as follows:
yi(ωTxi+b)...1-i,i＝1,2,...,n
i...0,i＝1,2,...,n
solving the solution of the above function may be equivalent to maximizing the following function, introducing the Lagrangian multiplier αi：
And the constraint conditions are satisfied:
0 embroidering αic，i＝1，2，...，n
In the formula, k (x)iyj) Is a kernel function, and the classification function finally obtained is:
after the optimal hyperplane is obtained through training data, for a given unknown sample, the value of the indicating function is obtained only by substituting the formula, and the classification of the sample can be judged according to the value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010528645.8A CN111751714A (en) | 2020-06-11 | 2020-06-11 | Radio frequency analog circuit fault diagnosis method based on SVM and HMM |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010528645.8A CN111751714A (en) | 2020-06-11 | 2020-06-11 | Radio frequency analog circuit fault diagnosis method based on SVM and HMM |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111751714A true CN111751714A (en) | 2020-10-09 |
Family
ID=72676029
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010528645.8A Pending CN111751714A (en) | 2020-06-11 | 2020-06-11 | Radio frequency analog circuit fault diagnosis method based on SVM and HMM |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111751714A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112381051A (en) * | 2020-11-30 | 2021-02-19 | 闽江师范高等专科学校 | Plant leaf classification method and system based on improved support vector machine kernel function |
| CN112731098A (en) * | 2020-12-03 | 2021-04-30 | 西安电子科技大学 | Radio frequency low-noise discharge circuit fault diagnosis method, system, medium and application |
| CN113188570A (en) * | 2021-04-27 | 2021-07-30 | 西南石油大学 | Attitude error calibration method of inclinometer while drilling based on support vector classifier and K-proximity method |
| CN113486720A (en) * | 2021-06-08 | 2021-10-08 | 浙江大学 | Video playing content inference method based on high-frequency noise of intelligent terminal device switching power supply |
| CN114858982A (en) * | 2022-07-11 | 2022-08-05 | 深圳市文浩科技有限公司 | Directional clamping winding type building material measuring system and method |
| CN116263878A (en) * | 2022-10-09 | 2023-06-16 | 北京理工大学 | Lithium ion battery thermal runaway risk prediction method and device |
| CN116340834A (en) * | 2023-05-26 | 2023-06-27 | 成都凯天电子股份有限公司 | Data correction system and method for superposition of measured signals and direct current bias signals |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102661866A (en) * | 2012-05-11 | 2012-09-12 | 天津工业大学 | Engine fault identification method based on time-domain energy and support vector machine |
| CN103198052A (en) * | 2013-04-09 | 2013-07-10 | 山东师范大学 | Active learning method based on support vector machine |
| CN104502103A (en) * | 2014-12-07 | 2015-04-08 | 北京工业大学 | Bearing fault diagnosis method based on fuzzy support vector machine |
| CN106371610A (en) * | 2016-09-23 | 2017-02-01 | 重庆金瓯科技发展有限责任公司 | Method for detecting driving fatigue based on electroencephalogram |
| CN108445855A (en) * | 2018-04-27 | 2018-08-24 | 惠州市宝捷信科技有限公司 | Injection molding machine formulation parameter optimization method based on K-means |
| CN109308316A (en) * | 2018-07-25 | 2019-02-05 | 华南理工大学 | A kind of adaptive dialog generation system based on Subject Clustering |
| CN110232416A (en) * | 2019-06-13 | 2019-09-13 | 中国人民解放军空军工程大学 | Equipment failure prediction technique based on HSMM-SVM |
-
2020
- 2020-06-11 CN CN202010528645.8A patent/CN111751714A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102661866A (en) * | 2012-05-11 | 2012-09-12 | 天津工业大学 | Engine fault identification method based on time-domain energy and support vector machine |
| CN103198052A (en) * | 2013-04-09 | 2013-07-10 | 山东师范大学 | Active learning method based on support vector machine |
| CN104502103A (en) * | 2014-12-07 | 2015-04-08 | 北京工业大学 | Bearing fault diagnosis method based on fuzzy support vector machine |
| CN106371610A (en) * | 2016-09-23 | 2017-02-01 | 重庆金瓯科技发展有限责任公司 | Method for detecting driving fatigue based on electroencephalogram |
| CN108445855A (en) * | 2018-04-27 | 2018-08-24 | 惠州市宝捷信科技有限公司 | Injection molding machine formulation parameter optimization method based on K-means |
| CN109308316A (en) * | 2018-07-25 | 2019-02-05 | 华南理工大学 | A kind of adaptive dialog generation system based on Subject Clustering |
| CN110232416A (en) * | 2019-06-13 | 2019-09-13 | 中国人民解放军空军工程大学 | Equipment failure prediction technique based on HSMM-SVM |
Non-Patent Citations (2)
| Title |
|---|
| 刘志刚: "射频放大模块故障预测方法的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 郑志文等: "基于隐马尔科夫模型和支持向量机的模拟电路早期故障诊断", 《计算机测量与控制》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112381051A (en) * | 2020-11-30 | 2021-02-19 | 闽江师范高等专科学校 | Plant leaf classification method and system based on improved support vector machine kernel function |
| CN112381051B (en) * | 2020-11-30 | 2023-12-19 | 闽江师范高等专科学校 | Plant leaf classification method and system based on improved support vector machine kernel function |
| CN112731098A (en) * | 2020-12-03 | 2021-04-30 | 西安电子科技大学 | Radio frequency low-noise discharge circuit fault diagnosis method, system, medium and application |
| CN112731098B (en) * | 2020-12-03 | 2022-04-29 | 西安电子科技大学 | Radio frequency low-noise discharge circuit fault diagnosis method, system, medium and application |
| CN113188570A (en) * | 2021-04-27 | 2021-07-30 | 西南石油大学 | Attitude error calibration method of inclinometer while drilling based on support vector classifier and K-proximity method |
| CN113486720A (en) * | 2021-06-08 | 2021-10-08 | 浙江大学 | Video playing content inference method based on high-frequency noise of intelligent terminal device switching power supply |
| CN113486720B (en) * | 2021-06-08 | 2023-12-08 | 浙江大学 | Video playing content deducing method based on high-frequency noise of switching power supply of intelligent terminal equipment |
| CN114858982A (en) * | 2022-07-11 | 2022-08-05 | 深圳市文浩科技有限公司 | Directional clamping winding type building material measuring system and method |
| CN114858982B (en) * | 2022-07-11 | 2022-11-01 | 深圳市文浩科技有限公司 | Directional clamping winding type building material measuring system |
| CN116263878A (en) * | 2022-10-09 | 2023-06-16 | 北京理工大学 | Lithium ion battery thermal runaway risk prediction method and device |
| CN116340834A (en) * | 2023-05-26 | 2023-06-27 | 成都凯天电子股份有限公司 | Data correction system and method for superposition of measured signals and direct current bias signals |
| CN116340834B (en) * | 2023-05-26 | 2023-10-03 | 成都凯天电子股份有限公司 | Data correction system and method for superposition of measured signals and direct current bias signals |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111751714A (en) | Radio frequency analog circuit fault diagnosis method based on SVM and HMM | |
| CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
| CN110738247B (en) | Fine-grained image classification method based on selective sparse sampling | |
| CN114564982B (en) | Automatic identification method for radar signal modulation type | |
| CN113255573B (en) | Pedestrian re-identification method based on mixed cluster center label learning and storage medium | |
| CN110794360A (en) | Method and system for predicting fault of intelligent electric energy meter based on machine learning | |
| CN112821559B (en) | Non-invasive household appliance load depth re-identification method | |
| CN112560596B (en) | Radar interference category identification method and system | |
| CN111428768A (en) | Hellinger distance-Gaussian mixture model-based clustering method | |
| CN111553186A (en) | Electromagnetic signal identification method based on depth long-time and short-time memory network | |
| CN113541834B (en) | Abnormal signal semi-supervised classification method and system and data processing terminal | |
| CN114897085A (en) | Clustering method based on closed subgraph link prediction and computer equipment | |
| CN112785015B (en) | Equipment fault diagnosis method based on case reasoning | |
| CN112801127A (en) | Cable partial discharge defect identification method based on oscillation waves | |
| CN117131449A (en) | Data management-oriented anomaly identification method and system with propagation learning capability | |
| CN113523904A (en) | Cutter wear detection method | |
| CN116842460A (en) | Cough-related disease identification method and system based on attention mechanism and residual neural network | |
| CN116400168A (en) | Power grid fault diagnosis method and system based on depth feature clustering | |
| CN112163613A (en) | Rapid identification method for power quality disturbance | |
| Yang et al. | Deep ensemble siamese network for incremental signal classification | |
| Liu et al. | An accurate method of determining attribute weights in distance-based classification algorithms | |
| CN113572732B (en) | Multi-step attack modeling and prediction method based on VAE and aggregated HMM | |
| Ambarsari et al. | Comparison Approaches of the Fuzzy C-Means and Gaussian Mixture Model in Clustering the Welfare of the Indonesian People | |
| CN116484206A (en) | SEIM-based unknown radiation source individual identification method and system | |
| CN115344693A (en) | Clustering method based on fusion of traditional algorithm and neural network algorithm |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20201009 |
|
| RJ01 | Rejection of invention patent application after publication |