A Phased Array Radar Transceiver Fault Identification Method Based on Double-Layer Classifier
The double-layer classifier with WSVM and SMDBN, optimized by MEO, addresses the limitations of single models in phased array radar transceiver fault identification, achieving high accuracy in fault diagnosis by distinguishing complex fault types.
Patent Information
- Application Number
- CN202411110137.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-14
AI Technical Summary
Existing single models for phased array radar transceiver fault identification are limited in their ability to handle a wide range of data and problems, leading to inaccurate fault identification and diagnosis.
A phased array radar transceiver fault identification method using a double-layer classifier composed of a weighted support vector machine (WSVM) as a coarse classifier and a sparse momentum deep belief network (SMDBN) as a fine classifier, with parameter configuration optimized by an improved balanced optimizer MEO that incorporates Tent chaotic mapping and Gaussian mutation.
The method enhances fault identification accuracy by effectively distinguishing easily confused fault types and improving generalization ability, achieving an average recognition rate of 96% in phased array radar transceiver fault diagnosis.
Smart Images

Figure CN118625275B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of radar transceiver fault identification, and in particular to a phased array radar transceiver fault identification method based on a double-layer classifier. Background Art
[0002] Phased array radar is a radar array composed of a large number of identical radiating units. Each radiating unit is independently controlled in phase and amplitude by a wave controller and a phase shifter, which can obtain an accurate and predictable radiation pattern and beam direction. Phased array radars are generally located in remote border areas and need to operate 24 hours a day. Component failure is inevitable in continuous high-load working mode and harsh environment. Unexpected failures outside the maintenance plan will at least affect the validity and detection performance of radar data, and even cause mission interruption. The transceiver is one of the components with the highest failure frequency in phased array radars. Due to the large number of fault sources, it is a challenge to establish a reliable expert knowledge base or an accurate model. In addition, the development of digitalization and signal processing technology of radar systems has provided favorable conditions for the research of data-based radar fault diagnosis methods.
[0003] Some research results have been achieved in data-driven fault diagnosis. Typical data-driven methods include machine learning and deep learning. In terms of the application of machine learning methods, the prior art studies the fault diagnosis of vehicle automatic steering actuators, uses the system model to generate the residual signal of the support vector machine (SVM) training data, and proposes an undersampling method to improve the diagnostic performance; some scholars have also proposed a hybrid model of fault feature distance stack autoencoder and support vector machine, in which the autoencoder is used to extract bearing features and the support vector machine is used to classify normal data and fault data; some scholars have proposed using discrete wavelet transform to extract early fault diagnosis features of gearboxes, and then constructing artificial neural networks for fault classification. Since the learning process of machine learning methods is relatively shallow, it is usually necessary to combine feature engineering technologies such as Fourier transform, wavelet analysis, and empirical mode decomposition. Some scholars have proposed using long short-term memory networks to adaptively learn the feature information of the original multivariate time series signal to achieve wind turbine fault diagnosis. In order to learn more comprehensive wheelset bearing fault information under variable load and strong noise conditions, some scholars have proposed to construct a deep learning method called multi-branch multi-scale convolutional neural network; in addition to long short-term memory and convolutional neural networks, deep belief network DBN is also a deep learning model with the advantages of multi-source data fusion, automatic feature construction, and fast convergence speed.
[0004] The above research mainly focuses on the improvement of a single model. However, a single model can only handle specific types of data and problems, but cannot cover a wide range of situations of all types of data and problems. This may cause the model to be unable to solve some emerging problems and adapt to new data distributions. Therefore, how to overcome the shortcomings of a single model and ultimately improve the accuracy of fault identification and diagnosis is a key issue that needs to be solved urgently. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides a phased array radar transceiver fault identification method based on a double-layer classifier, which solves the technical problem that a single model can only process specific types of data and problems but cannot cover a wide range of all types of data and problems, thereby improving the accuracy of fault identification and diagnosis.
[0006] To solve the above technical problems, the present invention provides the following technical solution: a phased array radar transceiver fault identification method based on a double-layer classifier, the method comprising the following steps:
[0007] S1. Obtaining the operating data of the phased array radar transceiver;
[0008] S2. constructing a double-layer classifier based on the operating data of the phased array radar transceiver, wherein the double-layer classifier is composed of a coarse classifier and a fine classifier;
[0009] S3, configuring parameters of the two-layer classifier by improving the balanced optimizer MEO;
[0010] S4, performing fault identification and classification on easily distinguishable monitoring variables in the operation data of the phased array radar transceiver based on a coarse classifier;
[0011] S5. For the faults that cannot be directly identified in step S4, by calculating the Fisher discriminant ratio of each monitoring variable in the operation data, the key monitoring variable data is selected and identified based on the detailed classifier to obtain the accurate fault type.
[0012] Furthermore, in step S1, obtaining the operation data of the phased array radar transceiver includes: the phased array radar antenna receives the electromagnetic signal reflected from the target, and performs data fusion and output processing on the received signal through analog-to-digital conversion, filter matching, and pulse compression operations.
[0013] Furthermore, in step S2, a two-layer classifier is constructed, which specifically includes:
[0014] S21. Set a fault set ,set up Type Failure and The fault confusion degree of type fault is , fault confusion express Type fault is misjudged as The probability of type failure and Type fault is misjudged as The average value of the probability of type failure, that is:
[0015] ;
[0016] in, For the test sample, is the corresponding recognition result in the test sample, and the fault confusion refers to the similarity between different types of faults;
[0017] S22. The coarse classifier is a weighted support vector machine (WSVM), and the fine classifier is a sparse momentum deep belief network (SMDBN).
[0018] Furthermore, in step S3, the improved equilibrium optimizer MEO adopts Tent chaotic mapping and Gaussian mutation to improve the performance of the optimizer. In the early stage, Tent chaotic mapping is used to increase the population diversity and accelerate the convergence speed of the optimizer. In the later stage, Gaussian mutation is used to avoid the optimizer from converging too early.
[0019] The parameter configuration of the two-layer classifier based on the improved balanced optimizer MEO is as follows:
[0020] S31, initialize the improved balance optimizer MEO parameters, each individual position is composed of model configuration parameters, the feasible domain space Individuals are initialized through Tent chaotic mapping to form an initial population, namely:
[0021] ;
[0022] in, express Individuals in Position in a dimension; Represents the dimension of the optimization problem; and They are The upper and lower bounds of the dimension variable in the search space; is the Tent chaotic sequence, which is given by the following formula:
[0023] ;
[0024] in, and A random number between 0 and 1;
[0025] S32, establish a balanced state pool and select candidates;
[0026] Balance Pool The definition is as follows:
[0027] ;
[0028] in, represents the four particles with the highest concentrations selected after the initialization phase of the improved equilibrium optimizer MEO is completed; represents the average particle and is given by:
[0029] ;
[0030] In the iterative process of the improved equilibrium optimizer MEO, each particle is selected with the same probability from the five candidate particles in the balanced state pool to generate the global optimal solution;
[0031] S33, improve the attention of the improved balanced optimizer MEO in the process of parameter configuration of the two-layer classifier;
[0032] S34, executing Gaussian mutation update;
[0033] First, mutant individuals are selected from the current population, Gaussian distribution is applied to each selected individual to generate new parameters and ensure that the parameter value after mutation is within the predefined range;
[0034] Secondly, calculate the fitness of the new individual after mutation, and decide whether to accept the new individual after mutation according to the fitness value;
[0035] Then, the new individuals that have undergone mutations are cross-operated with individuals in the original population to generate new candidate solutions;
[0036] Finally, after a new round of selection and update, the population is updated, and the improved equilibrium optimizer MEO enters the next generation of iterations. After multiple iterations, it gradually approaches the global optimal solution;
[0037] S35, judging whether the end condition is met, if the current iteration exceeds the maximum iteration, outputting the optimal individual position in the population; otherwise, executing step S32;
[0038] S36. Obtain parameter configuration according to the optimal individual position in the population, and retrain the coarse classifier or the fine classifier.
[0039] Further, in step S33, the attention of the improved balanced optimizer MEO in the process of parameter configuration of the double-layer classifier is improved, specifically including:
[0040] S331, Index Item It is an important indicator to measure the exploration and development capability of the improved balanced optimizer MEO, and its expression is:
[0041] ;
[0042] in, is a constant that controls the optimizer's exploration capability, is the symbolic function, and is a random number between 0 and 1. is a coefficient that is updated with the number of iterations and is given by:
[0043] ;
[0044] in, and Respectively represent the current number of iterations and the maximum number of iterations; is a constant that controls the optimizer's exploit capabilities;
[0045] S332, introduction of mass generation rate ,Right now:
[0046] ;
[0047] ;
[0048] in, is the control parameter vector of the generation rate; is the average particle concentration; is the current particle concentration; and A random number between 0 and 1; is a constant whose value is 0.5;
[0049] Each particle is updated as follows:
[0050] ;
[0051] in, is the unit volume.
[0052] Furthermore, step S4 specifically includes the following steps:
[0053] S41, update the support vector machine SVM model, the SVM model is:
[0054] ;
[0055] in, is the weight, It is The recognition results corresponding to the test samples are is the hyperplane normal vector, is the transpose of the hyperplane normal vector, is the penalty factor, is the slack variable, is a nonlinear mapping function. is the deviation, is the number of samples;
[0056] S42, map the sample set from the input space to the high-dimensional feature space through kernel mapping, construct the Lagrangian function and decision function, whose expressions are as follows:
[0057] ;
[0058] ;
[0059] in, is the constructed Lagrangian function; is the slack variable; is the Lagrange multiplier; and For the Lagrange multiplier of samples; Kernel function to realize nonlinear transformation; represents a symbolic function;
[0060] S43, determine the radial basis function, whose expression is:
[0061] ;
[0062] in, Expressed as a natural constant The exponential function with base , is the hyperparameter of the radial basis function, For vector and The distance between
[0063] S44, weighting the support vector machine SVM to obtain a decision function of the weighted support vector machine WSVM;
[0064] Let the total number of positive and negative samples be and Indicates that the centers of positive and negative samples are defined as and , positive sample radius And the radius of negative class samples It can be regarded as the maximum distance between positive and negative samples and the corresponding class center;
[0065] Therefore, the weighting coefficient can be:
[0066] ;
[0067] in, , , ; It should be noted that the introduction here This is to avoid the denominator in the formula being zero;
[0068] With the weighting factor The distribution of , the SVM model becomes:
[0069] ;
[0070] To solve the weighted coefficient, the Lagrangian function is constructed as follows:
[0071] ;
[0072] Calculation parameters, The partial derivative of is:
[0073] ;
[0074] in, is the partial derivative; is the above weighting coefficient;
[0075] Finally, the decision function of the weighted support vector machine WSVM is:
[0076] .
[0077] Furthermore, it is characterized in that step S5 specifically includes:
[0078] S51, using Fisher discriminant ratio to select key monitoring variable data;
[0079] variable The Fisher discriminant ratio is:
[0080] ;
[0081] in, represents the number of categories, Respectively represent variables The mean and variance of
[0082] S52, construct sparse momentum deep belief network SMDBN;
[0083] The Sparse Momentum Deep Belief Network (SMDBN) is a deep learning model consisting of three Boltzmann machines (RBMs) and a softmax layer.
[0084] The parameter update process of SMDBN is:
[0085] ;
[0086] in, is the connection weight between the visible layer and the hidden layer, and are the biases of the visible layer and the hidden layer respectively, is the number of iterations, is the momentum factor and , is the learning rate, is the mathematical expectation of the training data, Indicates The mathematical expectation of the sample distribution after step Gibbs sampling, For the neurons in the visible layer, For the hidden layer neurons, is the regularization factor, and Given by:
[0087] ;
[0088] in, is the number of neurons in the hidden layer, For the hidden layer activation probability; is the variance factor, For the neurons in the visible layer;
[0089] S53. Input the key monitoring variable data selected based on the Fisher discriminant ratio into the constructed sparse momentum deep belief network SMDBN to identify the accurate fault.
[0090] By means of the above technical solution, the present invention provides a phased array radar transceiver fault identification method based on a double-layer classifier, which has at least the following beneficial effects:
[0091] 1. The present invention constructs a double-layer classifier of a coarse classifier and a fine classifier, which is conducive to identifying and distinguishing easily confused fault types. The balanced optimizer MEO improved by Tent chaotic mapping and Gaussian mutation can determine the reasonable configuration parameters of the double-layer classifier.
[0092] 2. In the double-layer classifier of the present invention, the weighted support vector machine WSVM is used as a coarse classifier to identify and separate easily distinguishable faults, and the classification performance is improved by improving the balance optimizer MEO; the weighted support vector machine WSVM can reduce the impact of outliers on the total training error.
[0093] 3. In the two-layer classifier of the present invention, a sparse momentum deep belief network SMDBN is constructed as a fine classifier to identify fault types that cannot be directly identified by the weighted support vector machine WSVM as a coarse classifier, and its classification performance is improved by improving the balance optimizer MEO; the sparse momentum deep belief network SMDBN enhances the generalization ability of the model by adding sparse terms and momentum terms in parameter updates, thereby improving the accuracy of fault identification and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0094] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0095] Figure 1 A schematic diagram of an application object of the present invention is shown;
[0096] Figure 2 The overall framework flow chart of the present invention is shown;
[0097] Figure 3 The change of fitness function value of different optimizers of the present invention during the optimization process is shown;
[0098] Figure 4 The confusion matrix of the recognition rate (%) of the validation set using the MEO-WSVM model of the present invention is shown;
[0099] Figure 5 The double-layer classifier structure of the radar transceiver of the present invention is shown;
[0100] Figure 6 The Fisher discriminant ratios of the 26 monitoring variables of the present invention are shown;
[0101] Figure 7 The confusion matrix (%) of the recognition rate of the test set using the double-layer classifier of the present invention is shown;
[0102] Figure 8 The recognition rates of different classifiers of the present invention in ablation experiments are shown. DETAILED DESCRIPTION
[0103] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.
[0104] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, so the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0105] Please refer to Figure 1-Figure 8 This embodiment proposes a phased array radar transceiver fault identification method based on a double-layer classifier. By constructing a double-layer classifier of a coarse classifier and a fine classifier, the problem that a single model can only process specific types of data and problems, but cannot cover all types of data and problems, is solved. In order to improve the accuracy of fault identification, a balanced optimizer MEO improved by Tent chaotic mapping and Gaussian mutation is designed to determine the reasonable configuration parameters of the double-layer classifier and improve its classification performance. In the double-layer classifier of the present invention, a sparse momentum deep belief network SMDBN is constructed as a fine classifier to identify the fault type that the weighted support vector machine WSVM as a coarse classifier cannot directly identify. The sparse momentum deep belief network SMDBN enhances the model generalization ability by adding sparse terms and momentum terms in parameter updates, thereby improving the accuracy of fault identification and diagnosis. Figure 2 As shown, the method comprises the following steps:
[0106] S1. Obtaining operation data of a phased array radar transceiver.
[0107] As a preferred embodiment of step S1, obtaining the operating data of the phased array radar transceiver includes: the phased array radar antenna receives the electromagnetic signal reflected from the target, and performs data fusion and output processing on the received signal through analog-to-digital conversion, filter matching, and pulse compression operations.
[0108] S2. Construct a double-layer classifier based on the operating data of the phased array radar transceiver, wherein the double-layer classifier is composed of a coarse classifier and a fine classifier.
[0109] As a preferred embodiment of step S2, a two-layer classifier is constructed based on the operation data of the phased array radar transceiver, specifically including:
[0110] S21. Set a fault set ,set up Type Failure and The fault confusion degree of type fault is , fault confusion express Type fault is misjudged as The probability of type failure and Type fault is misjudged as The average value of the probability of type failure, that is:
[0111] ;
[0112] in, For the test sample, is the corresponding recognition result in the test sample, and the fault confusion refers to the similarity between different types of faults;
[0113] S22. The coarse classifier is a weighted support vector machine (WSVM), and the fine classifier is a sparse momentum deep belief network (SMDBN).
[0114] In this embodiment, by constructing a two-layer classifier of a coarse classifier and a fine classifier, the problem that a single model can only process specific types of data and problems is solved, which is conducive to identifying and distinguishing easily confused fault types.
[0115] S3. The parameters of the double-layer classifier are configured by improving the balanced optimizer MEO. The improved balanced optimizer MEO adopts Tent chaotic mapping and Gaussian mutation to improve the performance of the optimizer, thereby improving the classification performance of the coarse classifier and the fine classifier; in the early stage, Tent chaotic mapping is used to increase the population diversity and speed up the convergence of the optimizer, and in the later stage, Gaussian mutation is used to avoid the optimizer from converging too early;
[0116] As a preferred embodiment of step S3, the parameters of the double-layer classifier are configured based on the improved balanced optimizer MEO as follows:
[0117] S31, initialize the improved balance optimizer MEO parameters, each individual position is composed of model configuration parameters, the feasible domain space Individuals are initialized through Tent chaotic mapping to form an initial population, namely:
[0118] ;
[0119] in, express Individuals in Position in a dimension; Represents the dimension of the optimization problem; and They are The upper and lower bounds of the dimension variable in the search space; is the Tent chaotic sequence, which is given by the following formula:
[0120] ;
[0121] in, and A random number between 0 and 1;
[0122] During initialization, the individual positions of the coarse classifier are determined by the penalty factor and kernel parameters Therefore, the spatial dimension of MEO is is 2; for a single individual position of a detailed classifier, it is determined by the number of hidden layer nodes. , learning rate , Momentum Factor , variance factor and the regularization factor Composition, therefore, the spatial dimension of MEO is 7. By calculating the fitness value (fault recognition error rate) of each individual, the optimal individual position is recorded .
[0123] S32. Establish a balanced state pool and select candidates. The balanced state pool is the balance pool, which is the candidate solution provided in the optimization process. It consists of four particles with relatively good fitness values generated in the current initialization phase and an average particle. Candidates are a set of new possible solutions generated in the optimization process. These solutions will be evaluated and may be added to the balanced state pool. The balance pool is usually a group of pools used to temporarily save the balance state during the balancing process to alleviate short-term imbalance problems. It may share some functions with the balanced state pool, but it mainly focuses on immediate balance. The balanced state pool focuses more on long-term state preservation during the optimization process, and records the better state found during the entire optimization process.
[0124] Balance Pool The definition is as follows:
[0125] ;
[0126] in, represents the four particles with the highest concentrations selected after the initialization phase of the improved equilibrium optimizer MEO is completed; represents the average particle and is given by:
[0127] ;
[0128] In the iterative process of the improved equilibrium optimizer MEO, each particle is selected with the same probability from the five candidate particles in the balanced state pool to generate the global optimal solution;
[0129] S33. Improve the attention of the improved balanced optimizer MEO in the process of parameter configuration of the double-layer classifier; In the process of the optimizer configuring the classifier, improving attention means paying more attention to the exploration and optimization of specific candidate solutions or specific parts of the parameter space during the optimization process. Among them, the index term It is an important indicator to measure the exploration and development capability of the optimizer, and its expression is:
[0130] ;
[0131] in, is a constant that controls the optimizer's exploration capability, is the symbolic function, and is a random number between 0 and 1. is a coefficient that is updated with the number of iterations and is given by:
[0132] ;
[0133] in, and Respectively represent the current number of iterations and the maximum number of iterations; is a constant that controls the optimizer's development capability; in order to improve the optimizer's development capability, the quality generation rate is introduced ,Right now:
[0134] ;
[0135] ;
[0136] in, is the control parameter vector of the generation rate; is the average particle concentration; is the current particle concentration; and A random number between 0 and 1; is a constant whose value is 0.5; then, each particle is updated as follows:
[0137] ;
[0138] in, is the unit volume.
[0139] S34, executing Gaussian mutation update;
[0140] First, select mutant individuals from the current population, apply Gaussian distribution to each selected individual to generate new parameters and ensure that the parameter value after mutation is within the predefined range; secondly, calculate the fitness of the new individual after mutation, and decide whether to accept the new individual after mutation based on the fitness value. A Gaussian distribution with a mean of 0 and a variance of 1 is used to generate a new individual dimension:
[0141] ;
[0142] in, represents new particles that change within the same dimension, is the particle with the lowest concentration in the current dimension. In order to determine whether the concentration of the mutated particle is better than the original optimal particle, a greedy strategy is adopted after the particle mutation. That is, when the concentration of the mutated particle is better than the current particle, the concentration is updated; otherwise, the original particle is retained.
[0143] ;
[0144] Then, the new individuals that have undergone mutations are cross-operated with individuals in the original population to generate new candidate solutions; subsequently, the new population is evaluated and the best individual is selected again. This stage combines global search and local search to further improve the overall quality of the population. Finally, after a new round of selection and update, the balanced state pool (population) is updated, and the improved equilibrium optimizer MEO enters the next generation of iterations. Through multiple iterations, the optimizer gradually approaches the global optimal solution.
[0145] S35, judging whether the end condition is met, if the current iteration exceeds the maximum iteration, outputting the optimal individual position in the population; otherwise, executing step S32;
[0146] S36. Obtain parameter configuration according to the optimal individual position in the population, and retrain the coarse classifier or the fine classifier.
[0147] In this embodiment, in order to improve the accuracy of fault identification, a balanced optimizer MEO improved by Tent chaotic mapping and Gaussian mutation is used to determine the reasonable configuration parameters of the double-layer classifier and improve its classification performance. In the implementation process of this embodiment, Gaussian mutation update is an important step, and its role is to explore local areas in the solution space by slightly adjusting parameters, thereby improving the optimization effect.
[0148] S4, performing fault identification and classification on easily distinguishable monitoring variables in the operation data of the phased array radar transceiver based on a coarse classifier;
[0149] As a preferred embodiment of step S4, the following steps are specifically included:
[0150] S41. Update the support vector machine (SVM) model. SVM is a shallow machine learning model that has the advantage of processing nonlinear and high-dimensional data. Its basic principle is to transform data samples from low-dimensional space to high-dimensional space through kernel mapping, and then use linear partitioning to establish classification boundaries. Radar transceiver fault identification is a nonlinear multi-classification problem. The SVM model is established as follows:
[0151] ;
[0152] in, is the weight, It is The recognition results corresponding to the test samples are is the hyperplane normal vector, is the transpose of the hyperplane normal vector, is the penalty factor, is the slack variable, is a nonlinear mapping function. is the deviation, is the number of samples;
[0153] S42. Map the sample set from the input space to the high-dimensional feature space through kernel mapping, and construct the Lagrangian function and decision function, whose expressions are as follows:
[0154] ;
[0155] ;
[0156] in, is the constructed Lagrangian function; is the slack variable; is the Lagrange multiplier; and For the Lagrange multiplier of samples; Kernel function to realize nonlinear transformation; represents a symbolic function;
[0157] S43, determine the radial basis function, the radial basis function has good generalization performance, only one parameter needs to be determined. Therefore, the radial basis function selected by the present invention is expressed as:
[0158] ;
[0159] in, Expressed as a natural constant The exponential function with base , is the hyperparameter of the radial basis function, For vector and The distance between
[0160] S44, weighting the support vector machine SVM to obtain a decision function of the weighted support vector machine WSVM. In the above support vector machine SVM, sample points are considered equally. When there are outliers in the data set, the classification performance may be affected. In order to reduce the impact of outliers on the total training error, a lower weight can be given to the outliers when constructing the SVM.
[0161] Let the total number of positive and negative samples be and Indicates that the centers of positive and negative samples are defined as and , positive sample radius And the radius of negative class samples It can be regarded as the maximum distance between positive and negative samples and the corresponding class center;
[0162] Therefore, the weighting coefficient can be:
[0163] ;
[0164] in, , , ; It should be noted that the introduction here This is to avoid the denominator in the formula being zero;
[0165] With the weighting factor The distribution of , the SVM model becomes:
[0166] ;
[0167] To solve the weighted coefficient, the Lagrangian function is constructed as follows:
[0168] ;
[0169] Calculation parameters, The partial derivative of is:
[0170] ;
[0171] in, is the partial derivative; is the above weighting coefficient;
[0172] Finally, the decision function of the weighted support vector machine WSVM is:
[0173] .
[0174] In this embodiment, by weighting the support vector machine SVM, a weighted support vector machine WSVM is obtained as a coarse classifier to identify and separate easily distinguishable faults, and at the same time, the classification performance of the balance optimizer MEO is improved by improving the balance optimizer MEO; the use of the weighted support vector machine WSVM can reduce the impact of outliers on the total training error, thereby improving the recognition accuracy.
[0175] S5. For the faults that cannot be directly identified in step S4, by calculating the Fisher discriminant ratio of each monitoring variable in the operation data, the key monitoring variable data is selected and identified based on the detailed classifier to obtain the accurate fault type.
[0176] As a preferred embodiment of step S5, the following steps are specifically included:
[0177] S51. For the faults that cannot be directly distinguished in step S4, a sparse momentum deep belief network (SMDBN) with strong feature extraction capability is used for fine classification. Before fine classification, the Fisher discriminant ratio is used to select key monitoring variable data. Variable The Fisher discriminant ratio is:
[0178] ;
[0179] in, represents the number of categories, Respectively represent variables The mean and variance of the Fisher discriminant ratio; usually, the larger the value of the Fisher discriminant ratio, the better.
[0180] S52. Construct a sparse momentum deep belief network SMDBN. The sparse momentum deep belief network SMDBN is a deep learning model composed of three Boltzmann machines RBMs and a softmax layer. The parameter update process of SMDBN is:
[0181] ;
[0182] in, is the connection weight between the visible layer and the hidden layer, and are the biases of the visible layer and the hidden layer respectively, is the number of iterations, is the momentum factor and , is the learning rate, is the mathematical expectation of the training data, Indicates The mathematical expectation of the sample distribution after step Gibbs sampling, For the neurons in the visible layer, For the hidden layer neurons, is the regularization factor, and Given by:
[0183] ;
[0184] in, is the number of neurons in the hidden layer, For the hidden layer activation probability; is the variance factor, For the neurons in the visible layer;
[0185] S53. Input the key monitoring variable data selected based on the Fisher discriminant ratio into the constructed sparse momentum deep belief network SMDBN to identify the accurate fault.
[0186] In this embodiment, a sparse momentum deep belief network (SMDBN) is constructed as a fine classifier to identify fault types that cannot be directly identified by the weighted support vector machine (WSVM) as a coarse classifier, and its classification performance is improved by improving the balance optimizer (MEO). The sparse momentum deep belief network (SMDBN) enhances the generalization ability of the model by adding sparse terms and momentum terms in parameter updates, thereby improving the accuracy of fault identification and diagnosis.
[0187] Finally, the present invention takes a certain type of phased array radar online monitoring data provided by a research institute as an example to verify the method proposed in this application. Figure 1 As shown, the sampling interval is 1 minute, and one of the 8-channel T / R modules is selected as the research object. A total of 1800 labeled samples are collected, 26 monitoring variables, and 300 samples for each of the 6 working conditions. The 6 operating states refer to 5 fault states and 1 normal state. In turn, they can be coded as 1-6. Please note that although the normal state is not a fault, it is represented by fault code 6 for ease of description. The 1800 labeled samples are divided into test set and training set at a ratio of 1:5. That is, 1500 labeled samples are used to train the classifier, and 300 labeled samples are used as the test set.
[0188] In the two-layer classifier, an improved balanced optimized weighted support vector machine MEO-WSVM is established in the first layer for fault identification, and six types of faults are separated at a time, namely PU, PO, TF, RF, TRF and NS. In the experiment, ten-fold cross validation is used to divide the data set into 10 parts, 9 of which are used as training data and the remaining 1 is used as validation data. Each experiment will get the corresponding recognition rate, and the validation set will be replaced ten times. The average value of the results will be used as an estimate of the classifier accuracy. In the improved balanced optimizer MEO, the population size is set to 30 and the maximum number of iterations is 100. In the weighted support vector machine WSVM, the value range of the penalty factor and the kernel parameter is limited to between 0.01-100.
[0189] like Figure 3 As shown in Figure 2, the changes in the fitness function values of the three optimizers during the optimization process are shown. Figure 3 It can be seen that the improved balance optimizer MEO has higher search accuracy. This is mainly because MEO uses Tent chaos mapping to increase the diversity of the population in the early stage, and further uses Gaussian mutation to prevent the optimizer from falling into premature convergence in the later stage. Therefore, MEO can break through the local optimum. Next, the well-trained MEO-WSVM is used for the validation set. Figure 4 As shown, Figure 4 is the confusion matrix of the MEO-WSVM validation set recognition rate (%). Figure 4 , the confusion degree between the six fault categories can be calculated (see Table 1). As can be seen from Table 1, , Since the confusion between RF and TRF is high, and the confusion between RF and NS is also high, RF, TRF, and NS can be considered as a group, while the remaining PU, PO, and TF can be considered as three separate categories. Therefore, a two-layer classifier can be constructed, such as Figure 5 shown.
[0190] Table 1 Confusion degree of six types of faults based on MEO-WSVM (%)
[0191] Fault PU PO TF RF TRF NS PU 0.00 PO 0.00 0.00 TF 2.00 0.00 0.00 RF 0.00 1.00 0.00 0.00 TRF 0.00 0.00 2.00 9.00 0.00 NS 0.00 3.00 0.00 16.00 1.00 0.00
[0192] According to the construction principle of the two-layer classifier, RF, TRF and NS need to be reclassified. In this case, the key state monitoring data is selected by calculating the Fisher discriminant ratio of 26 monitoring variables, and then a sparse momentum deep belief network SMDBN is constructed to determine the fault type. Similar to WSVM, reasonable SMDBN configuration parameters are determined by MEO. Figure 6 As shown, Figure 6The Fisher discriminant ratios of 26 monitoring variables are shown. The present application considers that monitoring variables with Fisher discriminant ratios greater than 0.2 are more important, so monitoring variables 1, 2, 3, 4, 7, 11, 15, 19 and 24 are used as inputs to the improved balanced optimizer-sparse momentum deep belief network model MEO-SMDBN. Figure 7 This is the final classification result of the six types of faults. Figure 7 It can be seen that the first-layer classifier, the improved balanced optimized weighted support vector machine (MEO-WSVM), can accurately classify PU, PO and TF, with fault recognition rates of 96.00%, 96.00% and 100.00% respectively. For the three easily confused faults of RF, TRF and NS, the fault recognition rates of the second-layer classifier, the improved balanced optimizer-sparse momentum deep belief network MEO-SMDBN, are 96.00%, 94.00% and 94.00% respectively. The results show that the average recognition rate of the proposed two-layer classifier is as high as 96%.
[0193] The present invention constructs a two-layer classifier for phased array radar transceiver fault diagnosis. In the two-layer classifier, the weighted support vector machine WSVM is used as a coarse classifier to separate easily distinguishable faults, while the sparse momentum deep belief network SMDBN is used as a fine classifier to identify faults that WSVM cannot directly distinguish. The reasonable configuration parameters of SVM are determined by MEO DBN (deep belief network). In order to verify the effectiveness of each layer module in the proposed two-layer classifier, an ablation experiment is designed. The ablation experiment results are shown in Figure 2. Figure 8 shown.
[0194] In the coarse classifier, different SVMs have different recognition rate levels, such as Figure 8 As shown in (a) in the figure, the recognition rate of WSVM is improved to 90% compared with the baseline support vector machine SVM, and the recognition rate of MEO-SVM is further improved to 90.67%. In particular, when the weighted strategy and configuration optimization are applied simultaneously, the recognition rate of MEO-WSVM reaches 91.67%, significantly exceeding the baseline model and the single improved model.
[0195] In the fine classifier, by comparing the recognition rates of different DBNs, the impact of various improvements can be clearly observed. Figure 8 As shown in (b), the recognition rate of the baseline DBN model is 92%. After a single improvement, the recognition rate of the sparse momentum deep belief network SMDBN is increased to 93%, and the recognition rate of MEO-DBN is 93.33%, indicating that a single improvement is effective in improving the recognition rate. However, it is worth noting that when considering multiple improvement measures, the recognition rate of the MEO-SMDBN model is further improved to 94%, which is significantly better than the baseline model and the single improvement model.
[0196] Finally, the influence of combining high-performance coarse classifiers with fine classifiers on recognition performance is discussed. Figure 8 As shown in (c) in the figure, the recognition rate of the coarse classifier alone is 91.67%, and the recognition rate of the fine classifier alone is 94%. However, when the two are combined to form a two-layer classifier, the recognition rate is significantly improved to 96%. This shows that there is a certain degree of complementarity between the coarse classifier and the fine classifier, and the combination of the two can effectively improve the overall recognition performance.
[0197] In summary, the verification results on the phased array radar transceiver dataset show that the proposed two-layer classifier can effectively identify easily confused faults, achieve more accurate fault identification results, and thus reduce the fault diagnosis cost.
[0198] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.
[0199] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, which can be embodied in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, apparatus or device and execute instructions), or used in combination with these instruction execution systems, apparatuses or devices.
[0200] The above implementation methods have been described in detail. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.
Claims
1. A phased array radar transceiver fault identification method based on a double-layer classifier, characterized in that: The method comprises the following steps: S1. Obtaining the operating data of the phased array radar transceiver; S2. constructing a double-layer classifier based on the operating data of the phased array radar transceiver, wherein the double-layer classifier is composed of a coarse classifier and a fine classifier; The specific process of building a two-layer classifier is: S21. Set a fault set F, and let the fault confusion between the i-th type fault and the j-th type fault be I ij , fault confusion degree I ij It represents the average of the probability that the i-th type fault is misjudged as the j-th type fault and the probability that the j-th type fault is misjudged as the i-th type fault, that is: Among them, x is the test sample, y is the corresponding recognition result in the test sample, and the fault confusion refers to the similarity between different types of faults; S22, the coarse classifier is a weighted support vector machine WSVM, and the fine classifier is a sparse momentum deep belief network SMDBN; S3. Parameter configuration of the double-layer classifier is performed by improving the balanced optimizer MEO. The improved balanced optimizer MEO adopts Tent chaos mapping and Gaussian mutation to improve the performance of the optimizer. In the early stage, Tent chaos mapping is used to increase the population diversity and accelerate the convergence speed of the optimizer. In the later stage, Gaussian mutation is used to avoid the optimizer from converging too early. The parameter configuration of the two-layer classifier based on the improved balanced optimizer MEO is as follows: S31. Initialize the parameters of the improved equilibrium optimizer MEO. Each individual position is composed of model configuration parameters. N individuals in the feasible domain space are initialized by Tent chaotic mapping to form an initial population, namely: in, represents the position of the i-th (i=1,2,...,N) individual in the d-th (d=1,2,...,D) dimension; D represents the dimension of the optimization problem; and They are the upper and lower limits of the d-dimensional variable in the search space; is the Tent chaotic sequence, which is given by the following formula: Among them, r1 and r2 are random numbers between 0 and 1; S32, establish a balanced state pool and select candidates; Balance Pool Q eq_pool The definition is as follows: Q eq_pool ={Q eq1 ,Q eq2 ,Q eq3 ,Q eq4 ,Q eq_ave }; Among them, {Q eq1 ,Q eq2 ,Q eq3 ,Q eq4 } represents the four particles with the highest concentrations selected after the initialization phase of the improved equilibrium optimizer MEO is completed; Q eq_ave represents the average particle and is given by: In the iterative process of the improved equilibrium optimizer MEO, each particle is selected with the same probability from the five candidate particles in the balanced state pool to generate the global optimal solution; S33, improve the attention of the improved balanced optimizer MEO in the process of parameter configuration of the two-layer classifier; S34, executing Gaussian mutation update; First, select mutant individuals from the current population, apply Gaussian distribution to each selected individual to generate new parameters and ensure that the parameter value after mutation is within the predefined range; Secondly, calculate the fitness of the new individual after mutation, and decide whether to accept the new individual after mutation according to the fitness value; Then, the new individuals that have undergone mutations are cross-operated with individuals in the original population to generate new candidate solutions; Finally, after a new round of selection and update, the population is updated, and the improved equilibrium optimizer MEO enters the next generation of iterations. After multiple iterations, it gradually approaches the global optimal solution; S35, judging whether the end condition is met, if the current iteration exceeds the maximum iteration, outputting the optimal individual position in the population; otherwise, executing step S32; S36, obtaining parameter configuration according to the optimal individual position in the population, and retraining the coarse classifier or the fine classifier; S4, performing fault identification and classification on easily distinguishable monitoring variables in the operation data of the phased array radar transceiver based on a coarse classifier; The specific steps include: S41, update the support vector machine SVM model, the SVM model is: Among them, w is the weight, y i is the recognition result corresponding to the i-th test sample, p is the hyperplane normal vector, and p T is the transpose of the hyperplane normal vector, C is the penalty factor, ξ i is the slack variable, is a nonlinear mapping function, q is the deviation, and l is the number of samples; S42, map the sample set from the input space to the high-dimensional feature space through kernel mapping, construct the Lagrangian function and decision function, whose expressions are as follows: Where L is the constructed Lagrangian function; ζ is the slack variable; α, β are Lagrangian multipliers; α i and β i is the Lagrange multiplier of the i-th sample; K(x,x i ) is the kernel function for realizing nonlinear transformation; sgn represents the sign function; S43, determine the radial basis function, whose expression is: K(x,x i )=exp(-g‖x,x i ‖ 2 ); Where exp represents the exponential function with the natural constant e as the base, g is the hyperparameter of the radial basis function, ‖x,x i ‖ is the vector x and x i The distance between S44, weighting the support vector machine SVM to obtain a decision function of the weighted support vector machine WSVM; Let the total number of positive and negative samples be l + and l - Indicates that the centers of positive and negative samples are defined as and Positive sample radius r + And the radius of negative samples r - It can be regarded as the maximum distance between positive and negative samples and the corresponding class center; Therefore, the weighting coefficient can be: in, χ=10 -6 ; It should be noted that χ is introduced here to avoid the denominator in the formula being zero; With the weighting coefficient μ i The distribution of , the SVM model becomes: To solve the weighted coefficient, the Lagrangian function is constructed as follows: Calculate the parameters p, q, ξ i The partial derivative of is: in, is the partial derivative; μ i is the above weighting coefficient; Finally, the decision function of the weighted support vector machine WSVM is: S5. For the faults that cannot be directly distinguished in step S4, by calculating the Fisher discriminant ratio of each monitoring variable in the operation data, selecting key monitoring variable data and identifying the key monitoring variable data based on the detailed classifier, the accurate fault type is obtained; The specific steps include: S51, using Fisher discriminant ratio to select key monitoring variable data; The Fisher discriminant ratio for variable k is: Among them, z represents the number of categories, o ik ,o jk , They represent the mean and variance of variable k respectively; S52, construct sparse momentum deep belief network SMDBN; The Sparse Momentum Deep Belief Network (SMDBN) is a deep learning model consisting of three Boltzmann machines (RBMs) and a softmax layer. The parameter update process of SMDBN is: Where W is the connection weight between the visible layer and the hidden layer, a and b are the biases of the visible layer and the hidden layer respectively, τ is the number of iterations, κ is the momentum factor and 0≤κ<1, η is the learning rate, <> d is the mathematical expectation of the training data, <> k represents the mathematical expectation of the sample distribution after the k-th step Gibbs sampling, v i is the i-th visible layer neuron, h j is the jth hidden layer neuron, σ is the regularization factor, and Given by: Among them, m is the number of neurons in the hidden layer, u j is the activation probability of the jth hidden layer; ρ is the variance factor, v i is the i-th visible layer neuron; S53. Input the key monitoring variable data selected based on the Fisher discriminant ratio into the constructed sparse momentum deep belief network SMDBN to identify the accurate fault.
2. The phased array radar transceiver fault identification method based on a double-layer classifier according to claim 1 is characterized in that: In step S1, obtaining the operation data of the phased array radar transceiver includes: the phased array radar antenna receives the electromagnetic signal reflected from the target, and performs data fusion and output processing on the received signal through analog-to-digital conversion, filter matching, and pulse compression operations.
3. The phased array radar transceiver fault identification method based on a double-layer classifier according to claim 1 is characterized in that: In step S33, the attention of the improved balanced optimizer MEO in the process of parameter configuration of the double-layer classifier is improved, specifically including: S331. Index term U is an important indicator to measure the exploration and development capability of the improved balance optimizer MEO, and its expression is: U=a1sign(r-0.5)(e -λt -1); Where a1 is a constant that controls the optimizer's exploration capability, sign is a sign function, r and λ are random numbers between 0 and 1, and t is a coefficient that is updated with the number of iterations, given by: Among them, Iter and Max_iter represent the current number of iterations and the maximum number of iterations respectively; a2 is a constant that controls the development capability of the optimizer; S332, introduce the mass generation rate G, that is: G=GCP(Q eq -λQ)·U; Among them, GCP is the control parameter vector of the generation rate; Q eq is the average particle concentration; Q is the current particle concentration; r3 and r4 are random numbers between 0 and 1; GP is a constant with a value of 0.5; Each particle is updated as follows: Where V is the unit volume.
