A phased array radar element fault diagnosis method based on dual-stream DCNN

Abstract

The present invention's phased array radar element fault diagnosis method based on dual-stream DCNN constructs a planar array mathematical model, analyzes the changes in element excitation from the perspectives of amplitude, phase error and position deviation according to different element fault types, uses array steering vectors to give array echo signal models for these types of element faults, and proposes a dual-stream DCNN diagnostic model, which uses training samples of array echo signals as input, calculates error matrices from them through data preprocessing, and converts grayscale images of equal size; then uses two branches to extract the directional pattern features of the faulty array from the image, combines the features extracted by the two branches, generates a new feature map, and then extracts the features for classification. After multiple iteration cycles of training, the model can accurately locate the position of the faulty element. The present invention compares the performance of the diagnostic model under different SNRs, can quickly locate the faulty components in the planar element, and ensures the reliability of the radar.

Description

Technical Field

[0001] The invention belongs to the technical field of radar testing, and in particular relates to a method for locating a faulty array element in a phased array radar. Background Art

[0002] Phased array radars are widely used in military and civilian fields, especially military phased array radars, which are getting larger and larger, reaching thousands or even tens of thousands of array elements. Due to the high frequency of radar use, it may work 24 hours a day. Each channel of the active phased array has a separate transmitter and receiver. After long-term use, the components of the transceiver channel are prone to aging, resulting in the failure or failure of the corresponding array element, the radar antenna's directional pattern changes, and the detection performance is affected. Therefore, health management of phased array radars, timely detection of faulty array elements and repairs can avoid equipment failures and performance degradation caused by the gradual deterioration of system hardware during its service life. Early fault diagnosis can effectively prevent the radar from having more false alarms and missed detections due to performance degradation, and improve the system's security capabilities.

[0003] The radar radiation pattern of a phased array radar is synthesized by the excitation of multiple array elements, and has the characteristics of high gain and low side lobes. A faulty array element will change the characteristics of the antenna pattern, especially the sidelobe gain change of the pattern. Transmitters and receivers are usually composed of amplifiers, filters, mixers, and switches. After aging of the device, the gain and delay time change, which is manifested as the I / Q amplitude error and phase error of the output complex signal. In addition, the surface deflection, installation problems and surface vibration of the antenna array may cause the position of the array element to change. Amplitude and phase errors and position deviations cause the array pattern to change. Therefore, extracting features from the pattern can detect faults and locate the faulty array element. Summary of the invention

[0004] In order to manage the health of phased array radar systems, the present invention proposes a method for quickly diagnosing and locating radar faulty array elements by identifying array radiation patterns. By analyzing the main causes of array element failure, it is concluded that array element failure is manifested as channel amplitude error, phase error, and array element position deviation. A mathematical model of the array element radiation pattern under amplitude, phase error, and position deviation is established, and an echo signal model when the array element fails is derived. A dual-stream DCNN model for extracting features from echo signals is given. In view of the problems that echo signals are difficult to process and features are not obvious, a general preprocessing method is given.

[0005] The technical solution adopted by the present invention is:

[0006] Establish the directional pattern model of the faulty array. The array directional pattern is synthesized by the excitation of the normal array element and the faulty array element, and the uniform planar array is used to simplify the calculation derivation. According to the fault type of the faulty array element, the excitation change of the array element is analyzed, and the directional pattern is derived from the array steering vector.

[0007] An array echo model is established. The array pattern information needs to be obtained from the echo signal. The target echo is measured at multiple azimuths in the array far field. A narrowband impulse signal is used to simulate signal sampling. The noise signal is additive white Gaussian noise.

[0008] A method for diagnosing faults using a dual-stream DCNN model is proposed. In order to solve the problem that the fault array directional pattern does not change significantly, a convolutional neural network model is used to extract its features. The array echo signal is used as the training sample of the dual-stream DCNN diagnosis model. The input sample is preprocessed to reduce the amount of calculation. The directional pattern features are extracted through two branch structures. K-fold cross validation is used for testing to reduce the deviation under random data distribution. After multiple iterations, the training is completed and the performance of the model in locating the faulty array element is evaluated.

[0009] The specific steps are as follows:

[0010] Step 1: Construct a mathematical model of a planar array. Use a uniform planar array, simplify the operation relationship, and calculate the array element excitation and array synthesis pattern.

[0011] Step 2: Measure the target echoes in multiple directions in the far field and establish the array echo mathematical model. Based on the principle of phased array pattern synthesis, deduce the relationship between array element excitation and echo signal, and calculate the array steering vector.

[0012] Step 3: According to the type of faulty array element, establish the corresponding array element excitation mathematical model under amplitude, phase error and position deviation, calculate the synthetic directional pattern under array element failure, and generate echo signals under different SNRs.

[0013] Step 4: Build a dual-stream DCNN diagnostic model, use the array echo signal as the training sample of the model input, and complete the training after multiple iterations.

[0014] Step 5: Evaluate the diagnostic performance of the model and use cross-validation technology to evaluate the diagnostic performance of the model under different SNRs.

[0015] Each training process of the two-stream DCNN model is divided into the input and preprocessing stage, the feature extraction stage, and the classification stage. The characteristics are data preprocessing and two branches to extract features, using an adaptive gradient algorithm optimizer and classification cross entropy as the loss function. The training process is divided into the following steps:

[0016] In the first step, the input and preprocessing stage, the error matrix is ​​calculated based on the amplitude and phase information of the input echo signal, and then it is organized into a grayscale image of fixed size to reduce the amount of calculation and subsequent feature extraction.

[0017] The second step is the feature extraction stage, which includes convolutional layers, batch normalization layers, and pooling layers. Two branch structures are used to extract features, and then the extraction results of the two branches are combined to generate a new feature map, and then the features are extracted for classification.

[0018] In the third step, the classification phase, the fully connected layer flattens all features and obtains the output result through the ReLU activation function. In each iteration, the performance of the model is evaluated using accuracy, precision, recall rate, and f1 value, and the optimal model is saved. Training is completed after multiple iterations.

[0019] The beneficial effect of the present invention is that a mathematical model of array element excitation under amplitude, phase error and position deviation is established, and a mathematical model of the fault array pattern and a mathematical model of the echo signal are further derived. A method for diagnosing faults using a dual-stream DCNN model is proposed, and the echo signal is converted into a grayscale image using data preprocessing, which reduces the amount of calculation and facilitates subsequent feature extraction. It is also generally applicable to unknown data. The feature extraction adopts a two-branch structure, which is conducive to retaining data features and improving the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a principle flow chart of the present invention.

[0021] Figure 2 Schematic diagram of the planar phased array model used.

[0022] Figure 3 Radiation patterns of the 10x10 array under different fault conditions, (a) and (d) are fault-free arrays;

[0023] (b) is an array with a fault in the edge element (1,10); (e) is an array with a fault in the center element (5,5); (c)

[0024] (f) Respective differences between fault-free and faulty modes.

[0025] Figure 4 The proposed two-stream DCNN diagnosis model.

[0026] Figure 5 Similarity comparison between faulty and non-faulty elements in a 10x10 array.

[0027] Figure 6 The performance evaluation results of the diagnostic model under different SNRs. DETAILED DESCRIPTION

[0028] The principle flow chart of the present invention is as follows Figure 1 As shown, the fault diagnosis method of the present invention is further introduced below through specific embodiments in conjunction with the accompanying drawings.

[0029] Furthermore, the detailed process of step 1 is as follows:

[0030] Configure a uniform planar array mathematical model, assuming that the array elements of the planar array antenna are uniformly arranged on a rectangular grid, such as Figure 2 (a), depicting M x Line and M y The spacing between columns, rows and array elements is d x and d y The array arrangement, the physical structure of each array element and the array element excitation determine the performance of the array synthetic beam. The phased array used in the present invention has a structure such as Figure 2 As shown in (b), the spacing between adjacent array elements in each row and column is equal.

[0031] Assuming that all array elements are identical and mutual coupling effects are ignored, the radiation pattern of a planar array antenna can be expressed as the product of a unit factor and two linear array factors:

[0032]

[0033] where a mn is the excitation of the mnth array element, λ is the wavelength, d x , d y are the array element spacing along the x and y directions, respectively.

[0034] Furthermore, the detailed process of step 2 is:

[0035] Step 2.1 introduces the planar array structure in step 1, such as Figure 2 As shown in (b), the far-field target echo is measured and K narrow-band impulse signals are generated simultaneously. Where t = 1, 2, ..., T is the sample index. For an ideal array, the received far-field echo can be modeled as:

[0036]

[0037] Among them, z(t) represents the echo vector, s(t) is the narrowband signal vector, n(t) has a mean of zero and a variance of Additive white Gaussian noise vector that is uncorrelated with the signal.

[0038] Step 2.2 calculates the array steering vector. The direction matrices along the x and y directions are:

[0039]

[0040] The corresponding steering vector is:

[0041]

[0042] in The Khatri-Rao product and the Kronecker product are denoted as ⊙ and

[0043] Furthermore, the detailed process of step 3 is as follows:

[0044] Step 3.1 Establish an element fault model in a phased array. The directivity of an array antenna is a function of radiation intensity and is determined by the shape of the antenna’s pattern. Deterioration of the pattern shape will adversely affect the directivity of the array. When an element fails, the radiation pattern of the entire array will be distorted. There may be only slight differences in the main lobe width and side lobe level between a faulty array and a fault-free array, which are difficult to distinguish through initial observation. Therefore, it is necessary to extract the features of each fault state to distinguish them. Common element fault types can be modeled in terms of amplitude, phase, and position to illustrate the impact of the faulty element on the composite pattern. By

[0045] The composite pattern of the array A total (θ, φ), can be regarded as the sum of the excitations of each array element, including the faulty array element A error (θ, φ) and ideal array element A desired (θ,φ):

[0046] A total (θ,φ)=A desired (θ,φ)+A error (θ,φ) (5)

[0047] Considering the amplitude and phase disturbances, the excitation on a single array element can be expressed as:

[0048] a mn =a mn (1+Δ mn )exp(jδ mn ) (6)

[0049] where Δ mn and δ mn Indicates small changes in the amplitude and phase of the excitation.

[0050] Considering the influence of position deviation, in practice, the position of each array element may have a slight deviation Δ from the ideal position i :

[0051]

[0052] Where N x , N y represents the random deviation value. At this time, the steering vector of the array can be expressed as:

[0053]

[0054] Step 3.2 If the array structure is as in step 1, and the array element spacing in the x and y directions is set to The steering vector can be further expressed as:

[0055]

[0056] in represents the Schur-Hadamard product. The steering vector with position deviation is:

[0057]

[0058] in

[0059]

[0060] Use F ap is a matrix representing the amplitude and phase deviation, and the steering vector can be expressed as:

[0061]

[0062] Step 3.3 The performance of the phased array radar will gradually decrease under the influence of the faulty array element. After considering the above errors, the received signal of the array antenna is:

[0063] z(t)=FA(θ,φ)s(t)+n(t) (13)

[0064] where F represents the matrix of magnitude, phase, and position errors.

[0065] Furthermore, the detailed process of step 4 is as follows:

[0066] Step 4.1 Generate training samples. According to the phased array modeling method in step 1, a traditional 10×10 planar array with an element spacing of λ / 2 is used. Each element in the array is simulated to fail once, resulting in 100 different fault conditions. In order to generate training data for the model, 1,000 Monte Carlo simulation far-field samplings are performed at (θ,φ)=[0,0] to generate radiation pattern samples for 100 element fault states. In the same way, a data sample is generated when all elements are normal, such as Figure 3 As shown in the figure, these 101 training data sets represent the far-field radiation information generated by the phased array at a given signal-to-noise ratio. The generated noise samples correspond to the SNR range of 0dB to 25dB, with a step size of 5dB, and each SNR data set includes 101×1000 instances.

[0067] Step 4.2 builds a dual-stream DCNN diagnostic model. The diagnostic model structure proposed in the present invention is as follows: Figure 4, which is mainly divided into the input and preprocessing stage, the feature extraction stage and the classification stage. Each training sample contains the amplitude and phase information of the echo signal. During the preprocessing process, the error matrix is ​​obtained and a grayscale image of 64x64 size is generated. After the sample data is processed, it is divided into two branches for feature extraction, and each branch model has five layers. The input of the model is a 64×64 matrix, which is convolved in a convolution kernel of size 3×3, with a convolution step of 1, generating 32 feature maps. The rectified linear unit (ReLU) performs a restriction operation on each input element in the convolution layer, assigning zero to all values ​​below zero. Subsequently, the batch normalization layer keeps the data samples between the intermediate layers in the same distribution, ensuring fast learning and improved performance. The pooling layer size is 2×2, and the data is downsampled.

[0068] The above layers are copied for the next feature extraction. The two branches of the DCNN model are then combined to generate a feature map of size 16×16×128. Next, a convolutional layer with a convolution kernel of size 3×3 and a convolution step of 1 is used to generate 64 feature maps, which are then subjected to ReLU and batch normalization operations. After downsampling through a 2×2 pooling layer, the fully connected layer flattens all features into one dimension. The subsequent classification stage includes two hidden layers with ReLU activation function. Finally, the output layer includes 101 units and a Softmax activation function.

[0069] Step 4.3 completes the model training experiment. The experiment uses an adaptive gradient algorithm optimizer and categorical cross entropy as the loss function, with an initial learning rate of 0.001, a batch size of 64, and a learning rate that decreases exponentially after the first five epochs. In each iteration, the performance of the model is evaluated and the optimal model is saved. The model proposed in the present invention is implemented using Keras 2.3.1API and Python 3.7 on a system equipped with a 23.70GB NVIDIA GeForce RTX 3090 GPU, and the model is trained for 50 epochs.

[0070] Furthermore, the detailed process of step 5 is as follows:

[0071] Step 5.1 compares the eigenvalue results and similarity. When an array element fails, the radiation pattern will be distorted. In order to facilitate the analysis of the degree of change in the antenna pattern when an array element fails in the array, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index are used for measurement. PSNR is the ratio of the highest achievable signal strength to the distortion noise power that affects the quality of its representation. PSNR values ​​close to infinity indicate a higher image similarity, while low values ​​indicate a significant difference between the images. The SSIM index is a metric used to compare the similarity of two images, ranging from 0 to 1, where a value of 0 indicates no similarity between the images and a value of 1 indicates a high degree of similarity. Figure 5 The eigenvalue results of array element failures at different locations under different SNRs of the model are given.

[0072] Step 5.2 evaluates the performance of identifying faulty array elements under different signal-to-noise ratios. Accuracy, precision, recall, and f1-score are used as performance evaluation indicators of the fault classification model.

[0073]

[0074] Where N c is the number of groups, TP and FP are the number of true and false positives, respectively. And, TN and FN are the number of true and false negatives, respectively. In order to reduce the bias caused by random data allocation, K-fold cross validation is used. All sample data are divided into K subsets. In each iteration, K-1 subsets are used as training samples, and the remaining 1 is used as a test sample to evaluate the performance of the classification model. The experimental results using the 5-fold cross validation technique are shown in Table 1. The performance evaluation results of the model under different SNRs are shown in Figure 6 shown.

[0075] It should be understood that the examples and implementation modes described herein are for illustrative purposes only and are not intended to limit the present invention. Those skilled in the art may make various modifications or changes based on the examples and implementation modes. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0076] Table 1 Diagnostic performance of the diagnostic model under different SNRs

[0077]

[0078]

Claims

1. A phased array radar element fault diagnosis method based on dual-stream DCNN, characterized in that: The following steps are included: Step 1: Construct a planar array mathematical model, use a uniform planar array, simplify the operation relationship, calculate the array element excitation and array synthesis pattern, The step 1 configures a uniform planar array mathematical model, assuming that the array elements of the planar array antenna are uniformly arranged on a rectangular grid, and describes M x Line and M y The array element spacing between columns, rows and columns is d x and d y , using a phased array, the spacing between adjacent array elements in each row and column is equal, assuming that all array elements are identical, and ignoring mutual coupling effects, the radiation pattern of the planar array antenna is expressed as the product of a unit factor and two linear array factors: where a mn is the excitation of the mnth array element, λ is the wavelength, d x , d y are the array element spacing along the x and y directions respectively; Step 2: Measure the target echoes in multiple directions in the far field, establish the array echo mathematical model, deduce the relationship between array element excitation and echo signal based on the phased array pattern synthesis principle, and calculate the array steering vector. Step 2 introduces the planar array structure in step 1 to measure the far-field target echo and generate K narrowband impulse signals at the same time. Where t = 1, 2, ..., T is the sample index. For an ideal array, the received far-field echo is modeled as: z(t)=A(θ,φ)s(t)+n(t) Among them, z(t) represents the echo vector, s(t) is the narrowband signal vector, n(t) has a mean of zero and a variance of An additive white Gaussian noise vector that is uncorrelated with the signal; The step 2 calculates the array steering vector, and the direction matrices along the x and y directions are respectively: ， The corresponding steering vector is: ， in The Khatri-Rao product and the Kronecker product are denoted as ⊙ and Step 3: According to the type of faulty array element, a corresponding array element excitation mathematical model under amplitude, phase error and position deviation is established, the synthetic directional pattern under array element failure is calculated, and echo signals under different SNRs are generated; Step 4: Build a dual-stream DCNN diagnostic model and use the array echo signal as the training sample for the model input. After multiple iterations, the training is completed. The training process is divided into the following steps: The first step is the input and preprocessing stage. According to the amplitude and phase information of the input echo signal, the error matrix is ​​calculated and then sorted into a grayscale image of fixed size to reduce the amount of calculation and subsequent feature extraction. The second step is the feature extraction stage, which includes convolutional layer, batch normalization layer and pooling layer. Two branch structures are used to extract features, and then the extraction results of the two branches are combined to generate a new feature map, and then the features are extracted for classification; The third step is the classification stage. The fully connected layer flattens all features and obtains the output result through the ReLU activation function. In each iteration, the performance of the model is evaluated by accuracy, precision, recall rate, and f1 value. The optimal model is saved and the training is completed after multiple iterations. Step 5: Evaluate the diagnostic performance of the model and use cross-validation technology to evaluate the diagnostic performance of the model under different SNRs.

2. The method for phased array radar element fault diagnosis based on dual-stream DCNN according to claim 1 is characterized in that: The step 3 establishes an element fault model in the phased array, and models the amplitude, phase and position to explain the influence of the faulty element on the synthetic pattern. total (θ, φ), is considered as the sum of the excitations of each element, including the faulty element A. error (θ, φ) and ideal array element A desired (θ,φ): A total (θ,φ)=A desired (θ,φ)+A error (θ,φ), Considering the amplitude and phase disturbances, the excitation on a single array element is expressed as: a mn =a mn (1+D mn )exp(jδ mn ), Where Δ mn and δ mn represents small changes in the excitation amplitude and phase, Considering the influence of position deviation, in practice, the position of each array element may have a slight deviation Δ from the ideal position i : Where N x , N y represents the random deviation value, and the steering vector of the array is expressed as:

3. The method for phased array radar element fault diagnosis based on dual-stream DCNN according to claim 1 is characterized in that: The received signal of the array antenna is: z(t)=FA(θ,φ)s(t)+n(t), where F represents the matrix of magnitude, phase, and position errors.

4. The method for phased array radar element fault diagnosis based on dual-stream DCNN according to claim 1 is characterized in that: Each training process of the dual-stream DCNN diagnostic model in step 4 is divided into an input and preprocessing stage, a feature extraction stage, and a classification stage.

5. The method for phased array radar element fault diagnosis based on dual-stream DCNN according to claim 1 is characterized in that: The step 5 compares the eigenvalue results and the similarity, and adopts the peak signal-to-noise ratio PSNR and the structural similarity SSIM index measurement, where PSNR is the ratio of the highest achievable signal strength to the distortion noise power that affects its representation quality, and SSIM index is an indicator for comparing the similarity of two images.

6. The method for phased array radar element fault diagnosis based on dual-stream DCNN according to claim 1 is characterized in that: The step 5 evaluates the performance of identifying faulty array elements under different signal-to-noise ratios, using accuracy, precision, recall and f1-score as performance evaluation indicators of the fault classification model.

Images