A method for antenna array fault diagnosis based on data tolerance analysis extension

Abstract

The present invention belongs to the field of antenna array technology, and specifically relates to an antenna array fault diagnosis method based on data tolerance analysis expansion. In the data processing stage, far-field radiation data in multiple directions are measured in the far-field area to generate power information, and the measured power pattern is preprocessed using tolerance analysis to obtain the upper and lower bounds of the power pattern considering the array element position error, thereby expanding the data set. In the training stage, the training of the fault diagnosis model is completed, the expanded data set is used as the input of the neural network, the position and number of failed array elements are set as output, and two neural networks are trained. In the application stage, data is collected in the same direction as the data measured in the training stage, and the obtained data to be tested is input into the model to preliminarily determine the number of failed array elements and the probability of array element failure, and then data compensation is performed on the test results according to the radiation compensation formula to determine the position of the failed array elements in turn.

Description

Technical Field

[0001] The invention belongs to the technical field of antenna arrays, and in particular relates to an antenna array fault diagnosis method based on data tolerance analysis expansion. Background Art

[0002] Antenna arrays are special antennas composed of multiple array units arranged in a certain position. Antenna arrays are often used in automotive radar, 5G communication, satellite communication, and drones. Generally, the more antenna units, the better the radiation performance. For example, in phased array radars, the number of antenna units is hundreds or thousands. As the scale increases, the failure probability of antenna units will also increase. The failure of a certain number of antenna units will cause the main lobe to be severely widened, the side lobe level to increase, and the gain to decrease, affecting the array performance. In order to eliminate the impact of antenna unit failure, it is very important to accurately locate the number and position of failed array elements. Existing array fault diagnosis methods can be divided into source reconstruction method, very near field method, and parameter model method. Existing methods mostly consider antenna array element fault diagnosis under ideal or Gaussian noise error conditions. In actual engineering applications, antenna arrays are affected by installation accuracy, component aging, or process problems, and there are certain random errors compared with ideal conditions, including frequency offset, radiation phase error, and array position error. The accuracy of array fault diagnosis using existing methods in the presence of random errors is reduced. The parameter model method has a reliable diagnostic effect and a faster diagnostic speed. Its training phase is limited by the problem of data sets, and the accuracy is also reduced in the case of random errors. With limited data sets, it is very important to preprocess the data and perform error analysis.

[0003] There are currently three mainstream error analysis methods: numerical simulation, statistical theory-based methods, and interval arithmetic-based methods. Among them, the interval arithmetic-based method controls the error within a limited interval, and the upper and lower bounds of the radiation pattern distortion can be obtained only through limited interval operations. This type of research is also called tolerance analysis. As a powerful mathematical analysis tool, tolerance analysis takes into account the random errors in the antenna array and effectively expands the limited data set. Therefore, combined with tolerance analysis, the data set required by the expanded parameter model method can quickly diagnose arrays with random errors. Summary of the invention

[0004] The present invention proposes an antenna array fault diagnosis method based on a deep neural network extended by data tolerance analysis, which is suitable for antenna arrays with certain random errors. The existing random errors will reduce the accuracy of array fault diagnosis using the existing parameter model method. In order to improve the diagnostic accuracy, an error model is established, and tolerance analysis is used to obtain the upper and lower bounds of the power (radiation) pattern distortion through interval operations. The data is preprocessed according to the obtained upper and lower bounds to obtain reliable and accurate data intervals, and a batch of data sets processed by tolerance analysis are obtained. The model training is completed in the training stage before fault diagnosis, and the location and number of faulty array elements are determined based on the data set expanded by tolerance analysis. Compared with the traditional non-parametric model algorithm, it has better diagnostic effect and faster diagnostic speed for error conditions; compared with the algorithm of the parameter model method, it requires fewer measurement samples, pre-processes the error conditions, has a wider data set, and has more reliable diagnostic results, which is suitable for array fault diagnosis in error conditions.

[0005] The scheme of the present invention is:

[0006] An antenna array fault diagnosis method based on data tolerance analysis extension, such as Figure 1 As shown, the following steps are included:

[0007] S1. Data processing: Measure far-field radiation data in multiple directions in the far-field area to generate power information. Use tolerance analysis to pre-process the measured power pattern to obtain the upper and lower bounds of the power pattern considering the array element position error, thereby expanding a single data set into multiple data sets to obtain an extended data set.

[0008] The specific method for obtaining the upper and lower bounds of the power pattern through tolerance analysis is as follows: determine the position interval of each array element, obtain the upper and lower bounds of the power pattern through tolerance analysis, wherein the power pattern is composed of the real part interval and the imaginary part interval of the far-field radiation, and obtain the upper and lower bounds of the real part interval and the imaginary part interval of the far-field radiation by calculating the upper and lower bounds of the phase interval, and then scale the coupling of the real and imaginary parts of the far-field radiation data, and finally obtain the interval boundary of the power pattern when there is a position error;

[0009] S2, training a fault diagnosis model, the input of the fault diagnosis model is the extended data set, and the output is the position and number of failed array elements;

[0010] S3, measure the far-field data at the sampling points in the same direction as in S1, input the data into the trained neural network, confirm the position of the array element that is most likely to fail, and then complete the radiation to gradually determine the positions of all possible failed array elements.

[0011] Specifically, in S1, N groups of measurements are defined in different scenes, and each group measures radiation data in K different directions in the far field area. The far field radiation formula can be expressed as:

[0012]

[0013] Where f is the operating frequency, c is the speed of light, M is the total number of array elements, and d is the speed of light. i is the position of the ith array element, x i is the excitation of the ith array element, θ k It is the angle between the observation angle in the kth direction and the positive z axis of the array reference coordinate system.

[0014] Given that the operating frequency f and the array element position d have similar effects in the far-field radiation formula, the following example uses the case where there is an error in the array position to illustrate this method. Define the actual array position of the i-th array element as d i , define the interval of the position of the i-th array element as (·) I Represents interval, position vector and are the lower and upper boundaries of the i-th array element position interval, L is the lower boundary, U is the upper boundary, is the midpoint of the i-th array element position interval, represents the maximum offset of the position of the ith array element under the influence of the error. Therefore, the maximum offset of each array element can form a maximum offset vector △d = [△d 1 ,△d 2 ,...,△d M ] T , determine the maximum offset of each array element and the fixed interval size.

[0015] Create an interval function for far-field radiation:

[0016]

[0017] in, and represent the real and imaginary intervals of far-field radiation, respectively. is the phase interval,

[0018] The power pattern of the antenna array is defined as:

[0019]

[0020] in, and represent the real and imaginary intervals of far-field radiation, respectively. is the phase interval, f is the operating frequency, c is the speed of light, is the interval of the position of the ith array element, and are the lower and upper boundaries of the i-th array element position interval, respectively, and x i is the excitation of the ith array element, θ k is the angle between the observation angle in the kth direction and the positive z axis of the array reference coordinate system;

[0021] Solving for P by Tolerance Analysis I (θ k ) L (θ k ) and the upper bound P U (θ k ), specifically:

[0022] First calculate the phase interval The upper bound of With the lower bound

[0023] Recalculate the interval function

[0024]

[0025]

[0026] Summing all upper and lower bounds yields and Upper and lower bounds of an interval:

[0027]

[0028]

[0029] The real and imaginary parts of the far-field radiation data are coupled and scaled to obtain the interval boundary of the power pattern when there is a position error:

[0030]

[0031]

[0032] Specifically, step S2 is as follows:

[0033] Define the vector of power data in the nth group of scenarios as P n After obtaining the upper and lower bounds of the interval, the single power pattern is expanded to S according to the set step size, and the vector representation is P n,s ∈{P n,1 ,P n,2 ,...,P n,S}, the array excitation vector composed of the nth group of scenes is M is the total number of array elements, and the failure indication vector is constructed using array excitation. when but otherwise, The label r of the number of failed array elements n , r n Calibration g n The number of 1s;

[0034] Combine the power data with the labels of the number of failed array elements to obtain the training set of the number of failed array elements:

[0035] {(P 1,1 ,r 1 ),...,(P 1,S ,r 1 ),(P 2,1 ,r 2 ),...,(P 2,S ,r 2 ),...,(P N,1 ,r N ),...,(P N,S ,r N )}

[0036] Train a neural network f that correlates power data and the number of failed elements 1 , whose mapping relationship is a single-label neural network f 1 :r n =f 1 (P n,s ), neural network f 1 Use multi-classification logarithmic loss function;

[0037] The power data is combined with the array excitation vector to obtain a training set of array element failure locations:

[0038] {(P 1,1 ,g 1 ),...,(P 1,S ,g 1 ),(P 2,1 ,g 2 ),...,(P 2,S ,g 2 ),...,(P N,1 ,g N ),...,(P N,S ,g N )}

[0039] Train a neural network f to learn the relationship between the power data with errors and the locations of the failed elements 2 , whose mapping relationship is the multi-label neural network f 2 :g n =f 2 (P n,s ), neural network f 2Use the binary classification logarithmic loss function.

[0040] Specifically, S3 is as follows: the measured data obtained is defined as P tested =[P(θ 1 ),P(θ 2 ),...,P(θ K )] T , K is the measurement point, which is input into the single-label neural network f for determining the number of failed array elements. 1 The predicted number of failed array elements is obtained from

[0041] P tested Input to the neural network f used to determine the location of the failed array element 2 In the above, the element failure probability vector p is determined as [p 1 ,p 2 ,...,p M ], where p i , i∈[1,M] represents the probability of failure of the ith array element; sort the obtained array element failure probabilities, and take the maximum and second largest failure probabilities as p u and p v , and considered that one of the array elements failed;

[0042] Iterate the following steps until

[0043] Using p u and p v , the uth and vth array elements are compensated, and the kth far-field radiation data after compensation are:

[0044]

[0045]

[0046] set up After compensation and The power data obtained replaces P tested Input to the neural network f 2 We get the new probability vector and Compare [u] and p [v] The maximum value in p [u] In the above example, the uth array element is determined to be invalid; otherwise, the vth array element is determined to be invalid and u and v are updated.

[0047] The beneficial effects of the present invention are:

[0048] Compared with traditional methods, the present invention has a faster diagnosis speed in the application stage and a higher accuracy rate in the presence of errors; compared with existing parameter model-based methods, the extended data set of the present invention improves the reliability and versatility of diagnosis, has a higher accuracy rate, and the use of compensation methods also effectively reduces the sample data that needs to be collected in the training stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flowchart of the present invention.

[0050] Figure 2 This is the diagnosis accuracy result of the failed array element position.

[0051] Figure 3 Diagnosis results of faulty array element position under different position errors. DETAILED DESCRIPTION

[0052] The present invention is further explained and illustrated below in conjunction with the accompanying drawings and embodiments.

[0053] The technical solution of the present invention is: in the data processing stage, the far-field radiation data in multiple directions are measured in the far-field area to generate power information, and the measured power pattern is preprocessed using tolerance analysis to obtain the upper and lower bounds of the power pattern considering the array element position error, thereby expanding the data set. In the training stage, the training of the fault diagnosis model is completed, the expanded data set is used as the input of the neural network, the position and number of failed array elements are set as output, and two neural networks are trained. In the application stage, the far-field data at the sampling points in the same direction as the training stage are measured, and the data is input into the trained neural network to confirm the position of the array element that is most likely to fail, and then the radiation is completed to gradually determine the positions of all possible failed array elements. Therefore, the technical solution of the present invention is an antenna array fault diagnosis method based on a deep neural network extended by data tolerance analysis.

[0054] Example

[0055] The fault array in this example is a uniform linear array with 36 elements, and the element spacing is d = 0.5λ, where λ is the wavelength. Assume that the number of failed elements in the array is at most 4. The processing process of this example is:

[0056] Data processing stage:

[0057] Step 1: Within the range of pitch angle θ∈[-89°, 89°], measure the amplitude and phase of the far-field radiation field of the array under test at intervals of 2°, and a total of 90 groups of amplitude and phase are considered as a group. i =1, the excitation of the failed array element is x i= 0. Under the conditions of determining the number of array element failures, 800 sets of far-field radiation data under random failure scenarios at different array element positions were measured to generate corresponding power data.

[0058] Step 2: Determine the maximum offset vector of each array element position under the influence of error △d = [△d 1 ,△d 2 ,...,△d M ] T According to engineering practice, the maximum error ratio between each array element and the theoretical value is 3%, that is,

[0059] Step 3: Scale the coupling of the real and imaginary parts of the far-field radiation to obtain the interval boundary of the power pattern when there is a position error, which is:

[0060]

[0061]

[0062] According to the upper and lower bounds of the power pattern obtained, the step size of the error ratio is taken as 0.1%, and the single power pattern data is expanded to 30. The 800 sets of power data under random failure scenarios of different array element positions obtained in step 1 are expanded to obtain 24,000 sets of power data under random failure scenarios of different position errors and different array element positions.

[0063] Training phase:

[0064] Step 4: Group the data expanded in step 3 according to the number of array element failures to obtain 8 training sets, each containing 3000 training samples. n and the number of failed array elements label r n Power data P n Do classification.

[0065] Step 5: Combine the samples in step 4 and rearrange them to obtain Then divided into the training set of the number of array element failures and training set of array element failure positions Input single label neural network f for determining the number of failed array elements 1 and the neural network f used to determine the location of the failed array element 2 . Neural Network f 1 and f 2 The hyperparameters used are shown in Table 1.

[0066] Model application phase:

[0067] Step 6: Within the range of pitch angle θ∈[-89°, 89°], measure the amplitude and phase of the far-field radiation field of the array to be tested at intervals of 2°, collect far-field radiation data at 90 measurement points, and obtain the power data P to be tested. tested , the power data to be measured P tested Input to the trained neural network f 1 The predicted number of failed array elements is obtained

[0068] Step 7: The power data to be measured P tested Input to the trained neural network f 2 In the above, the element failure probability vector p is determined as [p 1 ,p 2 ,...,p 36 ], where p i , i∈[1,36] represents the probability of failure of the ith array element. Sort the obtained array element failure probabilities and take the maximum and second largest failure probabilities as p u and p v , and it is considered that one of the array elements has failed.

[0069] Step 8: Using p u and p v , the uth and vth array elements are compensated. The radiation in the kth measurement direction after compensation is:

[0070]

[0071]

[0072] Step 9: Setup After compensation and The power data obtained replaces the original measured data and is input into the neural network f 2 In the above equation, we get the new probability vector p [u] =[p [u] 1 ,p [u] 2 ,...,p [u] L ] and p [v] =[p [v] 1 ,p [v] 2 ,...,p [v] L ], compare p [u] and p [v] The maximum value in p [u]In the above example, the uth array element is considered invalid; otherwise, the vth array element is considered invalid. Update u, v.

[0073] Step 10: Repeat steps 8-9 until

[0074] The present invention proposes an antenna array fault diagnosis method based on a deep neural network extended by data tolerance analysis for a uniform array, which contains 36 array elements with an array element spacing of 0.5λ, where λ is the wavelength. The tolerance analysis method is used to generate 24,000 array samples, of which 16,800 samples are used to train the neural network and 7,200 are used for diagnostic testing. In the far field area, sampling is performed at measurement points with an interval of 2° within the range of pitch angles θ∈[-89°,89°], with a total of 90 measurement field points. The maximum position error is 0.015λ. The results of the diagnosis accuracy of the failed array element position are as follows: Figure 2 As shown in the figure, the diagnosis results of the faulty array element position under different position errors are as follows Figure 3 shown.

[0075] The present invention proposes an antenna array fault diagnosis method based on a deep neural network extended by data tolerance analysis. In the data processing stage, the far-field radiation data in multiple directions are measured in the far-field area to generate power information, and the measured power pattern is preprocessed using tolerance analysis to obtain the upper and lower bounds of the power pattern considering the array element position error, thereby expanding the data set. In the training stage, the training of the fault diagnosis model is completed, the expanded data set is used as the input of the neural network, the position and number of failed array elements are set as output, and two neural networks are trained. In the application stage, data is collected in the same direction as the data measured in the training stage, and the obtained data to be tested is input into the model to preliminarily determine the number of failed array elements and the probability of array element failure, and then the data compensation is performed on the test results according to the radiation compensation formula to determine the position of the failed array elements in turn.

Claims

1. An antenna array fault diagnosis method based on data tolerance analysis extension, It is characterized in that The following steps are involved: S1. Data processing: Measure far-field radiation data in multiple directions in the far-field area to generate power information. Use tolerance analysis to pre-process the measured power pattern to obtain the upper and lower bounds of the power pattern considering the array element position error, thereby expanding a single data set into multiple data sets to obtain an extended data set. The specific method for obtaining the upper and lower bounds of the power pattern through tolerance analysis is as follows: determine the position interval of each array element, and obtain the upper and lower bounds of the power pattern through tolerance analysis. The power pattern is composed of the real and imaginary intervals of the far-field radiation. First, calculate the upper and lower bounds of the phase interval, then calculate the interval function, and sum all the upper and lower bounds to obtain and The upper and lower bounds of the interval, and Respectively represent the real and imaginary intervals of far-field radiation, θ k is the angle between the observation angle in the kth direction and the positive z axis of the array reference coordinate system; then, by scaling the coupling between the real and imaginary parts of the far-field radiation data, the interval boundary of the power pattern when there is a position error is finally obtained; S2, training a fault diagnosis model, the input of the fault diagnosis model is the extended data set, and the output is the position and number of failed array elements; S3, measure the far-field data at the sampling point in the same direction as S1, input the data into the trained neural network, confirm the position of the array element that is most likely to fail, and then compensate for the radiation to gradually determine the positions of all possible failed array elements. Specifically, define the measured data as P tested =[P(θ 1 ),P(θ 2 ),...,P(θ K )] T , K is the measurement point, which is input into the single-label neural network f for determining the number of failed array elements. 1 The predicted number of failed array elements is obtained from f 1 It is a neural network between the correlation power data and the number of failed array elements; Input P tested into the neural network f for determining the position of the failed element 2 to determine the element failure probability vector p = [p 1 , p 2 ,..., p M , where f 2 is a neural network for learning the relationship between the power data with errors and the position of the failed element, and p i , i ∈ [1, M] represents the probability of the i-th element failing; sort the obtained element failure probabilities, and mark the maximum and the second maximum failure probabilities as p u and p v , and assume that one of the elements has failed; Iterate the following steps until Using p u and p v , the uth and vth array elements are compensated, and the kth far-field radiation data after compensation are: Where f is the operating frequency, c is the speed of light, and x u is the excitation of the u-th array element, x v is the excitation of the vth array element, d u is the actual position of the uth array element, d v is the actual position of the vth array element; set up After compensation and The power data obtained replaces P tested Input to the neural network f 2 We get the new probability vector and Compare [u] and p [v] The maximum value in p [u] In the above example, the uth array element is determined to be invalid; otherwise, the vth array element is determined to be invalid and u and v are updated.

2. According to claim 1, a method for antenna array fault diagnosis based on data tolerance analysis expansion, It is characterized in that In S1, the power pattern of the antenna array is defined as: in, is the phase interval, f is the operating frequency, c is the speed of light, is the interval of the position of the ith array element, and are the lower and upper boundaries of the i-th array element position interval, L is the lower boundary, U is the upper boundary, and x i is the excitation of the ith array element; Solving for P by Tolerance Analysis I (θ k ) L (θ k ) and the upper bound P U (θ k ), specifically: First calculate the phase interval The upper bound of With the lower bound Recalculate the interval function Summing all upper and lower bounds yields and Upper and lower bounds of an interval: The real and imaginary parts of the far-field radiation data are coupled and scaled to obtain the interval boundary of the power pattern when there is a position error:

3. According to claim 2, a method for antenna array fault diagnosis based on data tolerance analysis expansion, It is characterized in that Step S2 is specifically as follows: Define the vector of power data in the nth group of scenarios as P n After step S1 is expanded to S, the vector is represented as P n,s ∈{P n,1 ,P n,2 ,...,P n,S }, the array excitation vector composed of the nth group of scenes is M is the total number of array elements, and the failure indication vector is constructed using array excitation. when but otherwise, The label r of the number of failed array elements n , r n Calibration g n The number of 1s; Combine the power data with the labels of the number of failed array elements to obtain the training set of the number of failed array elements: {(P 1,1 ,r 1 ),...,(P 1,S ,r 1 ),(P 2,1 ,r 2 ),...,(P 2,S ,r 2 ),...,(P N,1 ,r N ),...,(P N,S ,r N )} Train a neural network f that correlates power data and the number of failed elements 1 , whose mapping relationship is a single-label neural network f 1 :r n =f 1 (P n,s ), neural network f 1 Use multi-classification logarithmic loss function; The power data is combined with the array excitation vector to obtain a training set of array element failure locations: {(P 1,1 ,g 1 ),...,(P 1,S ,g 1 ),(P 2,1 ,g 2 ),...,(P 2,S ,g 2 ),...,(P N,1 ,g N ),...,(P N,S ,g N )} Train a neural network f for learning the relationship between power data with errors and the positions of failed array elements 2 , and its mapping relationship is a multi-label neural network f 2 : g n = f 2 (P n,s ), and the neural network f 2 uses the binary classification logarithmic loss function.

Images