Radio frequency circuit board fault diagnosis method based on data driving

Abstract

The invention discloses a radio frequency circuit board fault diagnosis method based on data driving, and the method comprises the steps: presetting a fault, inputting an excitation signal, obtaining a plurality of groups of time domain data, carrying out the dual filtering of the time domain data, obtaining frequency domain data, carrying out the feature extraction of the frequency domain data, obtaining a fault feature of a fault type, and carrying out the fault diagnosis. Then, for all fault types, the fault types serve as labels, the corresponding fault features serve as training samples and are input into a classifier, the classifier is trained until convergence is achieved, finally, for a radio frequency circuit board, the same data processing is adopted, the obtained fault features are sent into the trained classifier which is trained to be converged, and the fault types are obtained. And completing fault diagnosis. According to the method, extra internal measuring points do not need to be added, fault diagnosis is carried out only through input and output ports of the radio frequency circuit board, excitation is applied to the radio frequency circuit board, radio frequency output signal data are collected, the data are analyzed through a machine learning model, and therefore the fault type of the radio frequency circuit board is accurately recognized.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and more specifically, relates to a data-driven radio frequency circuit board fault diagnosis method, which does not require additional measurement points, but instead utilizes the input and output ports of the radio frequency circuit board itself for fault diagnosis. Background Art

[0002] RF circuit boards are widely used in communications, radar, medical equipment, automotive electronics and other fields. Their internal structure mainly includes key components such as RF amplifiers, π-type attenuators and filters. Due to the high operating frequency, high power and complex electromagnetic field environment of RF circuit boards, they are more prone to component aging, damage, nonlinear and uncertainty failures than traditional analog circuits. If these faults are not isolated and located in time, they may cause the overall failure of the RF circuit board. Therefore, it is particularly important to establish fault diagnosis for RF circuit boards.

[0003] Current RF circuit board fault diagnosis methods usually involve disassembling the circuit board to expose the internal structure and directly measuring the voltage, current, and waveform information at both ends of the components to evaluate the aging or damage of the components. However, due to the high-frequency characteristics of RF circuits, adding measurement points directly inside the RF circuit board may affect the impedance and electromagnetic field of the circuit. In addition, this method is not suitable for large-scale fault diagnosis and requires high professional skills from technicians.

[0004] Therefore, there is an urgent need to study a non-invasive RF circuit board fault diagnosis method that does not require adding measurement points inside the RF circuit board. It can use the input and output ports of the RF circuit board itself to achieve automatic fault diagnosis without the need for professional operation. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a data-driven RF circuit board fault diagnosis method for batch automated fault detection of RF circuit boards, without adding additional internal measurement points, and only through the input and output ports of the RF circuit board itself to perform fault diagnosis. Specifically, by applying excitation to the RF circuit board and collecting RF output signal data, the data are analyzed using a machine learning model, so as to accurately identify the fault type of the RF circuit board.

[0006] To achieve the above-mentioned object of the invention, the present invention is based on a data-driven radio frequency circuit board fault diagnosis method, which is characterized by comprising the following steps:

[0007] (1) Model the RF circuit board, select important components for monitoring, and introduce a fault model on the RF circuit board. According to the fault model, a predetermined fault is introduced on the RF circuit board, that is, a fault type is selected;

[0008] (2) For the selected fault type, select a swept-frequency signal as the excitation signal f in (t) and inject it into the input end of the RF circuit board, and collect the time-domain signal f out (t) from the output end of the RF circuit board to obtain the time-domain data f(n). Repeat the above process to obtain multiple sets of time-domain data f(n), thereby forming a systematic data set, where t is time and n is the sampling point number sequence;

[0009] (3) Perform double filtering on a set of collected time-domain data f(n)

[0010] 3.1) Construct a pure-noise time-domain signal s extended (n) with the same length as the time-domain data f(n)

[0011] 3.1.1) Extract the envelope signal from the time-domain data f(n)

[0012] a) Calculate the discrete Fourier transform to obtain the frequency-domain data F(k):

[0013]

[0014] where N is the length of the discrete Fourier transform and k represents the frequency point;

[0015] b) Calculate the frequency response of the Hilbert transform

[0016]

[0017] H(k) is the transformation matrix;

[0018] c) Calculate the signal after the Hilbert transform

[0019]

[0020] d) Calculate the envelope A(n):

[0021]

[0022] 3.1.2) Extract the pure-noise data s(n)

[0023] Determine a threshold a. For the data points in the time-domain data f(n) that satisfy A(n) < a, they are determined as pure-noise data. In this way, the pure-noise data s(n) is extracted, where n1 < n < n2, 0 < n1 < n2 < n0, and n0 is the number of sampling points of the time-domain data f(n);

[0024] 3.1.3) Generate a pure-noise time-domain signal s extended (n)

[0025] The pure noise data s(n) is periodically extended to generate a pure noise time domain signal s of equal length extended (n):

[0026] s extended (n) = s n ((nmodT)+n1)

[0027] Where n = 0, 1, 2, ... n0-1;

[0028] 3.2) Adaptive filtering

[0029] a) For pure noise time domain signal s extended (n) Filter to obtain noise data

[0030]

[0031] Where w(n) is the coefficient vector of the adaptive filter, w(n) = [w0(n), w1(n), …, w M-1 (n)] T , M is the order of the adaptive filter, w m (n), m = 0, 1, 2, ... M-1 are the coefficients of the adaptive filter, s extended (n)=[s extended (n),s extended (n-1),…,s extended (n-M+1)] T ;

[0032] b) Calculate the error signal e(n):

[0033]

[0034] c) Using the least mean square error criterion (LMS) algorithm, the mean square value of the error signal is minimized by iteratively adjusting the adaptive filter coefficients:

[0035] J(n)=E[e 2 (n)]

[0036] Where E[·] represents the expected value, and the adaptive filter coefficients are iteratively adjusted as:

[0037] w(n+1)=w(n)+μs extended (n)e(n)

[0038] Where μ is the step size factor, which controls the rate and stability of updating the adaptive filter coefficients;

[0039] d) Use the error signal e(n) as the time domain data after adaptive filtering

[0040] 3.3) Fourier Transform

[0041] For the adaptively filtered time domain data Perform Fourier transform to obtain frequency domain data

[0042]

[0043] Where N is the length of the discrete Fourier transform, and k represents the frequency point;

[0044] 3.4) Sliding maximum filtering

[0045] For frequency domain data Perform sliding maximum filtering. For each frequency domain point k, the sliding window contains W frequency domain samples. The output of the maximum filtering is the frequency domain data is the maximum value within the window:

[0046]

[0047] The value range of k needs to ensure that km is within the effective frequency range, that is, 0≤km <N;

[0048] (4) Frequency domain data Perform feature extraction,

[0049] 4.1) Extract key indicators describing frequency domain waveforms as sample features

[0050] 4.2) Extract features that characterize the differences between different fault types, that is, extract inter-sample features;

[0051] 4.2.1) Build a "base spectrum diagram" for each fault type

[0052] a) Calculate the mean and residual

[0053] First, randomly select L frequency domain data after secondary filtering from the fault type. The mean of a single frequency domain data is recorded as x i , i = 1, 2, ..., L, the mean of the L frequency domain data means is recorded as Then the mean value of the i-th frequency domain data is x i The residual v i :

[0054]

[0055] b) Calculate the standard deviation σ using the Bessel formula x:

[0056]

[0057] c) Use the Wright test method to eliminate gross errors in the frequency domain data. That is, if the absolute value of the residual of a frequency domain data is greater than 3 times the standard deviation, that is, |v i |>3σ x , the frequency domain data is a gross error and is removed, and the process returns to step a) to reselect, calculate and remove, until L frequency domain data are randomly selected and filtered twice without gross errors;

[0058] d) The mean of the L frequency domain data that do not contain gross errors As the "basic spectrum diagram" of this fault type;

[0059] 4.2.2) Extract the Pearson correlation coefficient: For the “base spectrum” of the cth fault type, it is recorded as And a "basic spectrum diagram" with all fault types Calculate the Pearson correlation coefficient and get C related features r cj ,j=1,2,…,C:

[0060]

[0061] Among them, x ci is the i-th frequency domain data of the c-th fault type that does not contain gross errors, x ji is the jth piece of frequency domain data among the L pieces of frequency domain data that do not contain gross errors and correspond to the jth fault type;

[0062] C relevant features r cj That is, the inter-sample characteristic of the cth fault type;

[0063] (5) The sample features of the c-th fault type are concatenated with the inter-sample features to obtain the fault features of the c-th fault type. For all fault types, the fault type is used as a label and the corresponding fault features are used as training samples. The fault features are input into the classifier and the classifier is trained until convergence.

[0064] (6) Fault diagnosis

[0065] For an RF circuit board, select a frequency sweep signal as the excitation signal f in (t) is injected into the input end of the RF circuit board, and the time domain signal f is collected from the output end of the RF circuit board out(t), and obtain the time domain data f(n). Repeat the above process to obtain multiple sets of time domain data f(n), thereby forming a systematic data set. Then, according to the methods of steps (3), 4.1), and 4.2.1), the sample features and the "base spectrum diagram" are obtained, which are compared with the "base spectrum diagram" of all fault types. The Pearson correlation coefficient is calculated to obtain C related features, and then the sample features are concatenated with the inter-sample features to obtain the fault features, which are sent to the trained classifier in step (5) to obtain the fault type and complete the fault diagnosis.

[0066] The object of the invention of the present invention is achieved in this way:

[0067] The data-driven RF circuit board fault diagnosis method of the present invention obtains multiple groups of time domain data by predetermining the fault and inputting the excitation signal, then performs double filtering on the time domain data to obtain frequency domain data, then performs feature extraction on the frequency domain data to obtain a fault feature of a fault type, and then for all fault types, the fault type is used as a label, and the corresponding fault feature is used as a training sample, which is input into the classifier, and the classifier is trained until convergence. Finally, for a piece of RF circuit board, the same data processing is used to obtain its fault feature, which is sent to the trained classifier, i.e., the trained converged classifier, to obtain the fault type and complete the fault diagnosis. The present invention does not need to add additional internal measurement points, and only performs fault diagnosis through the input and output ports of the RF circuit board itself. By applying excitation to the RF circuit board and collecting RF output signal data, these data are analyzed using a machine learning model, thereby accurately identifying the fault type of the RF circuit board. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flow chart of a specific implementation method of the radio frequency circuit board fault diagnosis method based on data drive of the present invention;

[0069] Figure 2 It is a schematic diagram of the principle of adaptive filtering in the present invention;

[0070] Figure 3 This is a specific example of double filtering processing in the present invention, wherein the first and second figures are time domain figures and frequency domain figures respectively; the third and fourth figures are time domain figures and frequency domain figures after the first filtering respectively, and the fifth figure is a spectrum figure after sliding maximum filtering. DETAILED DESCRIPTION

[0071] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0072] Figure 1 It is a flow chart of a specific implementation of the data-driven radio frequency circuit board fault diagnosis method of the present invention.

[0073] In this embodiment, if Figure 1 As shown, the data-driven RF circuit board fault diagnosis method of the present invention includes the following steps:

[0074] Step S1: ADS modeling and fault type selection

[0075] Model the RF circuit board, select important components for monitoring, and introduce a fault model on the RF circuit board. According to the fault model, introduce a predetermined fault on the RF circuit board, that is, select a fault type.

[0076] In this embodiment, the RF faulty circuit board is modeled by using ADS (Advanced Design System).

[0077] Step S2: inject the excitation signal and collect multiple sets of time domain data f(n)

[0078] For the selected fault type, select the swept frequency signal as the excitation signal f in (t) is injected into the input end of the RF circuit board, and the time domain signal f is collected from the output end of the RF circuit board out (t), and obtain the time domain data f(n). Repeat the above process to obtain multiple sets of time domain data f(n), thereby forming a systematic data set, where t is time and n is the number of sampling points;

[0079] The entire fault diagnosis system consists of three main parts: the host computer, the RF transceiver terminal platform and the RF fault board (the RF circuit board after the predetermined fault is introduced). The host computer is responsible for controlling the RF transceiver terminal platform to generate the (RF) excitation signal f in (t) and injects it into the RF fault board. The RF signal output by the RF fault board is then collected and transmitted to the host computer for data display and fault diagnosis.

[0080] The host computer controls the RF terminal board through the serial port to inject an excitation signal into the RF fault board. In this embodiment, a sweep frequency signal of 1.68 to 1.69 GHz is used as the excitation signal f in (t). By adjusting the resistance value and plugging and unplugging the plug cap, the fault modes such as short circuit, open circuit and parameter drift of the main components (such as amplifiers and attenuators) in the RF circuit board are simulated. Only the original time domain signal is collected from the output end of the RF circuit board, and the fault type number F0~F13 is marked.

[0081] Step S3: Perform double filtering on a set of collected time-domain data f(n)

[0082] Perform double filtering on the collected time-domain data. First, perform preliminary filtering on the time-domain data in the time domain; subsequently, convert the data to the frequency domain through Fourier transform and then perform secondary filtering on the frequency-domain data. Compared with single filtering, this "double filtering" method can more flexibly select appropriate filtering schemes in the time domain and frequency domain, thereby effectively reducing the noise amplitude of the frequency-domain waveform.

[0083] The number of collected points is n0, and the time-domain data is represented as f(n), where 0 < n < n0 and contains multiple cycles. The information of these data is mainly divided into two categories: one is pure noise data s(n), where n1 < n < n2, 0 < n1 < n2 < n0, and the other contains both noise and valid data. Therefore, first, a pure noise time-domain signal s(n), where 0 < n < n0 - 1, with the same length as the original data needs to be constructed. The specific steps are as follows.

[0084] Step S3.1: Construct a pure noise time-domain signal s with the same length as the time-domain data f(n) extended (n)

[0085] Step S3.1.1: Extract the envelope signal from the time-domain data f(n)

[0086] a) Calculate the discrete Fourier transform to obtain the frequency-domain data F(k):

[0087]

[0088] where N is the length of the discrete Fourier transform and k represents the frequency point;

[0089] b) Calculate the frequency response of the Hilbert transform

[0090]

[0091] H(k) is the transformation matrix;

[0092] c) Calculate the signal after the Hilbert transform

[0093]

[0094] d) Calculate the envelope A(n):

[0095]

[0096] Step S3.1.2: Extract the pure noise data s(n)

[0097] Determine a threshold value \(a\). For the data points in the time-domain data \(f(n)\) that satisfy \(A(n)\lt a\), they are determined as pure noise data, and thus the pure noise data \(s(n)\) is extracted, where \(n_1\lt n\lt n_2\), \(0\lt n_1\lt n_2\lt n_0\), and \(n_0\) is the number of sampling points of the time-domain data \(f(n)\).

[0098] Step S3.1.3: Generate a pure noise time-domain signal \(s\) of equal length extended (n)

[0099] Periodically extend the pure noise data \(s(n)\) to generate a pure noise time-domain signal \(s\) of equal length extended (n):

[0100] s extended (n)=s n ((n\(\bmod\)T)+n1)

[0101] where \(n = 0,1,2,\cdots n_0 - 1\).

[0102] Step S3.2: Adaptive filtering

[0103] In this embodiment, as Figure 2 shown, the adaptive filtering process is as follows:

[0104] a), Filter the pure noise time-domain signal \(s\) extended (n) to obtain the noise data

[0105]

[0106] where \(w(n)\) is the coefficient vector of the adaptive filter, \(w(n)=[w_0(n),w_1(n),\cdots,w\) M-1 (n)] T , \(M\) is the order of the adaptive filter, \(w\) m (n), \(m = 0,1,2,\cdots M - 1\) are the coefficients of the adaptive filter, \(s\) extended (n)=[s extended (n),s extended (n - 1),\cdots,s extended (n - M + 1)] T ;

[0107] b), Calculate the error signal \(e(n)\):

[0108]

[0109] c), Adopt the least mean square error criterion (LMS) algorithm to minimize the mean square value of the error signal by iteratively adjusting the coefficients of the adaptive filter:

[0110] J(n)=E[e2 (n)]

[0111] Where E[·] represents the expected value, and the adaptive filter coefficients are iteratively adjusted as:

[0112] w(n+1)=w(n)+μs extended (n)e(n)

[0113] Where μ is the step size factor, which controls the rate and stability of updating the adaptive filter coefficients;

[0114] d) Use the error signal e(n) as the time domain data after adaptive filtering

[0115] Step S3.3: Fourier transform

[0116] For the adaptively filtered time domain data Perform Fourier transform to obtain frequency domain data

[0117]

[0118] Where N is the length of the discrete Fourier transform, and k represents the frequency point;

[0119] Step S3.4: Sliding maximum filtering

[0120] In the frequency domain, the sliding maximum filtering method is used for secondary filtering, which makes the time domain waveform smoother while retaining the peak characteristics in the frequency domain.

[0121] For frequency domain data Perform sliding maximum filtering. For each frequency domain point k, the sliding window contains W frequency domain samples. The output of the maximum filtering is the frequency domain data is the maximum value within the window:

[0122]

[0123] The value range of k needs to ensure that km is within the effective frequency range, that is, 0≤km <N。

[0124] In this embodiment, Figure 3 The figure shows the "double filtering" processing process of the time domain data collected for fault F0, where the first and second charts show the original time domain graph and frequency domain graph respectively; the third and fourth charts show the time domain graph and frequency domain graph after the first filtering; the fifth chart shows the spectrum graph after sliding maximum value filtering.

[0125] Step S4: Frequency domain data Perform feature extraction,

[0126] Step S4.1: Extract key indicators describing the frequency domain waveform as sample features

[0127] In this embodiment, the key indicators describing the sample frequency domain waveform include maximum value, range, mean, variance, rising edge width, falling edge width, bandwidth, peak factor, margin, shape factor, etc., where:

[0128] a) The rising edge width is the difference in the horizontal axis (n) corresponding to the signal changing from the extreme value (20%) to (80%).

[0129] [Rise Time Width=|n 20 -n 80 |

[0130] where n 80 The horizontal axis represents the signal reaching the extreme value (80%), n 20 The horizontal axis represents the extreme value of the signal.

[0131] b) The falling edge width is the difference in the horizontal axis corresponding to the signal changing from the extreme value (80%) to (20%).

[0132] [Fall Time Width=|n 80 -n 20 |

[0133] where n 80 The horizontal axis represents the signal reaching the extreme value (80%), n 20 The horizontal axis represents the extreme value (20%) reached by the signal (x(n)).

[0134] c) Width is the difference in the horizontal axis corresponding to the signal going from the extreme value (50%) to the next extreme value (50%).

[0135] [Bandwidth=|n1 50 -n2 50 |

[0136] Where n1 50 The horizontal axis represents the signal reaching the extreme value (50%), n2 50 The horizontal axis represents the next extreme value (50%) reached by the signal.

[0137] d) Peak Factor

[0138]

[0139] Where (RMS) is the root mean square value of the signal:

[0140] e) Waveform Factor

[0141]

[0142] f) Crest Factor

[0143]

[0144] Step S4.2: extracting features that characterize the differences between different fault types, i.e. extracting inter-sample features;

[0145] Step S4.2.1: Construct a “basic spectrum diagram” for each fault type

[0146] a) First, randomly select L frequency domain data after secondary filtering from the fault type. The mean of a single frequency domain data is recorded as x i , i = 1, 2, ..., L, the mean of the L frequency domain data means is recorded as Then the mean value of the i-th frequency domain data is x i The residual v i :

[0147]

[0148] b) Calculate the standard deviation σ using the Bessel formula x :

[0149]

[0150] c) Use the Wright test method to eliminate gross errors in the frequency domain data. That is, if the absolute value of the residual of a frequency domain data is greater than 3 times the standard deviation, that is, |v i |>3σ x , the frequency domain data is a gross error and is removed, and the process returns to step a) to reselect, calculate and remove, until L frequency domain data are randomly selected and filtered twice without gross errors;

[0151] d) The mean of the L frequency domain data that do not contain gross errors Serves as the “basic spectrum diagram” for this fault type.

[0152] In this embodiment, L=10, taking fault type F0 as an example, the specific results are shown in Table 1:

[0153]

[0154]

[0155] Table 1

[0156] Step S4.2.2: Extract the Pearson correlation coefficient: For the “base spectrum” of the cth fault type, it is recorded as And a "basic spectrum diagram" with all fault types Calculate the Pearson correlation coefficient and get C related features r cj ,j=1,2,…,C:

[0157]

[0158] Among them, x ci is the L corresponding to the cth fault type without gross errors ′ The i-th frequency domain data, x ji is the L corresponding to the jth fault type without gross errors ′ j pieces of frequency domain data.

[0159] C relevant features r cj That is, the inter-sample characteristic of the cth fault type.

[0160] In this embodiment, when extracting features, specific features of fault F0 are selected for detailed description, as shown in Table 2.

[0161]

[0162]

[0163] Table 2

[0164] In this embodiment, as shown in Table 1, the fault type includes 38 sample features feature1-38 and 14 inter-sample features, namely Pearson correlation coefficients pearson1-14.

[0165] Step S5: Training the classifier

[0166] The sample features of the c-th fault type are concatenated with the inter-sample features to obtain the fault features of the c-th fault type. For all fault types, the fault type is used as a label and the corresponding fault features are used as training samples. They are input into the classifier and the classifier is trained until convergence.

[0167] Step S6: Fault diagnosis

[0168] For an RF circuit board, select a frequency sweep signal as the excitation signal f in (t) is injected into the input end of the RF circuit board, and the time domain signal f is collected from the output end of the RF circuit board out(t), and obtain the time domain data f(n). Repeat the above process to obtain multiple sets of time domain data f(n), thereby forming a systematic data set. Then, process according to the methods of steps S3, S4.2, and S4.2.1 to obtain sample features and the "base spectrum diagram", which are compared with the "base spectrum diagram" of all fault types. The Pearson correlation coefficient is calculated to obtain C related features, and then the sample features are concatenated with the inter-sample features to obtain the fault features, which are sent to the trained classifier in step S5, i.e., the training converges, to obtain the fault type and complete the fault diagnosis.

[0169] Table 3 shows the diagnostic effect of the present invention.

[0170] Fault type Test samples Predict accurate samples Accuracy F0 40 37 0.925 F1 40 39 0.975 F2 40 40 1 F3 40 39 0.975 F4 40 40 1 F5 40 40 1 F6 40 40 1 F7 40 40 1 F8 40 40 1 F9 40 40 1 F10 40 39 0.975 F11 40 40 1 F12 40 40 1 F13 40 40 1

[0171] In this embodiment, for the fault prediction of 14 fault types F0-F13, 10 fault types have reached 100%, and the other four are also above 92.5%, achieving the expected effect.

[0172] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A data-driven RF circuit board fault diagnosis method, characterized in that: The following steps are involved: (1) Model the RF circuit board, select important components for monitoring, and introduce a fault model on the RF circuit board. According to the fault model, a predetermined fault is introduced on the RF circuit board, that is, a fault type is selected; (2) For the selected fault type, select the sweep frequency signal as the excitation signal f in (t) is injected into the input end of the RF circuit board, and the time domain signal f is collected from the output end of the RF circuit board out (t), and obtain the time domain data f(n). Repeat the above process to obtain multiple sets of time domain data f(n), thereby forming a systematic data set, where t is time and n is the number of sampling points; (3) Double filtering is performed on the collected time domain data f(n) 3.1) Construct a pure noise time domain signal s with the same length as the time domain data f(n) extended (n) 3.1.1) Extract the envelope signal from the time domain data f(n) a) Calculate the discrete Fourier transform to obtain the frequency domain data F(k): in, N is the length of discrete Fourier transform, k represents the frequency point; b) Calculate the frequency response of the Hilbert transform H(k) is the transformation matrix; c) Calculate the signal after Hilbert transform d) Calculate the envelope A(n): 3.1.2) Extract pure noise data s(n) Determine a threshold value a. For the data points in the time-domain data f(n) that satisfy A(n) < a, they are determined to be pure noise data, and thus the pure noise data s(n) is extracted, where n1 < n < n2, 0 < n1 < n2 < n0, and among them, n0 is the number of sampling points of the time domain data f(n); 3.1.3) Generate a pure noise time domain signal s of equal length extended (n) The pure noise data s(n) is periodically extended to generate a pure noise time domain signal s of equal length extended (n): s extended (n)=s n ((nmodT)+n1) Wherein, n=0, 1, 2, ... n0-1; 3.2) Adaptive filtering a) For pure noise time domain signal s extended (n) Filter to obtain noise data Wherein, w(n) is the coefficient vector of the adaptive filter, w(n)=[w0(n), w1(n), ..., w M-1 (n)] T , M is the order of the adaptive filter, w m (n), m = 0, 1, 2, ... M-1 is the coefficient of the adaptive filter, s extended (n)=[s extended (n), s extended (n-1), ..., s extended (n-M+1)] T ; b) Calculate the error signal e(n): c) Using the least mean square error criterion (LMS) algorithm, the mean square value of the error signal is minimized by iteratively adjusting the adaptive filter coefficients: J(n)=E[e 2 (n)] Where E[·] represents the expected value, and the adaptive filter coefficients are iteratively adjusted as: w(n+1)=w(n)+μs extended (n)e(n) Where μ is the step size factor, which controls the rate and stability of updating the adaptive filter coefficients; d) Use the error signal e(n) as the time domain data after adaptive filtering 3.3) Fourier Transform For the adaptively filtered time domain data Perform Fourier transform to obtain frequency domain data Where N is the length of the discrete Fourier transform, and k represents the frequency point; 3.4) Sliding maximum filtering For frequency domain data Perform sliding maximum filtering. For each frequency domain point k, the sliding window contains W frequency domain samples. The output of the maximum filtering is the frequency domain data is the maximum value within the window: Among them, the value range of k needs to ensure that km is within the effective frequency range, that is, 0≤km<N; (4) Frequency domain data Perform feature extraction, 4.1) Extract key indicators describing frequency domain waveforms as sample features 4.2) Extract features that characterize the differences between different fault types, that is, extract inter-sample features; 4.2.1) Build a "base spectrum diagram" for each fault type a) First, randomly select L frequency domain data after secondary filtering from the fault type. The mean of a single frequency domain data is recorded as x i , i = 1, 2, ..., L, the mean of the L frequency domain data means is recorded as Then the mean value of the i-th frequency domain data is x i The residual V i : b) Calculate the standard deviation σ using the Bessel formula x : c) Use the Wright test method to eliminate gross errors in the frequency domain data. That is, if the absolute value of the residual of a frequency domain data is greater than 3 times the standard deviation, that is, |v i |>3σ x , the frequency domain data is a gross error and is removed, and the process returns to step a) to reselect, calculate and remove, until L frequency domain data are randomly selected and filtered twice without gross errors; d) The mean of the L frequency domain data that do not contain gross errors As the "basic spectrum diagram" of this fault type; 4.2.2) Extract the Pearson correlation coefficient: For the "base spectrum" of the cth fault type, it is recorded as And the "base spectrum diagram" with all fault types Calculate the Pearson correlation coefficient and get C related features r cj , j = 1, 2, ..., C: in, x ci is the i-th frequency domain data of the c-th fault type that does not contain gross errors, x ji is the jth piece of frequency domain data among the L pieces of frequency domain data that do not contain gross errors and correspond to the th fault type; C relevant features r cj That is, the inter-sample characteristic of the cth fault type; (5) The sample features of the c-th fault type are concatenated with the inter-sample features to obtain the fault features of the c-th fault type. For all fault types, the fault type is used as a label and the corresponding fault features are used as training samples. The fault features are input into the classifier and the classifier is trained until convergence. (6) Fault diagnosis For an RF circuit board, select a frequency sweep signal as the excitation signal f in (t) is injected into the input end of the RF circuit board, and the time domain signal f is collected from the output end of the RF circuit board out (t), get the time domain data f(n), repeat the above process to obtain multiple sets of time domain data f(n), thus forming a systematic data set, and then process it according to the methods of steps (3), 4.1), and 4.2.1) to get the sample features and the "base spectrum diagram", which are compared with the "base spectrum diagram" of all fault types. The Pearson correlation coefficient is calculated to obtain C related features, and then the sample features are concatenated with the inter-sample features to obtain the fault features, which are sent to the trained classifier in step (5) to obtain the fault type and complete the fault diagnosis.

2. The data-driven RF circuit board fault diagnosis method according to claim 1 is characterized in that: The sample characteristics include maximum value, range, mean value, variance, rising edge width, falling edge width, bandwidth, peak factor, margin, and waveform factor.

Images