A multi-step life prediction method for radio frequency low noise amplifier circuit

Abstract

The application discloses a multi-step life prediction method for a radio frequency low-noise amplifier circuit, sweeps a frequency signal as an excitation signal of an input end of the radio frequency low-noise amplifier circuit to be tested during life prediction of the radio frequency low-noise amplifier circuit, then extracts circuit characteristic parameters at different frequencies and different time points to form a characteristic matrix, calculates a health score of the circuit according to the characteristic matrix to form a health score sequence, takes two GRU models as a generator and a discriminator of a generative adversarial network respectively, trains the generative adversarial network by using the health score sequence, extracts the generator from the trained generative adversarial network as a life prediction model, then extracts the characteristic matrix and calculates the health score sequence when life prediction is needed, and inputs the health score sequence into the life prediction model to iteratively predict a health score prediction value. The life prediction model is trained based on the generative adversarial network, and the life prediction accuracy of the radio frequency low-noise amplifier circuit is improved.

Description

A multi-step lifetime prediction method for RF low-noise amplifier circuits Technical Field

[0001] This invention belongs to the field of radio frequency low noise amplifier circuit technology, and more specifically, relates to a multi-step lifetime prediction method for radio frequency low noise amplifier circuits. Background Technology

[0002] Fault prediction and health management are important means of maintaining equipment safety and reliability, and saving maintenance costs. Research on fault prediction and health management in circuits has mainly focused on low-frequency analog circuits, with very little research on radio frequency (RF) circuits. The main theories and analysis methods for RF circuits are quite different from those for low-frequency analog circuits; therefore, research results on low-frequency analog circuits cannot be extended to RF circuits, making research on fault prediction and health management for RF circuits urgently needed.

[0003] Radio frequency (RF) circuits are easily affected by wiring, equipment, and the environment, leading to failures and significant losses. Therefore, extrapolating the degradation time series of RF circuits to the future allows for advance knowledge of the circuit's lifespan, providing sufficient time to decide whether to continue operating the circuit. The longer the extrapolation period, the more effectively the circuit can be guaranteed to operate normally. How to achieve multi-step lifespan prediction of RF circuits based on historical data is an important research problem.

[0004] There are three main strategies for multi-step prediction: direct strategies, iterative strategies, and multi-output strategies. Direct and iterative strategies are essentially single-step predictions. The difference lies in the training time: direct strategies train a model for each time step, while iterative strategies continuously replace the true values ​​with predicted values ​​to create new sequences, using a single model for multi-step predictions. Comparatively, iterative strategies have shorter training times but accumulate prediction errors, leading to a rapid decline in prediction performance as the prediction time step increases. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-step lifetime prediction method for radio frequency low noise amplifier circuits. The lifetime prediction model is obtained by training a generative adversarial network, thereby improving the lifetime prediction accuracy of radio frequency low noise amplifier circuits.

[0006] To achieve the above-mentioned objective, the multi-step lifetime prediction method for radio frequency low-noise amplifier circuits of the present invention includes the following steps:

[0007] S1: Set the parameters of the linear sweep signal of the RF low-noise amplifier circuit under test according to actual needs, including the center frequency f and the lower limit frequency f. low Upper limit frequency f high The frequency step Δf is used to calculate the number of frequencies N = (fhigh -f low ) / Δf+1;

[0008] S2: Using the linear sweep frequency signal as the excitation signal, the RF low-noise amplifier circuit under test is simulated and analyzed. The feature matrices of the circuit at different times are extracted, where the feature matrix F at time t is... t The size is N×M, where M represents the number of circuit characteristic parameters, and matrix F t The nth row vector is at frequency f low The circuit characteristic vector obtained under the excitation signal +nΔf in This indicates that at time t, at frequency f low The value of the m-th circuit characteristic parameter obtained under the excitation signal +nΔf, n=0,1,…,N-1, m=0,1,…,M-1, t=0,1,…,T, where T represents the number of time points;

[0009] S3: Calculate the corresponding health score based on the feature matrix at each time step, using the following method:

[0010] First, we examine the characteristic matrix F at time t. t The characteristic parameters of each circuit are normalized to obtain the normalized characteristic matrix.

[0011] Calculate the characteristic matrix at each time t′ With characteristic matrix Euclidean distance of the corresponding column vectors t′=1,2,…,T, and then the health score h at each time t′ is calculated using the following formula. t′ :

[0012]

[0013] The health scores at all times form a health score sequence H = {h1, h2, ..., h T};

[0014] S4: Divide the obtained health score sequence into several subsequences using a sliding window of length L and step size 1. The value of L is set according to the actual situation. Let the number of sequences obtained be K = T - L + 1. Treat each subsequence as a real sequence and denote the k-th real sequence as H. k ={h k ,h k+1 ,…,h k+L-1}, k = 1, 2, ..., K;

[0015] S5: Constructing a lifetime prediction model based on generative adversarial networks, the specific steps include:

[0016] S5.1: Set up two GRU models according to the actual situation, which serve as the generator and discriminator of the generative adversarial network respectively, and initialize the parameters of the generative adversarial network.

[0017] S5.2: For each training sequence H k ={h k ,h k+1 ,…,h k+L-1 Each of the G spurious sequences is generated using an iterative strategy, specifically as follows:

[0018] The training sequence H k ={h k ,h k+1 ,…,h k+L-1 Input generator, obtain based on training sequence H k Predicted health score at time k+L Constructed based on training sequence H k False sequences Then the spurious sequence Input generator, obtain based on training sequence H k Health score prediction at time k+L+1 Constructed based on training sequence H k False sequences This process is repeated G times to generate G predictions based on the training sequence H. k False sequences

[0019] S5.3: For each true sequence H k Search for fake sequences that are the same as the one generated in step S5.2 to form real-fake sequence pairs;

[0020] S5.4: Input each real / fake sequence pair into the discriminator, and the discriminator determines the probability that the real sequence or the fake sequence is real;

[0021] S5.5: Alternately update the parameters of the generator and discriminator;

[0022] S5.6: Determine whether the generative adversarial network has met the training termination condition. If it has, proceed to step S5.7; otherwise, return to step S5.2.

[0023] S5.7: Extract the generator from the trained generative adversarial network as a lifetime prediction model;

[0024] S6: When lifetime prediction is required, the linear sweep frequency signal is used as the excitation signal input to the RF low-noise amplifier circuit under test. The circuit characteristic parameters at the most recent L time moments are collected, and a feature matrix F of size L×M is constructed using the method in step S2. j j = 1, 2, ..., L; calculate the health score h of j at each time point using the method in step S3. j This constitutes the health score sequence H′={h1,h2,…,h L Input the health score sequence H′ into the lifespan prediction model, and iteratively predict the health score values ​​for the next R time points. The value of R is determined according to actual needs.

[0025] This invention discloses a multi-step lifetime prediction method for RF low-noise amplifier circuits. A swept-frequency signal is used as the excitation signal at the input of the RF low-noise amplifier circuit under test for lifetime prediction. Then, circuit characteristic parameters at different frequencies and times are extracted to form a feature matrix. Based on the feature matrix, a health score is calculated to form a health score sequence. Two GRU models are used as the generator and discriminator of a generative adversarial network (GAN), respectively. The GAN is trained using the health score sequence. The generator is extracted from the trained GAN as the lifetime prediction model. When lifetime prediction is needed, the feature matrix is ​​extracted, the health score sequence is calculated, and the lifetime prediction model is input for iterative prediction to obtain the predicted health score value.

[0026] The present invention has the following beneficial effects:

[0027] 1) This invention uses a swept frequency signal as the excitation signal at the input terminal when predicting the lifetime of the radio frequency low noise amplifier circuit under test. The swept frequency range of the swept frequency signal covers the entire operating frequency of the radio frequency circuit. The extracted feature matrix has feature parameters at different frequencies, which can better characterize the state of the radio frequency circuit within the operating frequency range, thereby effectively extracting richer circuit state information.

[0028] 2) This invention improves the accuracy of multi-step prediction by making the distribution of the predicted sequence as close as possible to the original sequence through the game between the generator and the discriminator during the training of the generative adversarial network. Attached Figure Description

[0029] Figure 1 is a flowchart of a specific implementation method for the multi-step lifetime prediction method of the radio frequency low noise amplifier circuit of the present invention.

[0030] Figure 2 is a flowchart of the prediction model constructed based on generative adversarial networks in this invention;

[0031] Figure 3 is a structural diagram of the generative adversarial network in this invention;

[0032] Figure 4 is a structural diagram of the radio frequency low noise amplifier circuit in this embodiment;

[0033] Figure 5 is a characteristic parameter curve of the RF low noise amplifier circuit based on ATF54143 in this embodiment under normal operating conditions.

[0034] Figure 6 is the loss function curve for lifetime prediction of the ATF54143-based RF low-noise amplifier circuit using the present invention in this embodiment;

[0035] Figure 7 is a comparison of the prediction results of the RF low-noise amplifier circuit based on ATF54143 in this embodiment. Detailed Implementation

[0036] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0037] Example

[0038] Figure 1 is a flowchart illustrating a specific implementation of the multi-step lifetime prediction method for RF low-noise amplifier circuits according to the present invention. As shown in Figure 1, the specific steps of the multi-step lifetime prediction method for RF low-noise amplifier circuits according to the present invention include:

[0039] S101: Set the parameters for the linear sweep frequency signal:

[0040] Set the parameters of the linear sweep signal of the RF low-noise amplifier circuit under test according to actual needs, including the center frequency f and the lower limit frequency f. low Upper limit frequency f high The frequency step Δf is used to calculate the number of frequencies N = (f high -f low ) / Δf+1.

[0041] S102: Extracting the feature matrix of the circuit:

[0042] A linear sweep frequency signal is used as the excitation signal to simulate and analyze the RF low-noise amplifier circuit under test. The feature matrices of the circuit at different times are extracted, where the feature matrix F at time t is... t The size is N×M, where M represents the number of circuit characteristic parameters, and matrix F t The nth row vector is at frequency f low The circuit characteristic vector obtained under the excitation signal +nΔf in This indicates that at time t, at frequency f low The value of the m-th circuit characteristic parameter obtained under the excitation signal +nΔf, n=0,1,…,N-1, m=0,1,…,M-1, t=0,1,…,T, where T represents the number of time points.

[0043] The circuit characteristic parameters can be set according to actual needs. In this embodiment, the circuit characteristic parameters include the S-parameters of the RF low-noise amplifier circuit under test. 11 ,S 12 ,S 21 ,S 22 Input VSWR, output VSWR, stability and noise figure.

[0044] S103: Calculate the health score sequence:

[0045] The health score is calculated based on the feature matrix at each time point, as follows:

[0046] First, we examine the characteristic matrix F at time t. t The characteristic parameters of each circuit are normalized to obtain the normalized characteristic matrix. The normalization formula can be expressed as:

[0047]

[0048] Among them, f m,max f m,min This represents the maximum and minimum values ​​of the m-th circuit characteristic parameter.

[0049] Calculate the characteristic matrix at each time t′ With characteristic matrix Euclidean distance of the corresponding column vectors t′=1,2,…,T:

[0050]

[0051] Then, the health score h at each time t is calculated using the following formula. t′ :

[0052]

[0053] The health scores at all times form a health score sequence H = {h1, h2, ..., h T}

[0054] S104: Segmenting the health score sequence:

[0055] The obtained health score sequence is divided into several subsequences using a sliding window of length L and step size of 1. The value of L is set according to the actual situation, and the number of sequences obtained is denoted as K = T - L + 1. Each subsequence is taken as the real sequence, and the k-th real sequence is denoted as H. k ={h k ,hk+1 ,…,h k+L-1}, k = 1, 2, ..., K.

[0056] S105: Constructing a lifetime prediction model based on generative adversarial networks:

[0057] To better predict the lifetime of RF low-noise amplifier circuits, this invention constructs a lifetime prediction model based on generative adversarial networks (GANs). Figure 2 is a flowchart of the prediction model constructed based on GANs in this invention. As shown in Figure 2, the specific steps of constructing the prediction model based on GANs in this invention include:

[0058] S201: Constructing a Generative Adversarial Network:

[0059] Two GRU (Gate Recurrent Unit) models are set up according to the actual situation, serving as the generator and discriminator of the generative adversarial network (GAN) respectively, and the parameters of the GAN are initialized. GRU (Gate Recurrent Unit) is a type of Recurrent Neural Network (RNN) that can quickly and accurately extract temporal features. Furthermore, GRU models have fewer parameters, lower complexity, and are easier to train compared to other RNN models. The working principle and specific structure of the GRU model will not be elaborated here. Figure 3 is a structural diagram of the GAN in this invention. As shown in Figure 3, in the GAN of this invention, the generator generates fake sequences based on the training set, and then feeds them and the corresponding real sequences into the discriminator for discrimination to obtain the true / false probabilities, thereby completing the training of the GAN.

[0060] S202: Generating fake sequences:

[0061] For each training sequence H k ={h k ,h k+1 ,…,h k+L-1 Each of the G spurious sequences is generated using an iterative strategy, specifically as follows:

[0062] The training sequence H k ={h k ,h k+1 ,…,h k+L-1 Input generator, obtain based on training sequence H k Predicted health score at time k+L Constructed based on training sequence H k False sequences Then the spurious sequence Input generator, obtain based on training sequence H k Health score prediction at time k+L+1 Constructed based on training sequence H k False sequences This process is repeated G times to generate G predictions based on the training sequence H. k False sequences

[0063] S203: Identify real-spurious sequence pairs:

[0064] For each real sequence H k Search for fake sequences that are the same as the one generated in step S202 to form real-fake sequence pairs.

[0065] S204: Discriminator discrimination:

[0066] Each real / fake sequence pair is input into the discriminator, which then determines the probability that the real or fake sequence is real.

[0067] S205: Parameter Update:

[0068] The parameters of the generator and discriminator are updated alternately; that is, the discriminator is updated several times, and then the generator is updated once.

[0069] In this embodiment, the loss function Gen during generator update is... loss The following formula is used for calculation:

[0070]

[0071] Where q = 1, 2, ..., Q, Q represents the number of real-spurious sequence pairs obtained in step S203, and x q , Let represent the real sequence and the fake sequence in the q-th real-fake sequence pair, respectively.

[0072] The loss function Disc during discriminator update loss The following formula is used for calculation:

[0073]

[0074] Among them, y q D represents the distribution of the true sequence, and D() represents the discrimination result of the sequence.

[0075] S206: Determine whether the generative adversarial network has reached the training termination condition. If it has, proceed to step S207; otherwise, return to step S202.

[0076] S207: Obtain the lifetime prediction model:

[0077] Extract the generator from the trained generative adversarial network as a lifetime prediction model.

[0078] S106: Lifespan Prediction:

[0079] The linear sweep frequency signal is used as the excitation signal input to the RF low-noise amplifier circuit under test. The circuit characteristic parameters at the most recent L time moments are collected, and a feature matrix F of size L×M is constructed using the method in step S102. j j = 1, 2, ..., L. The health score h of j at each time point is calculated using the method described in step S103. j This constitutes the health score sequence H′={h1,h2,…,h L The health score sequence H′ is input into the life expectancy prediction model, and the predicted health scores for the next R time points are obtained through iterative prediction. The value of R is determined according to actual needs.

[0080] To better illustrate the technical solution of this invention, a specific example is used to experimentally verify the invention. Figure 4 is a structural diagram of the RF low-noise amplifier circuit in this embodiment. As shown in Figure 4, the circuit used in this embodiment is an RF low-noise amplifier circuit based on ATF54143. The operating frequency of this circuit is 2.4-2.5GHz. Therefore, the center frequency f of the linear sweep signal is set to 2.45GHz, and the lower limit frequency f is set to 2.45GHz. low =2.4GHz, upper limit frequency f high =2.5GHz, frequency step Δf = 0.01GHz. The above linear sweep signal is used as the excitation signal at the input terminal for lifetime prediction of this RF circuit. The circuit characteristic parameters extracted in this embodiment include S-parameters S 11 ,S 12 ,S 21 ,S 22 Input VSWR, output VSWR, stability and noise figure, therefore matrix F t The nth row vector is represented as

[0081] The changes in the low-noise amplifier (LNA) voltage at different times are reflected in the transistor gate voltage. In this embodiment, to simulate the degradation of the LNA circuit during simulation analysis, the following formula is used to calculate the voltage ΔV applied to the transistor gate at time t. th Adjustments will be made:

[0082] ΔV th =C·t n

[0083] Where C is a parameter that depends on the process change and voltage, the time exponent n = 0.25, and t is the stress time.

[0084] Figure 5 is a graph showing the characteristic parameters of the RF low-noise amplifier circuit based on ATF54143 under normal operating conditions in this embodiment. The extracted feature matrix is ​​as follows:

[0085]

[0086] Then, the generative adversarial network is trained to extract the lifetime prediction model. Figure 6 shows the loss function curve for lifetime prediction of the ATF54143-based RF low-noise amplifier circuit using this invention in this embodiment. Then, lifetime prediction is performed on the ATF54143-based RF low-noise amplifier circuit.

[0087] Figure 7 is a comparison of the prediction results of the RF low-noise amplifier circuit based on ATF54143 in this embodiment. As shown in Figure 7, the prediction results of the method of the present invention are compared with the actual values ​​and the prediction results of the conventional GRU model. It can be seen that the prediction results of the method of the present invention are very close to the actual curve and are better than the prediction results of the conventional GRU model.

[0088] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A multi-step lifetime prediction method for radio frequency low-noise amplifier circuits, characterized in that, Includes the following steps: S1: Set the parameters of the linear sweep signal for the RF low-noise amplifier circuit under test, including the center frequency. Lower limit frequency upper limit frequency Frequency step The number of frequencies was calculated. S2: Using the linear sweep frequency signal as the excitation signal, the RF low-noise amplifier circuit under test is simulated and analyzed. The feature matrix of the circuit at different times is extracted, where the feature matrix at time... Feature matrix The size is ， The matrix represents the number of characteristic parameters of the circuit. The Middle Each row vector is at a frequency of The circuit feature vector obtained under the excitation signal; S3: Calculate the corresponding health score based on the feature matrix at each time step, the specific method is as follows: First, for time step... Feature matrix The characteristic parameters of each circuit are normalized to obtain the normalized characteristic matrix. ; Calculate each time point Feature matrix With characteristic matrix Euclidean distance of the corresponding column vectors ， Then, the following formula is used to calculate each time step. Health score ： Construct a health score sequence from the health scores at all times. S4: Transform the obtained health score sequence into a sequence of length [length missing]. Divide the sequence into several subsequences using a sliding window with a step size of 1, and record the number of subsequences obtained. Treat each subsequence as the real sequence, and denote the first subsequence as the first subsequence. a real sequence ， S5: Construct a lifetime prediction model based on generative adversarial networks. Specific steps include: S5.1: Setting up two GRU models, serving as the generator and discriminator of the generative adversarial network respectively; S5.2: For each training sequence... Generate using iterative strategies respectively A fake sequence, specifically, is obtained by using the training sequence... Input generator, get results based on training sequences The moment Health score prediction , constituted based on training sequences False sequences Then the spurious sequence Input generator, get results based on training sequences The moment Health score prediction , constituted based on training sequences False sequences ; thus iterative prediction Next, generate A training sequence-based False sequences ， S5.3: For each true sequence S5.2: Search for fake sequences with the same time interval from the fake sequences generated in step S5.2 to form real-fake sequence pairs; S5.4: Input each real-fake sequence pair into the discriminator, which determines the probability that the real or fake sequence is real; S5.5: Alternately update the parameters of the generator and discriminator; S5.6: Extract the generator from the trained generative adversarial network as the lifetime prediction model; S6: When lifetime prediction is required, input the linear sweep frequency signal as the excitation signal into the RF low-noise amplifier circuit under test to acquire the most recent The circuit characteristic parameters at each time point are used to construct a value of [value] using the method in step S2. Feature matrix ， The method in step S3 is used to calculate each time step. Health score , forming a health score sequence ; Sequence of health scores Input a lifetime prediction model, and iteratively predict the future. Predicted health score at each time point.

2. The multi-step lifetime prediction method for RF low-noise amplifier circuits according to claim 1, characterized in that, The circuit characteristic parameters in step S2 include the S-parameters of the radio frequency circuit under test. Input VSWR, output VSWR, stability and noise figure.

Images