A Method for Identifying Power Amplifier Fingerprint Features Based on Supervised Contrastive Learning

Abstract

The present invention discloses a power amplifier fingerprint feature recognition method based on supervised contrastive learning, belonging to the field of extraction and analysis of fingerprint features of integrated circuit devices. By using an improved supervised contrastive learning network algorithm, the measurability of features can be improved, making the features become practical fingerprint features to solve the problem of power amplifier fingerprint feature recognition. The present invention derives the expression of the harmonic amplitude under the excitation of a variable-power sine signal, obtains the constraint relation expression and the harmonic feature expression, and uses an improved supervised contrastive learning network algorithm to extract the constrained harmonic features and perform feature recognition.

Description

Technical Field

[0001] The present invention relates to the field of extraction and analysis of fingerprint features of integrated circuit devices, and particularly to a method for identifying power amplifier fingerprint features based on supervised contrastive learning. Background Art

[0002] A power amplifier (PA) is a very complex non-linear device, and the research on power amplifier modeling has a long history in both academic and industrial circles. How to effectively reduce its non-linear impact on the system while utilizing the amplification characteristics of the power amplifier is a common concern in the fields of integrated circuit design, wireless communication, etc. Especially in recent years, with the rapid development of communication technologies, solving the low-distortion wireless transmission of communication signals has become an urgent need to improve the transmission efficiency of base stations and the channel utilization rate, and a large amount of research has focused on power amplifier modeling.

[0003] Power amplifier modeling is generally divided into physical modeling and behavioral modeling. Physical modeling generally requires obtaining the circuit structure of the power amplifier, component characteristics, basic circuit laws, and related theoretical rules. If the above data is obtained, the characteristics of the power amplifier are already clear at a glance, losing the meaning of fingerprint extraction for the power amplifier. Therefore, the present invention adopts the method of behavioral modeling for design.

[0004] The power amplifier is the core component of a master-oscillator power amplifier (MOPA) radio transmitter and is also one of the main sources of the radio frequency characteristics of the transmitter. In order to effectively enhance the communication effect, the transmitter must use high gain to generate a strong transmission signal to overcome the 1 / R4 energy attenuation during space transmission. Once high-gain devices operate in a high-gain state, they will all produce very significant non-linear effects, such as high-order harmonics and cross-modulation, that is, the non-linear "unintentional modulation" of the amplifier. The occurrence of these "unintentional modulations" constitutes the basic idea of power amplifier fingerprint recognition: if characteristic quantities related to the inherent characteristics of the power amplifier device can be extracted from the output signal of the power amplifier, then based on this characteristic, the original characteristics of the power amplifier can still be recognized when the intentional modulation changes. It is difficult for traditional methods to extract these "unintentional modulations" of the power amplifier from subtle spectrum data. By using an improved supervised contrastive learning network algorithm, the measurability of the characteristics can be improved, making this characteristic a practical fingerprint characteristic to solve the problem of power amplifier fingerprint feature recognition. The present invention derives the expression of the harmonic amplitude under the excitation of a variable-power sine signal, obtains the constraint relation expression and the harmonic characteristic expression, and uses an improved supervised contrastive learning network algorithm to extract the constrained harmonic characteristics and perform feature recognition. Summary of the Invention

[0005] The purpose of the present invention is to improve the measurability of features by using an improved supervised contrastive learning network algorithm, so that the features become practical fingerprint features to solve the problem of power amplifier fingerprint feature recognition. According to the above problem, the present invention provides a power amplifier fingerprint feature recognition method based on supervised contrastive learning.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a power amplifier fingerprint feature recognition method based on supervised contrast learning, comprising the following steps:

[0007] S1. First, input unlabeled amplifier feature information into the network and extract unlabeled feature data;

[0008] S2. Then the unlabeled feature data is reduced in dimension by PCA;

[0009] S3. After the feature data set is reduced in dimension by PCA, the reduced data is clustered using an iterative self-organizing clustering algorithm. After the calculation is completed, it is determined whether the current clustering effect meets the initially set clustering standard, whether the number of iterations is completed, and the relationship between the number of cluster centers and the expected number of cluster centers. After iterative cyclic clustering training, the final clustering result and the pseudo-labels used for weighting in supervised comparative training are obtained;

[0010] S4. Use the pseudo-label weighted cross loss method for training. Assigning different weights to features of the same class and features of different classes can improve the efficiency of training. Then use unlabeled feature data to perform weight optimization training on the supervised contrastive learning network.

[0011] S5. Input the obtained sample data into the encoder. The encoder is used to obtain an abstract expression of the feature in a higher dimension. Here, the encoder uses ResNeSt as the feature extraction network.

[0012] S6. After the feature encoding is completed, the encoded data is subjected to loss space feature mapping, by mapping the high-dimensional features to the representation space applying the weighted contrast loss;

[0013] S7. The data features extracted from the sample and the corresponding positive data set are brought closer in feature space, and the data features extracted from the negative data set of the sample are brought further away in feature space. The similarity between the features is measured using cosine similarity. After the training is completed, the parameter values ​​of the neuron nodes in the supervised contrastive learning network are fixed as the initial values ​​of the network nodes for the next step of supervised learning fine-tuning.

[0014] S8. Add a linear classification layer behind the trained network, and perform fine-tuning training by minimizing the cross-entropy loss function as the supervision information. Use the minority of the labeled power amplifier fingerprint feature dataset Feature_label as the input data for the supervised training, and fine-tune the network with the fixed parameter node information as the training initial value. While fine-tuning the parameters on the basis of the initialized network, train the added linear classification layer at the same time;

[0015] S9. The network after fine-tuning can be used for fingerprint feature recognition of power amplifier devices.

[0016] The beneficial effects of the invention are as follows:

[0017] The present invention can improve the measurability of features from the use of the improved supervised contrastive learning network algorithm, making the features become practical fingerprint features to solve the problem of power amplifier fingerprint feature recognition. The present invention derives the expression of the harmonic amplitude under the excitation of a variable-power sine signal, obtains the constraint relation, introduces the Wu Wenjun method to constrain the harmonic features, and uses the improved supervised contrastive learning network algorithm to extract the constrained harmonic features and perform feature recognition. Brief Description of the Drawings

[0018] Figure 1 It is the supervised contrastive learning network framework of the present invention;

[0019] Figure 2 It is the flowchart of the working process of the present invention;

[0020] Figure 3 It is the schematic diagram of the power amplifier feature data classification network of the present invention. Detailed Embodiments

[0021] To make the technical means, creative features, achieved purposes and functions of the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.

[0022] To achieve the above object, the present invention considers using harmonics for feature extraction. First, it is necessary to describe the power amplifier as an instantaneous non-linear function using the Taylor series, and then analyze the harmonic features in the case of a sine excitation signal.

[0023] The general form of a narrowband signal is

[0024] u(t) = r(t)cos(2πf0t + θ) (1)

[0025] where r(t) is the baseband envelope of the signal u(t), f0 is the carrier frequency, and θ is an arbitrary initial phase. Then the output signal of the amplifier is:

[0026] y(t) = O{u(t)} (2)

[0027] Where \(O\{\cdot\}\) is the overall modulation characteristic of the amplifier, which consists of the amplitude-amplitude (AM-AM) transfer function and the amplitude-phase transfer function (AM-PM). Let \(O\{u(t)\}\) be expanded in a Taylor series with respect to \(u(t)\) to obtain the Taylor series model of the RF (Radio Frequency) amplifier.

[0028]

[0029] When analyzing the unintentional modulation of the entire spectrum, both the odd and even power terms in the above equation need to be considered simultaneously. Since this patent only analyzes the unintentional modulation within the signal operating frequency band, it can be known from the literature "Linear RF power amplifier design for wireless signals: a spectrum analysis approach" that the signal components generated by the even power are always far from the operating frequency band, so the even power terms can be ignored.

[0030] According to the fundamentals of software radio, the nonlinear distortion of the transmitter is mainly caused by the final power amplifier (or the final amplification link) of the transmitter, and there are no frequency-doubled harmonics and cross-modulation components in the input signal of the final power amplifier. Take the excitation sine signal of the power amplifier and truncate the Taylor series model. Taking the third order as an example:

[0031] When the order of the amplifier model is 3, for the excitation signal represented by Equation (1), let \(r(t)=A\), and denote \(\omega = 2\pi f_0\). The output signal \(y(t)\) can be expressed as:

[0032] \(y(t)=a_0 + a_1A\sin(\omega t+\theta)+a_2A\) 2 \(\sin\) 2 (\omega t+\theta)+a_3A\) 3 \(\sin\) 3 (\omega t+\theta)(4)\)

[0033] After trigonometric expansion, the expression of the logarithm \(A\) of the amplitude of the \(i\)-th harmonic component relative to the amplitude of the carrier component can be obtained as: t (r) The expression is:

[0034]

[0035]

[0036] It can be seen that the relative amplitudes of the harmonics are all functions of the amplitude of the excitation signal. Since the energy of the harmonics in \(y(t)\) is much smaller than the energy of the carrier, that is

[0037] \(a_1A\gg a_3A\) 3 (7)\)

[0038] From equations (6) and (5), we can obtain

[0039]

[0040] where C is a constant, and equation (8) is approximately a linear constraint relationship on a plane. Thus, under this condition, a characteristic quantity independent of the amplitude A of the excitation signal can be obtained from the relative amplitude of the harmonics.

[0041] To make the conclusion more general, the Taylor series model can be transformed into a special case of the Generalized Frequency Response Function (GFRF) model. GFRF is a direct generalization of the frequency response function of a linear system to a nonlinear system. This method assumes that the GFRF characteristics are inherent. When the input signal is known, the GFRF model can be used to obtain the system response, that is, the form of the output signal. The Taylor series is a very special class of polynomial-like nonlinear equations, so the expressions of harmonics can be obtained using the conclusions of nonlinear system theory:

[0042]

[0043] where D represents the differential operator, D pi represents the p i th order differential operator, M is the maximum differential order of the differential equation, N is the maximum power order, the subscripted a, b, c represent system parameters, and u(t) and y(t) are the input and output signals of the system respectively.

[0044] Let

[0045]

[0046]

[0047] Then, as a special case of equation (9), the pure input system has the following GFRF recurrence relation [7] .

[0048] The pure input system is

[0049]

[0050] Its GFRF is

[0051]

[0052]

[0053] where is the nth order Generalized Frequency Response Function (GFRF) of the system.

[0054] The Taylor series model of the amplifier belongs to the pure input system, and the DC component does not need to be considered. The Lth order Taylor series model of the amplifier is

[0055]

[0056] From equations (10) and (11), we have

[0057] A1(jω1) = 1 (16)

[0058] C n (jω1, jω2, …, jω n ) = -a n , n = 1, 2, ……, L (17)

[0059] Furthermore, from equations (13) and (14), we have

[0060]

[0061] When the input is a single sinusoidal signal, i.e., u(t) = Acos(ωt), the amplitude of the l-th harmonic of the nonlinear system is:[[]]

[0062]

[0063] Where Substituting (18) gives:[[]]

[0064]

[0065] Equation (20) is the expression for the harmonic amplitude at the output of the amplifier under the excitation of a single sinusoidal signal.[[]]

[0066] The logarithm of the relative amplitude of the first harmonic can be expressed as

[0067]

[0068] In particular, if |A| << 1, the relative amplitude is approximately

[0069]

[0070] Equation (22) shows that the logarithm of the relative amplitude has an approximately linear relationship with the logarithm of the input signal amplitude log(|A|). Taking log(|A|) as a parameter, the above equation is a set of linear equations about A l (r) The point (A l (r) , A2 (r) , …, A n (r))All fall on a straight line in the high-dimensional space. This system of equations reflects the non-linear characteristics of the power amplifier, and the constraint line is related to the amplifier. According to the approximation conditions of equations (22) and (8), it can be obtained that when N > 3, the condition for the constraint to be approximately linear is slightly stronger than when N = 3. Under the condition of meeting the approximation conditions, equation (8) can also be obtained from equation (22).

[0071] The characteristic definition expression of the fingerprint can be obtained from equation (22):

[0072]

[0073] Each single feature Feature l1,l2 The characterization can be independently used for classification. Multiple independent features can form a feature vector. For example:

[0074]

[0075] The obtained feature vector dataset Feature classifies and identifies fingerprint features through a method based on supervised contrast learning. The specific process includes:

[0076] First, input unlabeled power amplifier feature information into the network, corresponding to Figure 1 S101 in it. Perform PCA dimensionality reduction on the obtained high-dimensional feature dataset, which can reduce the time complexity while retaining the feature data information. Assume that m n-dimensional feature data (x (1) , x (2) , …, x (m) ) have all been normalized, that is The new coordinate system obtained after the projection transformation is {ω1, ω2, …, ω n ,}, where w is an orthonormal basis, that is If the feature data is reduced from n dimensions to n' dimensions, then some coordinates in the new coordinate system are discarded, and the new coordinate system is {ω1, ω2, …, ω n ',}, and the projection of the sample point x(i) in the n'-dimensional coordinate system is: where is the coordinate of x(i) in the i-th dimension in the low-dimensional coordinate system. For any feature x(i), its projection in the new coordinate system is W T x (i) , and the projection variance in the new coordinate system is x (i)T W T x (i) W. To maximize the sum of the projection variances of all samples, that is, to maximize That is:

[0077]

[0078] The Lagrange function can be used to obtain

[0079] J(W) = tr(W T XX T W) + λ(W T W - I) (26)

[0080] Taking the derivative with respect to w gives XX T W + λW = 0, which can be rearranged as: XX T W = -λW

[0081] From the above equation, it can be concluded that w is a matrix composed of n' eigenvectors of XX T and -λ is the eigenvalue of XX T . When we reduce the feature data from n dimensions to n', we need to find the eigenvectors corresponding to the largest n' eigenvalues. The matrix w composed of these n' eigenvectors is the matrix we need. For the original feature dataset, we use z (i) = W T x (i) to reduce the original dataset to n' dimensions with the minimum projection distance, corresponding to Figure 1 in S102

[0082] After the feature dataset is reduced by PCA, the reduced data is clustered using the iterative self-organizing clustering algorithm. For the input N feature samples {x i , i = 1, 2,..., n}, first select M clustering centers {z1, z2,..., z M} as the initial clustering centers, and set the initial parameter values, including the expected number of clustering centers K, the minimum distance dis_min between two clustering centers, the number of iterations of the operation, the standard deviation θ s of the sample distance distribution in a clustering, the minimum number θ N of samples in each clustering region. For each power amplifier feature sample, it is assigned to a different clustering center by the method of the nearest distance. After all samples are assigned, the positions of the clustering centers are corrected

[0083]

[0084] where N j represents the number of samples in the clustering domain S j . Then calculate the average distance between the samples in each clustering domain and the clustering center points

[0085]

[0086] Calculate the total average distance between all samples and their corresponding clustering centers

[0087]

[0088] After the calculation is completed, it is necessary to determine whether the current clustering effect meets the initially set clustering criteria, whether the number of iterations is completed, and at the same time, determine the relationship between the number of cluster centers and the expected number of cluster centers. If the number of cluster centers is less than half of the expected number of cluster centers, the splitting operation is performed. Similarly, if the number of cluster centers is more than twice the expected number, instead of performing the classification operation, the merging operation is performed. The specific operation of the splitting step is as follows: Calculate the standard deviation vector of all samples with respect to the corresponding cluster centers, σ j =(σ 1j ,σ 2j ,…,σ nj ). The calculation method of each element in this vector is:

[0089]

[0090] where i = 1, 2, …, n is the vector dimension, j = 1, 2, …, M is the number of clusters, and N j is the number of samples in the cluster. Solve the maximum vector σ jmax in the standard deviation vector. If it satisfies σ jmax > θ s , and

[0091]

[0092] then split z j into two new cluster centers and. The number of cluster centers is incremented by 1. After completing the classification operation, continue with the iterative clustering. If the splitting operation is not performed, calculate the distances between all the cluster center points, compare the obtained distances between the cluster center points with the minimum distance set initially, and sort the obtained distances in ascending order. Merge the clusters that do not meet the minimum distance between the cluster centers. After iterative loop clustering training, the final clustering result and the pseudo-labels used for weighted supervision in the contrast training are obtained, corresponding to Figure 1 in S103.

[0093] Since the traditional supervised contrast method usually regards the augmented sample aug(x(i)) of the sample x(i) as the positive example label and regards the remaining samples as negative example samples, this situation will cause the distance between samples of the same type to be pulled apart, thus affecting the final classification effect. In this paper, the method of pseudo-label weighted cross-loss is used for training. Assigning more weights to the features with consistent pseudo-labels can make the current training samples have a smaller distance in the feature space of samples of the same class. On the contrary, the distance in the feature space with different pseudo-labels is farther. Assigning different weights to the features of the same class and different classes in this step can improve the efficiency of this step of training.

[0094] Then, the unlabeled feature data is used to perform weight optimization training on the supervised contrastive learning network. After the clustering operation in the previous step, the feature data obtained can be expressed as Feature[x(i), y(i)], where x(i) represents a single feature and y(i) represents the pseudo-label corresponding to the feature. The feature data matched with the pseudo-label is subjected to three types of random enhancement processing, corresponding to Figure 1 In S104, the enhanced representation classified as one category in the corresponding pseudo-label is regarded as positive data, that is, when the label of Feature[x(i), y'] in the feature is consistent with that of another sample Feature[x(i), y"], it is regarded as the same category of features, and y' is the same as y", and the enhanced data that does not belong to the same category of features as the current sample is regarded as a negative sample, thus obtaining the positive and negative example data sets of the feature data after dimensionality reduction enhanced according to the pseudo-label classification.

[0095] The next step is to input the obtained sample data into the encoder. The role of the encoder is to obtain an abstract expression of the feature in a higher dimension. The operation of this step can be expressed as x(i)'=enco(x(i)), where enco(·) represents the encoding operation of the sample x(i). The encoder here uses ResNeSt as the feature extraction network, corresponding to Figure 1 S105 in.

[0096] After feature encoding is completed, the encoded data is mapped into a loss space. The loss space mapping network can be composed of a hidden layer of size 2048 and a linear layer of size 128. By mapping high-dimensional features to the representation space of the weighted contrast loss, the performance of calculating the loss can be effectively improved. Figure 1 In S106, the core of the training is to bring the data features extracted from the sample and the corresponding positive dataset closer in the feature space, and to increase the distance between the negative dataset of the sample in the feature space. The similarity between features is measured by cosine similarity. The similarity measurement formula between features is as follows:

[0097]

[0098] The training process corresponds to Figure 1 In S107, after the training is completed, the parameter values ​​of the neuron nodes in the supervised contrastive learning network are fixed as the initial values ​​of the network nodes for the next step of supervised learning fine-tuning.

[0099] Add a linear classification layer behind the trained network, and perform fine-tuning training by minimizing the cross-entropy loss function as the supervision information. Use the minority of the labeled power amplifier fingerprint feature dataset Feature_label as the input data for the supervised training, and fine-tune the network with the fixed parameter node information as the initial value of the training. On the basis of the initialized network, fine-tune the parameters and train the added linear classification layer at the same time. The process of this step corresponds to Figure 1 S108 in Figure 1 , and the fingerprint feature recognition of the power amplifier device can be performed through the fine-tuned network, corresponding to S109 in

[0100]

[0101] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0101] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for identifying power amplifier fingerprint features based on supervised contrastive learning, characterized in that, The steps of wireless positioning method are as follows: S1. First, input unlabeled amplifier feature information into the network and extract unlabeled feature data as harmonic features; S2. Then the unlabeled feature data is reduced in dimension by PCA; S3. After the feature data set is reduced in dimension by PCA, the reduced data is clustered using an iterative self-organizing clustering algorithm. After the calculation is completed, it is determined whether the current clustering effect meets the initially set clustering standard, whether the number of iterations is completed, and the relationship between the number of cluster centers and the expected number of cluster centers. After iterative cyclic clustering training, the final clustering result and the pseudo-labels used for weighting in supervised comparative training are obtained; S4. Use the pseudo-label weighted cross loss method for training, assign different weights to features of the same class and features of different classes, and then use unlabeled feature data to perform weight optimization training on the supervised contrastive learning network; S5. Input the obtained sample data into the encoder, where the encoder uses ResNeSt as the feature extraction network; S6. After the feature encoding is completed, the encoded data is subjected to loss space feature mapping, by mapping the high-dimensional features to the representation space applying the weighted contrast loss; S7. The data features extracted from the sample data and the corresponding positive data set are brought closer in feature space, and the data features extracted from the negative data set of the sample data are brought further away in feature space. The similarity between the features is measured using cosine similarity. After the training is completed, the parameter values ​​of the neuron nodes in the supervised contrastive learning network are fixed as the initial values ​​of the network nodes for the next step of supervised learning fine-tuning. S8. Add a linear classification layer after the trained network, use the minimization of the cross entropy loss function as the supervision information for training fine-tuning, use the labeled power amplifier fingerprint feature dataset Feature_label as the input data for supervised training, fine-tune the network with fixed parameter node information as the initial training value, fine-tune the parameters based on the initialized network and train the added linear classification layer at the same time; S9. The fine-tuned network can be used for fingerprint feature recognition of power amplifier devices.

Images