Power amplifier digital pre-distortion method and system based on minimum gating network

Abstract

The invention discloses a power amplifier digital pre-distortion method and system based on a minimum gating network, and relates to the technical field of gating networks, and the method comprises the steps: firstly, obtaining a baseband input signal, and constructing a multi-dimensional feature vector containing current and historical real part imaginary parts and first-order and third-order instantaneous envelopes; time sequence processing is carried out through a scalar minimum gating unit layer, a stable state and a normalized forgetting mechanism are introduced, and the modeling capability of long-time nonlinear dynamics is enhanced. Then attention weighting is carried out on the hidden state, and instantaneous phase sine and cosine features extracted from the input signal are fused to form a comprehensive feature; and generating a complex field pre-distortion signal through output layer mapping. According to the method, a time sequence modeling structure based on a scalar minimum gating single is introduced, and delay enhancement feature construction, an attention weighting mechanism and instantaneous phase feature fusion are combined, so that high-precision and low-complexity digital pre-distortion of nonlinear distortion of the power amplifier is realized.

Description

Technical Field

[0001] This invention relates to the field of gating network technology, and more specifically to a digital predistortion method and system for power amplifiers based on a minimum gating network. Background Technology

[0002] In wireless communication systems, power amplifiers (PAs) are nonlinear devices in the transmit link. The amplitude / amplitude (AM / AM) distortion, amplitude / phase (AM / PM) distortion, and memory effects they generate can lead to problems such as spectrum regeneration and adjacent channel leakage, severely impacting communication quality. Digital predistortion (DPD) technology is an effective method for predistorting signals in the digital domain to compensate for PA nonlinearity. Traditional DPD models, such as the memory polynomial (MP) model and the generalized memory polynomial (GMP) model, while structurally clear, have limited ability to model strong nonlinearity and complex memory effects, especially in broadband, high-power-efficiency scenarios where accuracy is insufficient. In recent years, deep learning-based neural network DPD models have shown stronger nonlinear fitting capabilities. Among them, recurrent neural networks (RNNs) and their variants, such as long short-term memory networks (LSTMs), have attracted attention in DPD modeling due to their ability to effectively handle time-series dependencies. However, LSTMs and similar structures have inherent drawbacks such as large parameter size, high computational complexity, and high hardware deployment resource consumption, making them unsuitable for application in real-time communication devices. Furthermore, existing neural network DPD models are often used as black boxes, lacking explicit correlation with the physical properties of PA, and the interpretability and robustness of the models need to be improved. Summary of the Invention

[0003] To address the aforementioned problems, this invention proposes a digital predistortion method and system for power amplifiers based on a minimum gate network.

[0004] On the one hand, digital predistortion methods for power amplifiers based on minimum gated networks include:

[0005] S1, obtain the baseband input signal at the current moment;

[0006] S2, delay enhancement is performed on the baseband input signal to construct a multi-dimensional input feature vector including the current signal component, the historical delayed signal component, and the signal envelope features;

[0007] S3, input the multidimensional input feature vector into the scalar minimum gate unit layer for temporal processing. The scalar minimum gate unit layer calculates the hidden state at the current time based on the multidimensional input feature vector at the current input time, the hidden state at the previous time, and the stabilization state.

[0008] S4: Apply attention weighting to the hidden state at the current moment to obtain weighted features; simultaneously, extract instantaneous phase features from the baseband input signal; fuse the weighted features and instantaneous phase features to form a comprehensive feature vector;

[0009] S5 maps the integrated feature vector through the output layer to generate the pre-distorted output signal.

[0010] Furthermore, multidimensional input feature vectors The calculation formula is as follows:

[0011] ;

[0012] in, ; Indicates a fixed delay; Represents the real part of a complex signal; Represents the imaginary part of a complex signal; For delay The historical significance of the item; For instantaneous amplitude envelope; This is the cube term of the instantaneous amplitude envelope; This is the magnitude envelope term of the delayed m term.

[0013] Furthermore, the formula for calculating the hidden state at the current moment is as follows:

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] ;

[0019] in, Input at time t; For the current input The weight matrix; and Indicates the previous short-term state The weight matrix; For the corresponding bias term; This represents the exponential activation function; Represents a logarithmic function; Represents the hyperbolic tangent function; This is the hidden state from the previous moment; Indicates the current hidden state; Indicates the Gate of Oblivion; This indicates the stable state at the previous moment; Indicates a stable state; This indicates the normalized forgetting gate; ⊙ indicates a candidate hidden state; ⊙ indicates element-wise multiplication.

[0020] Furthermore, the weighted features and instantaneous phase features are fused to form a comprehensive feature vector, calculated as follows:

[0021] ;

[0022] ;

[0023] ;

[0024] in, This indicates the output of the core sMGU layer. Attention-weighted features Represents the real part of a complex signal; Represents the imaginary part of a complex signal; Represents the arctangent function; Indicates instantaneous phase; Represents the comprehensive feature vector; This indicates a matrix concatenation operation.

[0025] Furthermore, the synthesized feature vector is mapped through the output layer to generate the predistorted output signal, calculated as follows:

[0026] ;

[0027] in, and The weights and biases of the output layer are used to complete the end-to-end nonlinear inverse mapping; This represents the output signal after predistortion.

[0028] Furthermore, the digital predistortion method for power amplifiers based on minimum gated networks is treated as the inverse function of the nonlinear transfer function of the power amplifier, and the calculation formula is as follows:

[0029] ;

[0030] in, The original input signal; This is the output signal of the predistorter; This represents the system response function of the predistorter; For power amplifier nonlinear function The inverse function of .

[0031] On the other hand, a power amplifier digital predistortion system based on a minimum gate network includes:

[0032] The input signal acquisition module is used to acquire the baseband input signal at the current moment;

[0033] The multidimensional feature construction module is used to perform delay enhancement on the baseband input signal and construct a multidimensional input feature vector that includes the current signal component, the historical delayed signal component, and the signal envelope feature.

[0034] The hidden state calculation module is used to input the multi-dimensional input feature vector into the scalar minimum gate unit layer for temporal processing. The scalar minimum gate unit layer calculates the hidden state at the current time step based on the current input, the hidden state at the previous time step, and the stable state.

[0035] The comprehensive feature vector calculation module is used to perform attention weighting on the hidden state at the current moment to obtain weighted features; at the same time, it extracts instantaneous phase features from the baseband input signal; and it fuses the weighted features and instantaneous phase features to form a comprehensive feature vector.

[0036] The predistortion signal generation module is used to map the comprehensive feature vector through the output layer to generate the predistorted output signal.

[0037] The present invention adopts the above technical solution and has the following beneficial effects:

[0038] (1) This invention constructs a delay-enhanced input that integrates current and historical signal components and multi-order amplitude characteristics, and uses a scalar minimum gate unit for stable timing modeling, thereby achieving a high-precision characterization of the dynamic memory effect and strong nonlinear distortion of the power amplifier.

[0039] (2) By introducing a logarithmic exponential stabilization mechanism and a normalization gating strategy, this invention effectively suppresses numerical instability during training and improves convergence speed and robustness.

[0040] (3) This invention achieves accurate compensation for amplitude-phase coupling distortion by explicitly extracting and fusing instantaneous phase information and attention-weighted features, and then mapping through a lightweight output layer, so that the predistorter truly approximates the inverse function of the power amplifier's nonlinear characteristics. Attached Figure Description

[0041] Figure 1 This is a flowchart of the digital predistortion method for power amplifiers based on a minimum gate network, according to an embodiment of the present invention.

[0042] Figure 2 This is a real-valued time-delay scalar minimum gating network according to an embodiment of the present invention;

[0043] Figure 3This is the smallest gating unit layer in this embodiment of the invention;

[0044] Figure 4 This is the overall system training framework of this invention embodiment;

[0045] Figure 5 This is a diagram of a power amplifier digital predistortion system based on a minimum gate network, according to an embodiment of the present invention. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0047] like Figure 1 As shown, the present invention provides a power amplifier digital predistortion method based on a minimum gate network, comprising:

[0048] S1, obtain the baseband input signal at the current moment.

[0049] S2 performs delay enhancement on the baseband input signal to construct a multi-dimensional input feature vector that includes the current signal component, historical delayed signal components, and signal envelope features.

[0050] Specifically, multidimensional input feature vector The calculation formula is as follows:

[0051] ;

[0052] in, ; Indicates a fixed delay; Represents the real part of a complex signal; Represents the imaginary part of a complex signal; For delay The historical significance of the item; For instantaneous amplitude envelope; This is the cube term of the instantaneous amplitude envelope; This is the magnitude envelope term of the delayed m terms. This layer can be understood as an enhancement layer.

[0053] S3 inputs the multidimensional input feature vector to the scalar minimum gate unit layer for temporal processing. The scalar minimum gate unit calculates the hidden state at the current time step based on the current input, the hidden state at the previous time step, and the stable state.

[0054] Specifically, in S3, the formula for calculating the hidden state at the current moment is as follows:

[0055] ;

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] in, Input at time t; For the current input The weight matrix; and Indicates the previous short-term state The weight matrix; For the corresponding bias term; This represents the exponential activation function; Represents a logarithmic function; Represents the hyperbolic tangent function; This is the hidden state from the previous moment; Indicates the current hidden state; Indicates the Gate of Oblivion; This indicates the stable state at the previous moment; Indicates a stable state; This indicates the normalized forgetting gate; ⊙ indicates a candidate hidden state; ⊙ indicates element-wise multiplication.

[0061] Specifically, such as Figure 2 As shown, the real-valued time-delay scalar minimum gated network RVTD-sMGU according to an embodiment of the present invention mainly includes an input enhancement layer, a core sMGU layer, a lightweight attention module, a phase extraction and fusion layer, and an output layer. The input on the left side of the figure contains the real part, imaginary part, amplitude envelope, and cubic terms of the signal at the current and multiple historical moments, forming rich temporal and nonlinear features. The middle part is a lightweight minimum gated unit, which performs efficient temporal modeling of these features through a stabilization mechanism and uses an attention mechanism to highlight key time information. The phase information of the original signal is extracted separately on the right side, and its sine and cosine components are concatenated with the attention-weighted features. Finally, the real and imaginary parts of the predistorted output signal are directly generated through a fully connected layer.

[0062] Specifically, in this embodiment, the enhanced input features are... The data is fed into the scalar minimum gate unit (sMGU) layer. The sMGU is a lightweight loop unit designed in this invention, and its structure diagram is shown below. Figure 3As shown, the scalar minimum gated unit (Min-GatedUnit) constructs a multi-dimensional input vector containing current / historical I / Q components and multi-order envelope features through a delay enhancement strategy; it uses the minimum gate mechanism to stably model the temporal dependency and integrates attention-weighted features with explicitly extracted instantaneous phase information (sinϕ(n) and cosϕ(n)) to form a more physically meaningful comprehensive feature; finally, the output layer completes the end-to-end nonlinear inverse mapping.

[0063] S4: Attention weighting is applied to the hidden state at the current moment to obtain weighted features; at the same time, instantaneous phase features are extracted from the baseband input signal; the weighted features and instantaneous phase features are fused to form a comprehensive feature vector.

[0064] Specifically, the weighted features and instantaneous phase features are fused to form a comprehensive feature vector, calculated using the following formula:

[0065] ;

[0066] ;

[0067] ;

[0068] in, This indicates the output of the core sMGU layer. Attention-weighted features Represents the real part of a complex signal; Represents the imaginary part of a complex signal; Represents the arctangent function; Indicates instantaneous phase; Represents the comprehensive feature vector; This represents the matrix concatenation operation (lightweight attention module and phase extraction and fusion layer).

[0069] Specifically, the comprehensive feature vector is mapped through the output layer to generate the predistorted output signal, and the calculation formula is as follows:

[0070] ;

[0071] in, and The weights and biases of the output layer are used to complete the end-to-end nonlinear inverse mapping; This represents the output signal after predistortion.

[0072] S5 maps the integrated feature vector through the output layer to generate the pre-distorted output signal.

[0073] Specifically, the digital predistortion method for power amplifiers based on minimum gated networks is treated as the inverse function of the power amplifier's nonlinear transfer function, and the calculation formula is as follows:

[0074] ;

[0075] in, The original input signal; This is the output signal of the predistorter; This represents the system response function of the predistorter; For power amplifier nonlinear function The inverse function of .

[0076] The resulting nonlinear characteristics are designed to complement the inherent nonlinear characteristics of the power amplifier. Therefore, under ideal compensation conditions, the cascaded system of the predistorter and power amplifier can be equivalent to a linear system, and its input-output relationship can be simplified as follows:

[0077]

[0078] In the formula, This is the output signal of the power amplifier. This represents the ideal linear gain (constant value) of the power amplifier.

[0079] Specifically, the overall system training framework is as follows: Figure 4 As shown, the process of using ILC to obtain an ideal predistortion signal and then supervising the training of the RVTD-sMGU network can be divided into three stages.

[0080] Phase 1: Iterative learning controls the generation of ideal training data.

[0081] 1. Convert the original baseband signal The signal is fed into the PA, and its nonlinear output is collected. .

[0082] 2. Calculate the output With the expected linear output ( The error between (the expected gain) and (the expected gain)

[0083]

[0084] 3. Apply ILC algorithms (such as P-type update laws) to iteratively update the input signal of PA:

[0085]

[0086] in For the number of iterations, The learning gain is calculated. After several iterations, the input signal converges to the ideal predistortion signal. .

[0087] 4. This results in a high-quality training dataset:

[0088] Phase 2: Supervised training of the RVTD-sMGU network.

[0089] Using the dataset generated in the first phase ,by For input, To achieve the target output, the RVTD-sMGU network undergoes end-to-end supervised training. The mean squared error (MSE) loss function is used.

[0090]

[0091] in The network predicts the output. All weight parameters of the network are optimized using a backpropagation algorithm (such as the Adam optimizer).

[0092] Third stage: Feedforward execution linearization.

[0093] After training, the trained RVTD-sMGU network parameters are fixed and deployed as a feedforward digital predistorter. In normal operating mode, the signal to be transmitted... The network generates a predistorted signal in real time when directly input. This signal drives the PA to produce a highly linear output. This enables efficient power amplifier linearization, significantly improving performance metrics such as adjacent channel leakage ratio (ACPR) and normalized mean square error (NMSE).

[0094] like Figure 5 As shown, this embodiment also discloses a power amplifier digital predistortion system based on a minimum gate network, including:

[0095] The input signal acquisition module 51 is used to acquire the baseband input signal at the current moment;

[0096] The multidimensional feature construction module 52 is used to perform delay enhancement on the baseband input signal and construct a multidimensional input feature vector including the current signal component, the historical delayed signal component, and the signal envelope feature.

[0097] The hidden state calculation module 53 is used to input the multi-dimensional input feature vector into the scalar minimum gate unit layer for temporal processing. The scalar minimum gate unit calculates the hidden state at the current time based on the current input, the hidden state at the previous time step, and the stable state.

[0098] The comprehensive feature vector calculation module 54 is used to perform attention weighting on the hidden state at the current time to obtain weighted features; at the same time, it extracts instantaneous phase features from the baseband input signal; and it fuses the weighted features and instantaneous phase features to form a comprehensive feature vector.

[0099] The predistortion signal generation module 55 is used to map the comprehensive feature vector through the output layer to generate the predistorted output signal.

[0100] The specific implementation of the power amplifier digital predistortion system based on minimum gate network is the same as that of the power amplifier digital predistortion method based on minimum gate network, and will not be described again in this embodiment.

[0101] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A digital predistortion method for power amplifiers based on a minimum gate network, characterized in that, Includes the following steps: S1, Obtain the baseband input signal at the current moment; S2, delay enhancement is performed on the baseband input signal to construct a multi-dimensional input feature vector including the current signal component, the historical delayed signal component, and the signal envelope features; S3, input the multidimensional input feature vector into the scalar minimum gate unit layer for temporal processing. The scalar minimum gate unit layer calculates the hidden state at the current time based on the multidimensional input feature vector at the current input time, the hidden state at the previous time, and the stabilization state. S4: Apply attention weighting to the hidden state at the current moment to obtain weighted features; simultaneously, extract instantaneous phase features from the baseband input signal; fuse the weighted features and instantaneous phase features to form a comprehensive feature vector; S5 maps the integrated feature vector through the output layer to generate the predistorted output signal.

2. The digital predistortion method for power amplifiers based on minimum gated networks according to claim 1, characterized in that, In S2, the multidimensional input feature vector The calculation formula is as follows: ； in, ; Indicates a fixed delay; Represents the real part of a complex signal; Represents the imaginary part of a complex signal; For delay The historical significance of the item; For instantaneous amplitude envelope; This is the cube term of the instantaneous amplitude envelope; This is the magnitude envelope term of the delayed m term.

3. The digital predistortion method for power amplifiers based on a minimum gate network according to claim 1, characterized in that, In S3, the formula for calculating the hidden state at the current moment is as follows: ； ； ； ； ； in, Input at time t; For the current input The weight matrix; and Indicates the previous short-term state The weight matrix; For the corresponding bias term; This represents the exponential activation function; Represents a logarithmic function; Represents the hyperbolic tangent function; This is the hidden state from the previous moment; Indicates the current hidden state; Indicates the Gate of Oblivion; This indicates the stable state at the previous moment; Indicates a stable state; This indicates the normalized forgetting gate; ⊙ indicates a candidate hidden state; ⊙ indicates element-wise multiplication.

4. The digital predistortion method for power amplifiers based on minimum gated networks according to claim 1, characterized in that, In S4, the weighted features and instantaneous phase features are fused to form a comprehensive feature vector, calculated as follows: ； ； ； in, This indicates the output of the core sMGU layer. Attention-weighted features Represents the real part of a complex signal; Represents the imaginary part of a complex signal; Represents the arctangent function; Indicates instantaneous phase; Represents the comprehensive feature vector; This indicates a matrix concatenation operation.

5. The digital predistortion method for power amplifiers based on a minimum gate network according to claim 4, characterized in that, The synthesized feature vector is mapped through the output layer to generate the predistorted output signal. The calculation formula is as follows: ； in, and The weights and biases of the output layer are used to complete the end-to-end nonlinear inverse mapping; This represents the output signal after predistortion.

6. The digital predistortion method for power amplifiers based on minimum gated networks according to claim 1, characterized in that, Also includes: The digital predistortion method for power amplifiers based on minimum gated networks is treated as the inverse function of the power amplifier's nonlinear transfer function, and the calculation formula is as follows: ； in, The original input signal; This is the output signal of the predistorter; This represents the system response function of the predistorter; For power amplifier nonlinear function The inverse function of .

7. A system for digital predistortion of power amplifiers based on a minimum gate network, characterized in that, include: The input signal acquisition module is used to acquire the baseband input signal at the current moment; The multidimensional feature construction module is used to perform delay enhancement on the baseband input signal and construct a multidimensional input feature vector that includes the current signal component, the historical delayed signal component, and the signal envelope feature. The hidden state calculation module is used to input the multi-dimensional input feature vector into the scalar minimum gate unit layer for temporal processing. The scalar minimum gate unit layer calculates the hidden state at the current time based on the current input, the hidden state at the previous time step, and the stable state. The comprehensive feature vector calculation module is used to perform attention weighting on the hidden state at the current moment to obtain weighted features; at the same time, it extracts instantaneous phase features from the baseband input signal; and it fuses the weighted features and instantaneous phase features to form a comprehensive feature vector. The predistortion signal generation module is used to map the comprehensive feature vector through the output layer to generate the predistorted output signal.

Images