A method and system for compensating for nonlinear distortion in the terahertz frequency band

Abstract

The present application relates to the field of wireless communication, especially to a kind of terahertz band nonlinear distortion compensation method and system, method includes, real-time acquisition power amplifier output end radio frequency signal, from it extracts the characteristic parameter reflecting nonlinear distortion state and generates the state vector of current time;State vector is input to the preset reinforcement learning network, and the compensation parameter vector of current time is output by the network;Compensation parameter vector is converted into pre-distortion coefficient by preset mapping relationship, and the coefficient is used to execute digital pre-distortion processing to input baseband signal;After pre-distortion processing, the signal is input to power amplifier, and its output signal is collected;Based on output signal, generate feedback information, and adjust the subsequent decision of reinforcement learning network according to feedback information;Finally complete the compensation of nonlinear distortion.The present application realizes the dynamic self-adaptive adjustment of pre-distortion parameter, and is suitable for the linearization processing of the power amplifier of the sixth generation mobile communication terahertz band.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a method and system for nonlinear distortion compensation in the terahertz band. Background Technology

[0002] Terahertz frequency bands, with their ultra-wide bandwidth, have become a key technology direction for sixth-generation mobile communication systems. However, terahertz power amplifiers exhibit significant nonlinear distortion during operation, leading to signal quality degradation and severely impacting the bit error rate and transmission efficiency of communication systems. Digital predistortion techniques are widely used to compensate for this nonlinear distortion. Existing terahertz nonlinear distortion compensation methods typically employ fixed parameters or semi-adaptive parameter architectures based on pre-stored static characteristics. These methods often exhibit significant fluctuations in compensation effectiveness when channel conditions, operating temperature, or power amplifier aging conditions change. Furthermore, existing methods have long parameter adjustment cycles or require accumulating a large number of output signal samples to complete an effective parameter update, making it difficult to match the rapid time-varying characteristics of terahertz channels.

[0003] In nonlinear distortion compensation systems, the precision of feature extraction directly affects the effectiveness of subsequent compensation parameters. Existing technologies often use simple power envelope detection or single-dimensional amplitude statistics to characterize the distortion state. When the power amplifier enters the saturation region or the bias conditions change significantly, a single statistic is insufficient to distinguish distortion differences within different amplitude ranges, resulting in insufficient responsiveness of the compensation parameters to local distortion features. Furthermore, existing parameter update mechanisms mostly employ fixed-step iterative methods, with each update performed independently. This fails to leverage historical adjustment experience to accelerate the convergence process. When channel conditions or power amplifier operating points change drastically, numerous iterations are required for reconvergence, making it difficult to meet the sub-millisecond adaptive adjustment requirements of sixth-generation mobile communication systems. Summary of the Invention

[0004] To address one or more problems in the prior art, the main objective of this application is to provide a method and system for nonlinear distortion compensation in the terahertz band.

[0005] To achieve the aforementioned objectives, this application proposes a terahertz band nonlinear distortion compensation method, the method comprising: The radio frequency signal at the output of the power amplifier is acquired in real time. Based on the radio frequency signal, characteristic parameters reflecting the nonlinear distortion state are extracted, and the state vector at the current moment is generated based on the characteristic parameters. The current state vector is input into a preset reinforcement learning network, and the current compensation parameter vector is output through the reinforcement learning network. Based on the output, the compensation parameter vector is converted into predistortion coefficients through a preset mapping relationship, and the predistortion coefficients are used to perform digital predistortion processing on the input baseband signal; Based on the processing results, the pre-distorted signal is input into the power amplifier, and the corresponding output signal of the power amplifier is acquired. Feedback information is generated based on the collected output signals, and the subsequent decisions of the reinforcement learning network are adjusted according to the feedback information. Based on the adjustment results, compensation for nonlinear distortion is completed.

[0006] This application also provides a terahertz band nonlinear distortion compensation system, including: The acquisition module is used to acquire the radio frequency signal at the output of the power amplifier in real time, extract the feature parameters reflecting the nonlinear distortion state based on the radio frequency signal, and generate the state vector at the current moment based on the feature parameters. The first input module is used to input the current state vector into a preset reinforcement learning network and output the compensation parameter vector at the current time through the reinforcement learning network. The conversion module is used to convert the compensation parameter vector into predistortion coefficients based on the output result through a preset mapping relationship, and to perform digital predistortion processing on the input baseband signal using the predistortion coefficients; The second input module is used to input the pre-distorted signal into the power amplifier according to the processing result, and to acquire the corresponding output signal of the power amplifier. An adjustment module is used to generate feedback information based on the collected output signal and adjust the subsequent decisions of the reinforcement learning network according to the feedback information. The completion module is used to compensate for nonlinear distortion based on the results of the adjustment.

[0007] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0008] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0009] The terahertz band nonlinear distortion compensation method and system of this application acquires the power amplifier output signal in real time and extracts a refined state vector, enabling the system to keenly perceive subtle changes in distortion characteristics. By leveraging the continuous action decision-making capability of a reinforcement learning network, a compensation parameter vector adapted to the current state can be quickly generated, avoiding the slow convergence speed problem of traditional iterative search algorithms. Through a closed-loop feedback adjustment mechanism, the results of each compensation step can be used to optimize subsequent decisions, forming a positive loop. This ensures that the compensation parameters no longer rely on offline measurement and static storage, but dynamically evolve with changes in the power amplifier's operating state, thus maintaining high compensation accuracy in the rapidly time-varying environment of the terahertz band. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a terahertz band nonlinear distortion compensation method according to an embodiment of this application. Figure 2 This is a flowchart illustrating a terahertz band nonlinear distortion compensation method according to an embodiment of this application. Figure 3 This is a schematic block diagram of a terahertz band nonlinear distortion compensation system according to an embodiment of this application; Figure 4 This is a schematic block diagram of the structure of a computer device according to an embodiment of this application.

[0011] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0013] Reference Figure 1 This application provides a terahertz band nonlinear distortion compensation method, the method comprising: S1. Real-time acquisition of the radio frequency signal at the output of the power amplifier; extraction of characteristic parameters reflecting the nonlinear distortion state based on the radio frequency signal; generation of the current state vector based on the characteristic parameters. S2. Input the current state vector into a preset reinforcement learning network, and output the compensation parameter vector at the current time through the reinforcement learning network; S3. Based on the output results, the compensation parameter vector is converted into predistortion coefficients through a preset mapping relationship, and the predistortion coefficients are used to perform digital predistortion processing on the input baseband signal; S4. Based on the processing result, input the pre-distorted signal into the power amplifier and collect the corresponding output signal of the power amplifier; S5. Generate feedback information based on the collected output signal, and adjust the subsequent decisions of the reinforcement learning network according to the feedback information; S6. Based on the adjustment results, complete the compensation for nonlinear distortion.

[0014] As described in steps S1-S3 above, the first step involves acquiring the RF signal at the power amplifier output in real time via a high-speed analog-to-digital converter. This acquisition process is performed at a rate of no less than 100 million samples per second to ensure the capture of the complete envelope information of the terahertz band signal. After quadrature demodulation, the acquired RF signal yields in-phase and quadrature components. Subsequently, the system extracts characteristic parameters reflecting the current nonlinear distortion state of the power amplifier from these components. For example, it can calculate the mapping deviation between the instantaneous amplitude of the signal and the amplitude of the input baseband signal, as well as the degree of rotation of the instantaneous phase relative to the phase of the input signal. It can also statistically analyze the slope changes of the distortion curve in different amplitude ranges. Normalizing these multidimensional characteristic parameters constitutes the state vector at the current moment. This state vector is equivalent to a digital snapshot of the current operating state of the power amplifier, which can more meticulously characterize the local features of distortion than simply monitoring the output power or bit error rate. The second step involves inputting the state vector into a pre-trained reinforcement learning network. This network employs a deep reinforcement learning architecture, enabling it to make decisions in a continuous action space. Unlike traditional fixed-parameter lookup tables or simple linear mappings, reinforcement learning networks can infer the most suitable compensation parameters for the current power amplifier state based on the input state vector. The network output is a vector of continuous values, where each element represents the adjustment amount of a certain dimension of the predistortion model. Since the network has learned a large number of power amplifier distortion characteristics under different operating conditions during training, only one forward computation is needed to obtain the compensation parameters during actual operation, eliminating the need for iterative search and thus controlling the inference latency to the sub-millisecond level. The third step is the conversion and predistortion processing of predistortion coefficients. The compensation parameter vector output by the reinforcement learning network is not a directly usable predistortion coefficient because the former usually has a low dimension, while the polynomial model that actually performs predistortion processing requires more coefficients. Therefore, the system converts the low-dimensional compensation parameter vector into high-dimensional predistortion coefficients through a preset mapping relationship. This mapping relationship can be linear, for example, mapping five input parameters to seven polynomial coefficients through a trainable transformation matrix. After the conversion, the system uses these coefficients to perform digital predistortion processing on the input baseband signal. The essence of a predistorter is to construct a transfer function that is opposite to the nonlinear characteristics of a power amplifier. When the original signal passes through the predistorter, its amplitude and phase are pre-distorted. When it is sent to the power amplifier, the nonlinear effect of the power amplifier cancels out this distortion, thus obtaining a linearly amplified signal at the output of the power amplifier.

[0015] As described in steps S4-S6 above, the fourth step involves power amplifier processing and output acquisition of the predistorted signal. The predistorted baseband signal is up-converted to a terahertz carrier frequency and fed into the power amplifier via a digital-to-analog converter (DAC). After amplification, the system again acquires the RF signal from the power amplifier output via a high-speed DAC. This acquisition operation is similar to that in the first step, but its purpose is to obtain the actual output result after the current predistortion compensation. The fifth step calculates indicators that measure the degree of signal distortion, such as error vector magnitude or adjacent channel leakage ratio, based on the acquired power amplifier output signal. These indicators are used as feedback information to quantitatively evaluate the effectiveness of the current predistortion compensation. Then, the system adjusts the subsequent decision-making behavior of the reinforcement learning network based on this feedback information. Specifically, if the feedback information indicates low distortion, it means the current compensation parameters are effective, and the network will reinforce the tendency to generate similar parameters; if the feedback information indicates high distortion, the network will try to generate different parameters. This adjustment process does not require manually set rules but is automatically completed by the reinforcement learning algorithm, allowing the compensation parameters to be continuously updated as the power amplifier characteristics drift slowly or change rapidly. Step six: Compensation completion. After the above five steps are executed cyclically, the system continuously adjusts the output of the reinforcement learning network based on the latest feedback information, ensuring that the pre-distortion coefficients always match the current nonlinear state of the power amplifier. When the distortion index indicated by the feedback information stabilizes within the preset acceptable range, the nonlinear distortion compensation under the current operating condition is considered complete. As the communication process continues, the system will continue to monitor the power amplifier output; if the distortion index deteriorates again, a new round of parameter adjustments will be automatically triggered.

[0016] As described above, compared to existing predistortion methods that use fixed or semi-adaptive parameters, this embodiment introduces a reinforcement learning-driven dynamic parameter adjustment mechanism. Existing technologies typically pre-calculate a set of predistortion coefficients based on the static characteristics of the power amplifier and maintain them unchanged or only perform coarse segmented switching during actual communication. This approach struggles to adapt to the rapid drift of terahertz power amplifier nonlinear characteristics due to changes in temperature, bias voltage, aging, and channel environment. When operating conditions change, the compensation effect of the fixed-parameter predistorter deteriorates drastically, leading to significant fluctuations in the error vector amplitude. This embodiment, by acquiring the power amplifier output signal in real time and extracting a refined state vector, enables the system to sensitively perceive subtle changes in distortion characteristics. Through the continuous action decision-making capability of the reinforcement learning network, a compensation parameter vector adapted to the current state can be quickly generated, avoiding the slow convergence speed problem of traditional iterative search algorithms. Through a closed-loop feedback adjustment mechanism, the results of each compensation can be used to optimize subsequent decisions, forming a positive cycle. This allows the compensation parameters to evolve dynamically as the power amplifier's operating state changes, rather than relying on offline measurements and static storage, thus maintaining high compensation accuracy in the rapidly changing environment of the terahertz band.

[0017] Reference Figure 2 In one embodiment, the step of extracting feature parameters reflecting the nonlinear distortion state includes: S11. Perform orthogonal decomposition on the acquired power amplifier output signal to obtain in-phase signal components and quadrature signal components; S12. Perform Hilbert transform on the in-phase signal component and the quadrature signal component to calculate the instantaneous amplitude and instantaneous phase values ​​of the signal; S13. Generate an amplitude distortion characteristic curve based on the correspondence between the instantaneous amplitude value and the amplitude value of the input baseband signal; S14. Generate a phase distortion characteristic curve based on the deviation between the instantaneous phase value and the phase value of the input baseband signal; S15. Divide the amplitude range of the input signal into multiple continuous amplitude intervals, and calculate the slope change of the amplitude distortion characteristic curve and the phase rotation of the phase distortion characteristic curve in each amplitude interval. S16. Combine the slope change and phase rotation calculated in each amplitude interval into a set of feature parameters to complete the feature parameter extraction.

[0018] As described above, the first step involves orthogonally demodulating the acquired power amplifier output RF signal, decomposing it into in-phase and quadrature signal components. These two components are perpendicular to each other and together completely describe the amplitude and phase information of the RF signal. Through orthogonal decomposition, the RF carrier signal, which was originally difficult to process directly, is converted into a complex signal in baseband form, creating conditions for subsequent Hilbert transform and instantaneous parameter calculation. The second step involves performing a Hilbert transform on these two components after obtaining the in-phase and quadrature components. The essence of the Hilbert transform is to construct an analytic signal that allows the instantaneous envelope and instantaneous phase of the signal to be accurately extracted. Through this transformation, the system can calculate the instantaneous amplitude and instantaneous phase values ​​of the signal at each moment. The instantaneous amplitude reflects the signal strength at the current moment, while the instantaneous phase reflects the current phase position of the signal waveform. These two parameters are the basis for subsequent analysis of the power amplifier's nonlinear characteristics. The third step involves comparing the calculated instantaneous amplitude value with the amplitude value of the input baseband signal point by point. Under ideal linear amplification, the output instantaneous amplitude should be proportional to the input amplitude. However, due to the nonlinear characteristics of the power amplifier, the output amplitude often compresses when the input amplitude is large. Plotting the input amplitude on the x-axis and the relationship between the instantaneous output amplitude and the input amplitude on the y-axis yields the amplitude distortion characteristic curve. This curve visually demonstrates the change in amplification gain of the power amplifier for different input power levels, especially when the input signal approaches or exceeds the power amplifier's saturation point, the curve becomes noticeably curved. The fourth step, similar to amplitude distortion, involves the power amplifier's nonlinear effects causing phase distortion. Subtracting the calculated instantaneous phase value from the phase value of the input baseband signal yields the phase deviation. Plotting the input amplitude on the x-axis and the phase deviation on the y-axis yields the phase distortion characteristic curve. This curve reflects the phase modulation effect of the power amplifier on the input signal. In an ideal power amplifier, the output phase should maintain a constant difference from the input phase, but due to the variation of parasitic parameters such as capacitance and inductance within the power amplifier with the operating point, significant phase rotation often occurs when the input amplitude is large. The fifth step involves statistically analyzing the distortion characteristics in different intervals. Obtaining only two distortion curves is insufficient, as the nonlinear characteristics of the power amplifier exhibit different behaviors in different input amplitude ranges. For example, in the low-power range, the power amplifier typically exhibits quasi-linear characteristics, with gain approaching a constant; in the medium-power range, the gain begins to compress slowly; and in the high-power range, the gain drops sharply and the phase rotates rapidly. If only an average slope is extracted from the entire curve, these local differences will be lost. Therefore, this implementation divides the amplitude range of the input signal into multiple consecutive amplitude intervals, calculating the slope change of the amplitude distortion characteristic curve and the phase rotation of the phase distortion characteristic curve within each interval. The slope change reflects the degree of gain compression within that interval, while the phase rotation reflects the severity of the phase shift within that interval.This interval-based statistical approach allows the system to accurately determine at which power level the power amplifier experiences the most severe distortion, and whether the distortion is primarily amplitude compression or phase rotation. The sixth step combines the slope changes and phase rotation values ​​calculated for each amplitude interval in a preset order, forming a set of feature parameters. This set is actually a multi-dimensional vector, where each dimension corresponds to a specific distortion metric within a particular power interval. Using this set as the state vector input to the subsequent reinforcement learning network, the network can obtain a comprehensive and detailed description of the power amplifier's current nonlinear distortion state.

[0019] In one embodiment, the step of inputting the current state vector into a preset reinforcement learning network and outputting the compensation parameter vector at the current time through the reinforcement learning network includes: The current state vector is input into the Actor network, which contains three fully connected layers. Each fully connected layer uses the LeakyReLU activation function. The original action vector is output through the Actor network. Based on the output, motion noise is added to the original motion vector to obtain a smoothed compensation parameter vector, which is used as the output of the reinforcement learning network. The current state vector and the compensation parameter vector are input together into the first Critic network to obtain the first Q-value estimate; the current state vector and the compensation parameter vector are input together into the second Critic network to obtain the second Q-value estimate. The smaller value between the first Q-value estimate and the second Q-value estimate is selected as the target Q-value, which is used for the delay policy update of the reinforcement learning network.

[0020] As described above, the first step involves feeding the current state vector into a deep neural network called the Actor network. This Actor network contains three fully connected layers, each using LeakyReLU as the activation function. Unlike the traditional ReLU function, which outputs zero directly when the input is negative, LeakyReLU retains a small slope in the negative interval. This ensures that the network does not completely lose its gradient flow capability even when processing negative inputs, which is particularly beneficial for negative eigenvalues ​​that may appear in the nonlinear compensation of terahertz power amplifiers. After receiving the state vector, the Actor network performs multiple nonlinear transformations and outputs a raw action vector. This raw action vector can be understood as the initial value of the compensation parameters that the network considers optimal in the current state, but it has not yet undergone any smoothing processing. The second step involves a potential problem in directly using the raw action vector output by the Actor network as the compensation parameter: if the network's estimation of the optimal action is not accurate enough in certain states, the output value may fluctuate drastically, causing the predistortion coefficient to fluctuate violently, which in turn worsens the compensation effect. Therefore, this implementation introduces an action noise addition stage after the Actor network output. The system superimposes a pre-defined noise term onto the original action vector to obtain a smoothed compensation parameter vector, which is then used as the compensation parameter for the final output of the reinforcement learning network. This noise addition strategy has two advantages. First, moderate noise can prevent the compensation parameters from getting stuck near a local optimum for a long time, helping the system explore better parameter combinations. Second, the noise acts as a low-pass filter, preventing drastic changes in output parameters between adjacent time steps, ensuring the smoothness of pre-distortion coefficient updates. Third, to evaluate the quality of actions generated by the Actor network, the system introduces two Critic networks, referred to as the first Critic network and the second Critic network. These two Critic networks have similar structures to the Actor network, but their tasks differ. Each Critic network receives two inputs: the current state vector and the current compensation parameter vector. The network outputs a scalar value, called the Q-value estimate, representing the long-term cumulative gain obtained by performing the action in the current state. A higher Q-value means that the compensation parameter combination is expected to bring less distortion and a more stable output. Using two Critic networks instead of one is to address the Q-value overestimation problem common in traditional reinforcement learning methods. Because the two networks use different initialization parameters and training samples, their Q-value estimates for the same action often differ, and this difference precisely reflects the uncertainty of the estimate. In the fourth step, after obtaining two Q-value estimates from the first and second Critic networks respectively, instead of averaging them, the smaller value is selected as the final target Q-value.This smaller-scale operation is a conservative estimation strategy that automatically ignores potential overestimations in both networks, adopting only the more cautious evaluation. The target Q-value is then used to guide the update direction of the Actor network. Furthermore, this implementation employs a delayed update mechanism: the Critic network can update more frequently, while the Actor network waits for the Critic network to update several times before updating itself. This staggered timing design prevents the Actor network from blindly following the wrong direction before the Critic network has stabilized, thereby improving the stability and convergence speed of the entire learning process.

[0021] Specifically, this embodiment constitutes a stable and efficient reinforcement learning decision-making closed loop. Existing technologies, when applying reinforcement learning to power amplifier nonlinear compensation, typically employ a single Critic network and an instantaneous update strategy. This exposes two prominent problems in the noisy, rapidly changing environment of the terahertz band. First, a single Critic network is prone to overestimating the Q-value of actions, leading to overconfidence and incorrect compensation parameter selection, and this error reinforces itself during the update process. Second, when the Actor network and Critic network update synchronously, if the Critic network's evaluation is inaccurate, the Actor network will adjust in the wrong direction, causing policy oscillations and jumps between compensation parameters at different times, introducing additional distortion. This embodiment effectively suppresses the risk of Q-value overestimation through dual Critic networks and a minimum-value operation. The two networks estimate independently, and the minimum value is taken as the target, effectively providing a safety boundary for decision-making. Even if one network temporarily deviates significantly, the lower estimate of the other network will take over, preventing the system from taking overly aggressive compensation actions. Meanwhile, the delayed update strategy ensures that the Actor network is only updated when the Critic network is relatively mature, reducing the number of ineffective or even harmful adjustments and improving sample utilization efficiency. Furthermore, the addition of action noise plays a crucial role. In traditional methods, parameter optimization often gets stuck in local optima; for example, the amplifier might be well compensated in one power range but still severely distorted in another, and the system cannot escape this suboptimal state. Action noise provides continuous exploration momentum, allowing the system to try different parameter combinations and thus discover a better global compensation scheme. Simultaneously, the smoothing effect of noise makes the changes in compensation parameters more continuous over time, avoiding the impact of abrupt changes in pre-distortion coefficients on signal quality.

[0022] In one embodiment, each fully connected layer of the Actor network, the first Critic network, and the second Critic network has 128 nodes; the negative slope coefficient of the LeakyReLU activation function is 0.2. This embodiment sets each fully connected layer in the Actor network, the first Critic network, and the second Critic network to contain 128 nodes. This value is not arbitrarily chosen, but rather a compromise determined after comprehensively considering the complexity of terahertz power amplifier nonlinear distortion modeling, the latency constraints of real-time inference, and the carrying capacity of hardware resources. If the number of nodes is too small, such as only 32 or 64, the network's expressive power will be limited, making it difficult to fit the complex nonlinear mapping relationships of the power amplifier under different input power ranges, different temperatures, and different aging states. The state vector contains feature information in up to a dozen dimensions, and these features have nonlinear interactions, requiring a sufficient number of neurons to construct sufficiently rich feature combinations. On the other hand, if the number of nodes is too large, such as reaching 256 or 512, although the expressive power is further enhanced, the amount of multiplication and accumulation operations required for each forward inference will increase exponentially, causing the network output latency to exceed the sub-millisecond constraint. On a real hardware platform, 128 nodes working in a three-layer structure can complete a full forward computation within tens of microseconds, while retaining sufficient nonlinear fitting capability to handle various distortion modes of terahertz power amplifiers. Secondly, the setting of the negative slope coefficient of the LeakyReLU activation function. In this implementation, the negative slope coefficient of the LeakyReLU activation function is set to 0.2. LeakyReLU is a modified linear unit; when the input is positive, the output equals the input, and when the input is negative, the output equals the input multiplied by a very small slope coefficient. This slope coefficient determines how much gradient information the network retains in the negative region. If the slope coefficient is set to zero, LeakyReLU degenerates into the traditional ReLU function; negative inputs are completely truncated, and the corresponding neurons no longer have any effect on subsequent computations. This leads to a phenomenon known as neuron death, where some neurons, once falling into the negative region, can no longer regain activity, and the network's expressive power gradually decreases. Setting the slope coefficient to 0.2 means that negative inputs can still propagate forward with one-fifth of their intensity, and the gradients of neurons will not completely disappear. The value of 0.2 is slightly larger than the common 0.01 or 0.1 because in the scenario of terahertz power amplifier compensation, some elements in the feature vector may frequently oscillate between positive and negative values. A larger negative slope ensures that this fluctuation information is not filtered out, thus helping the network to capture subtle changes in the sign of the distortion features. For example, the phase distortion characteristic curve may show negative phase rotation in certain power ranges. If the negative slope is too small, this negative information will be almost discarded, and the network will be unable to distinguish the different consequences of positive and negative rotation directions.

[0023] In one embodiment, the step of converting the compensation parameter vector into predistortion coefficients through a preset mapping relationship includes: Obtain the compensation parameter vector output by the reinforcement learning network; The compensation parameter vector is input into a differentiable mapping layer, which contains a trainable linear transformation matrix; A linear transformation is performed on the compensation parameter vector using the linear transformation matrix to obtain predistortion coefficients, wherein the number of predistortion coefficients is greater than the number of elements in the compensation parameter vector.

[0024] As described above, the first step is to obtain the compensation parameter vector output at the current time step from the reinforcement learning network. This vector has a low dimension, typically containing five consecutive numerical elements. The reason for designing a low dimension is that the reinforcement learning network needs to search for the optimal policy in a continuous action space. Each increase in the dimension of the action space drastically increases the search complexity, the number of samples required for training, and the convergence time. Using a low-dimensional compensation parameter vector can significantly reduce the training difficulty and inference latency of the reinforcement learning network, enabling it to complete a decision in the sub-millisecond range. The second step is to input the low-dimensional compensation parameter vector into a module called a differentiable mapping layer. This mapping layer contains a trainable linear transformation matrix. Differentiability means that the parameters of this transformation matrix can be jointly trained end-to-end with the reinforcement learning network using the error backpropagation algorithm. In other words, during training, the system not only adjusts the weights of the Actor and Critic networks but also simultaneously optimizes the values ​​of this mapping matrix, allowing the conversion from compensation parameters to predistortion coefficients to adaptively adapt to the characteristics of the current amplifier. The third step involves the differentiable mapping layer performing matrix multiplication on the input compensation parameter vector using a linear transformation matrix, outputting a higher-dimensional predistortion coefficient vector. The new vector has more elements than the original compensation parameter vector, for example, expanding from five input elements to seven or more output coefficients. These output coefficients directly correspond to the coefficients in the predistortion polynomial model, such as the static gain term, the third-order nonlinear term, and the memory effect term. The mathematical essence of the linear transformation is that the elements of the input vector are combined with different weights to generate each element of the output vector; each output coefficient is the sum of the input elements multiplied by different weights. This decouples the decision dimension from the execution dimension. Reinforcement learning networks only need to learn to output a few abstract control parameters, rather than directly outputting all polynomial coefficients. The mapping layer is responsible for concretizing these abstract parameters, translating low-dimensional decision instructions into high-dimensional execution instructions using trained linear combination rules. Since the linear transformation matrix is ​​trainable, the mapping layer is not just a fixed decoder. It can learn during training how to map abstract compensation parameters to specific polynomial coefficients. This allows reinforcement learning networks to focus on adjusting a few core control variables without worrying about the specific form of the predistortion model.

[0025] This implementation introduces a differentiable mapping layer, compressing the high-dimensional action space that originally required direct learning into a low-dimensional abstract action space, and then restoring it to high-dimensional coefficients through a trainable linear transformation. This approach firstly reduces the learning difficulty of the reinforcement learning network. The network only needs to search for the optimal policy in a five-dimensional continuous space, and the complexity of a five-dimensional space is far lower than that of a seven-dimensional or even higher-dimensional space. This allows the network to converge with fewer training samples and in a shorter training time. Secondly, it maintains the expressive power of the predistortion model. The number of predistortion coefficients in the final output is unlimited and can still be set to seven, nine, or higher dimensions as needed, thus ensuring that the multinomial model has sufficient degrees of freedom to fit complex nonlinear distortions. Furthermore, the joint training of the differentiable mapping layer and the reinforcement learning network ensures end-to-end optimality of the entire system. The mapping matrix is ​​not a fixed value set manually, but is automatically optimized through error backpropagation during training, thus enabling adaptive adjustment of the mapping rules according to the specific amplifier characteristics and predistortion model structure.

[0026] In one embodiment, the step of adjusting the subsequent decisions of the reinforcement learning network based on the feedback information includes: The error vector amplitude is calculated based on the acquired power amplifier output signal, and the error vector amplitude is used as feedback information. The current state vector, compensation parameter vector, feedback information, and the state vector of the next moment are combined into a data record and stored in a preset buffer pool; Data records are read from the buffer pool and sampled according to the priority of each data record. The priority is determined based on the absolute value of the difference between the output value of the first Critic network and the output value of the second Critic network corresponding to the compensation parameter vector in the data record. Gradient updates are performed on the Actor and Critic networks of the reinforcement learning network based on the sampled data records.

[0027] As described above, the first step involves calculating the error vector magnitude based on the acquired power amplifier output signal. The error vector magnitude is a comprehensive parameter measuring the difference between the actual transmitted signal and the ideal reference signal, reflecting both amplitude and phase errors. The system uses this value as feedback information to evaluate the effectiveness of the current pre-distortion compensation. A smaller error vector magnitude indicates more accurate compensation; conversely, a larger magnitude indicates insufficient compensation. Compared to simple power detection or signal-to-noise ratio estimation, the error vector magnitude more comprehensively reflects the impact of nonlinear distortion on signal quality, making it highly suitable as a reward signal source in reinforcement learning. The second step involves combining four key pieces of information into a complete data record: the current state vector, the current compensation parameter vector, the current feedback information (error vector magnitude), and the state vector for the next time step. These four elements constitute an empirical sample within the standard reinforcement learning framework, fully describing a transition from state to action, then to reward, and finally to a new state. The system stores this data record in a pre-defined buffer. The buffer is a circular storage area that can hold a certain amount of historical experience. By preserving a large amount of past experience, the system can repeatedly utilize this data in subsequent training, rather than discarding old samples after each update. This experience reuse mechanism greatly improves sample efficiency, enabling the network to learn as much knowledge as possible from limited online running data. In the third step, when reading data records from the buffer pool, instead of simple random sampling or a first-in-first-out order, biased sampling is performed based on the priority of each data record. The priority is calculated based on the absolute value of the difference between the output values ​​of the first and second Critic networks corresponding to the compensation parameter vector in that data record. Recalling the reinforcement learning decision-making methods described earlier, the Q-value estimates of the two Critic networks for the same action are usually different. When the difference is large, it indicates that the current network's evaluation of the action is highly uncertain. This uncertainty often occurs when the system encounters states or action combinations that are rarely seen before. Such experience samples have higher learning value for improving the network because the network needs to correct its judgments on these unfamiliar situations as quickly as possible. Conversely, if the outputs of the two Critic networks are very close, it indicates that the network has a relatively certain understanding of the good or bad nature of the action, and the priority of such samples can be relatively reduced. By setting sampling priorities based on the differences in the outputs of the two Critic networks, the system can automatically identify and prioritize learning the most informative experiences, thereby accelerating the learning process. The fourth step involves performing gradient updates on the Actor and Critic networks in the reinforcement learning network based on the sampled data records. Specifically, for the Critic network, the system calculates the temporal difference error between the target Q-value and the current Q-value, and then adjusts the network weights in the direction that reduces this error.For the Actor network, the system adjusts the weights in directions that increase the Critic network output, so that the network tends to output compensation parameters that yield higher Q-value evaluations. These two types of updates are performed alternately, gradually optimizing the network's decision-making ability.

[0028] It should be noted that the reinforcement learning network and the linear transformation matrix in the differentiable mapping layer used in this method require a certain amount of training to achieve effective decision-making capabilities. This implementation adopts a hybrid training strategy combining offline pre-training and online fine-tuning. In the offline pre-training stage, terahertz power amplifier models under various operating conditions are first constructed, including different bias voltage settings, different chip operating temperatures, and different device aging degrees. For each operating condition, the system randomly generates a large number of input baseband signals, obtains the corresponding output signals through power amplifier model simulation, and calculates the state vector and error vector amplitude according to the aforementioned feature extraction method. These data constitute empirical samples, each containing a state vector, action, reward, and the state vector at the next time step. The system uses these samples to jointly train the Actor network, the dual Critic network, and the linear transformation matrix in the differentiable mapping layer. The training objective is to minimize the error vector amplitude in the compensation parameters output by the Actor network, i.e., maximize the cumulative reward. After sufficient offline training, the network can learn how to adjust the predistortion coefficients under various common operating conditions. After the system is actually deployed and running, it enters the online fine-tuning stage. At this point, the system no longer relies on simulation models but instead uses feedback information generated by the actual power amplifier output signal. Following the aforementioned priority experience playback mechanism, the system stores experience samples generated during actual operation in a buffer pool and periodically samples from it for small-batch gradient updates. The update step size for online fine-tuning is typically smaller than that for offline pre-training to avoid drastic changes in network weights. Simultaneously, online fine-tuning also employs the aforementioned temperature compensation trigger mechanism, forcibly triggering updates when the power amplifier temperature drifts rapidly, enabling the network to quickly adapt to the new thermal environment. Through the combination of offline pre-training and online fine-tuning, the system possesses basic compensation capabilities from the initial deployment stage and can continuously optimize its performance during long-term operation.

[0029] In one embodiment, during the actual operation of a terahertz power amplifier, especially in the context of sudden high-power continuous transmission in an industrial internet environment, the junction temperature of the power amplifier chip may rise rapidly from room temperature to over 85 degrees Celsius within tens of milliseconds. This rapid temperature drift directly alters the gain compression point position and phase rotation characteristics of the power amplifier. Although the feature extraction method described above can capture instantaneous distortion features with a sliding window on the order of tens of microseconds, the extracted state vector reflects the distortion that has already occurred, while the nonlinear characteristics of the power amplifier drift with temperature changes at millisecond levels or even faster. The gain compression point shift and phase rotation acceleration caused by temperature changes far exceed the range that the micro-update rate of the sliding window can track. As a result, the compensation parameters output by the reinforcement learning network based on the current state vector always lag behind the actual nonlinear state. During the continuous rise in temperature, the error vector amplitude gradually deteriorates, making it difficult to converge to a acceptable level, and even leading to divergent oscillations. To address this problem, this embodiment adds a temperature compensation triggering mechanism to the step of adjusting the subsequent decisions of the reinforcement learning network based on feedback information. This mechanism does not rely on conventional timed trigger cycles, but directly monitors the temperature changes of the power amplifier chip itself. When the temperature drifts rapidly, it actively intervenes and forces the system to update network parameters immediately, thereby making up for the lag caused by simply relying on the characteristics of the output signal for judgment.

[0030] Specifically, the step of adjusting the subsequent decisions of the reinforcement learning network based on the feedback information further includes temperature compensation triggering, the steps of which include: The temperature value of the power amplifier chip is acquired in real time, and the temperature value at the previous moment is recorded. Calculate the temperature change between the current temperature value and the previous temperature value; The temperature change is compared with a preset temperature change threshold. If the temperature change exceeds the temperature change threshold, the parameters of the reinforcement learning network are updated once without waiting for a preset timed trigger period. When performing the parameter update, the feedback information corresponding to the power amplifier output signal collected at the current moment is used as a high-priority sample and directly used for this parameter update.

[0031] As described above, the first step involves acquiring the chip's current operating temperature in real time at preset sampling intervals using a temperature sensor installed on the power amplifier chip. Simultaneously, the system also needs to record the temperature value from the previous moment to calculate subsequent temperature trends. Temperature is a key physical quantity affecting the nonlinear characteristics of terahertz power amplifiers. When the chip temperature rises, the carrier mobility of the semiconductor device decreases, and the threshold voltage drifts, leading to significant changes in the power amplifier's gain compression point and phase characteristics. These changes are often on the order of milliseconds, much faster than conventional timed trigger cycles. Therefore, relying solely on fixed time intervals to update network parameters may not be sufficient to keep up with the rate of temperature change. The second step involves subtracting the temperature value acquired at the current moment from the temperature value from the previous moment to obtain the temperature change. This change can be positive (representing a temperature increase) or negative (representing a temperature decrease). Whether it's a temperature increase or decrease, as long as the change is large enough, it will cause a significant change in the power amplifier's nonlinear characteristics. For example, in an industrial internet environment, terahertz power amplifiers may experience prolonged high-power transmission, causing the chip junction temperature to rise from room temperature to over 85 degrees Celsius within tens of milliseconds. This drastic temperature rise can rapidly invalidate the previously trained pre-distortion coefficients. By calculating the temperature change, the system obtains an intuitive and sensitive triggering criterion. The third step compares the absolute value of the calculated temperature change with a preset temperature change threshold. This threshold can be calibrated based on the power amplifier's thermal characteristics, for example, set to ±5 degrees Celsius. If the absolute value of the temperature change does not exceed the threshold, it indicates that the power amplifier's thermal state is relatively stable, and the network parameters can continue to be updated according to the normal timed trigger cycle without additional intervention. However, if the temperature change exceeds the threshold, it means that the power amplifier's operating temperature has drifted significantly in a very short time, and its nonlinear characteristics have deviated from their previous state. In this case, the system no longer waits for the preset timed trigger cycle but immediately forces an update of the reinforcement learning network's parameters. This proactive, temperature-triggered update method fundamentally solves the lag problem of timed update mechanisms in scenarios with rapid temperature changes. Fourthly, when the temperature-triggered update is executed, the system marks the feedback information corresponding to the power amplifier output signal acquired at the current moment—that is, the error vector amplitude value—as a high-priority sample. Unlike the aforementioned priority experience playback mechanism, which sets priorities based on the output differences of the dual-critically-based networks, the priority here comes from the triggering of external physical events. The system directly uses this high-priority sample for this parameter update, rather than waiting for it to be randomly selected. The reason for this is that the moment the temperature change exceeds the threshold is precisely the moment when the power amplifier characteristics undergo a substantial change; the feedback information acquired at this time best reflects the distortion state under the new operating conditions. Prioritizing the use of this sample for the update allows the network to adapt to the new thermal environment as quickly as possible, avoiding the use of outdated compensation parameters during the transition period.

[0032] The temperature compensation triggering mechanism in this embodiment ensures that when rapid temperature drift occurs, temperature changes can be detected within tens of microseconds and parameter updates can be forcibly executed, while the most representative samples at present are directly used for learning. This mechanism complements the conventional timed triggering update: timed triggering is responsible for continuous optimization under stable conditions, while temperature triggering is responsible for dealing with sudden thermal shocks. The two work together to enable the entire adaptive compensation system to maintain sub-millisecond response speed and stable compensation accuracy even when facing extreme conditions such as long pulse sequences in the industrial Internet environment, effectively solving the problem of feature parameters lagging behind the actual nonlinear state, as well as the divergent oscillation of the error vector amplitude.

[0033] Reference Figure 3 This application also provides a terahertz band nonlinear distortion compensation system, comprising: Acquisition module 1 is used to acquire the radio frequency signal at the output of the power amplifier in real time, extract the feature parameters reflecting the nonlinear distortion state based on the radio frequency signal, and generate the state vector at the current moment based on the feature parameters. The first input module 2 is used to input the current state vector into a preset reinforcement learning network and output the compensation parameter vector at the current time through the reinforcement learning network. The conversion module 3 is used to convert the compensation parameter vector into predistortion coefficients based on the output result through a preset mapping relationship, and to perform digital predistortion processing on the input baseband signal using the predistortion coefficients; The second input module 4 is used to input the pre-distorted signal into the power amplifier according to the processing result, and to collect the corresponding output signal of the power amplifier. Adjustment module 5 is used to generate feedback information based on the collected output signal and adjust the subsequent decisions of the reinforcement learning network according to the feedback information; Complete module 6, which is used to compensate for nonlinear distortion based on the adjustment results.

[0034] As described above, it is understood that each component of the terahertz band nonlinear distortion compensation system proposed in this application can realize the function of any of the terahertz band nonlinear distortion compensation methods described above, and the specific structure will not be described in detail.

[0035] Reference Figure 4 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 4As shown, this computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores monitoring data and other data. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a terahertz band nonlinear distortion compensation method.

[0036] The processor described above executes the terahertz band nonlinear distortion compensation method, including: real-time acquisition of radio frequency signals at the output of a power amplifier; extraction of feature parameters reflecting the nonlinear distortion state based on the radio frequency signals; generation of a current-time state vector based on the feature parameters; inputting the current-time state vector into a preset reinforcement learning network; outputting a compensation parameter vector at the current-time through the reinforcement learning network; converting the compensation parameter vector into predistortion coefficients based on the output result through a preset mapping relationship; performing digital predistortion processing on the input baseband signal using the predistortion coefficients; inputting the predistorted signal into the power amplifier based on the processing result; acquiring the corresponding output signal of the power amplifier; generating feedback information based on the acquired output signal; adjusting the subsequent decisions of the reinforcement learning network based on the feedback information; and completing the nonlinear distortion compensation based on the adjustment result.

[0037] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a terahertz band nonlinear distortion compensation method, including the following steps: real-time acquisition of radio frequency signals at the output of a power amplifier; extraction of feature parameters reflecting the nonlinear distortion state based on the radio frequency signals; generation of a current-time state vector based on the feature parameters; inputting the current-time state vector into a preset reinforcement learning network; outputting a compensation parameter vector at the current-time through the reinforcement learning network; converting the compensation parameter vector into predistortion coefficients based on the output result through a preset mapping relationship; performing digital predistortion processing on the input baseband signal using the predistortion coefficients; inputting the predistorted signal into the power amplifier based on the processing result; acquiring the corresponding output signal of the power amplifier; generating feedback information based on the acquired output signal; adjusting the subsequent decisions of the reinforcement learning network based on the feedback information; and completing the nonlinear distortion compensation based on the adjustment result.

[0038] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0039] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0040] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for compensating nonlinear distortion in the terahertz frequency band, characterized in that, The method includes: The radio frequency signal at the output of the power amplifier is acquired in real time. Based on the radio frequency signal, characteristic parameters reflecting the nonlinear distortion state are extracted, and the state vector at the current moment is generated based on the characteristic parameters. The current state vector is input into a preset reinforcement learning network, and the current compensation parameter vector is output through the reinforcement learning network. Based on the output, the compensation parameter vector is converted into predistortion coefficients through a preset mapping relationship, and the predistortion coefficients are used to perform digital predistortion processing on the input baseband signal; Based on the processing results, the pre-distorted signal is input into the power amplifier, and the corresponding output signal of the power amplifier is acquired. Feedback information is generated based on the collected output signals, and the subsequent decisions of the reinforcement learning network are adjusted according to the feedback information. Based on the adjustment results, compensation for nonlinear distortion is completed.

2. The terahertz band nonlinear distortion compensation method according to claim 1, characterized in that, The step of extracting feature parameters reflecting the nonlinear distortion state includes: The acquired power amplifier output signal is quadratured to obtain in-phase and quadrature signal components; Perform Hilbert transform on the in-phase and quadrature signal components to calculate the instantaneous amplitude and instantaneous phase values ​​of the signal; An amplitude distortion characteristic curve is generated based on the correspondence between the instantaneous amplitude value and the amplitude value of the input baseband signal; A phase distortion characteristic curve is generated based on the deviation between the instantaneous phase value and the phase value of the input baseband signal. The amplitude range of the input signal is divided into multiple continuous amplitude intervals. Within each amplitude interval, the slope change of the amplitude distortion characteristic curve and the phase rotation of the phase distortion characteristic curve are calculated. The slope change and phase rotation calculated in each amplitude interval are combined into a set of feature parameters to complete the feature parameter extraction.

3. The terahertz band nonlinear distortion compensation method according to claim 1, characterized in that, The steps of inputting the current state vector into a preset reinforcement learning network and outputting the compensation parameter vector at the current time through the reinforcement learning network include: The current state vector is input into the Actor network, which contains three fully connected layers. Each fully connected layer uses the LeakyReLU activation function. The original action vector is output through the Actor network. Based on the output, motion noise is added to the original motion vector to obtain a smoothed compensation parameter vector, which is used as the output of the reinforcement learning network. The current state vector and the compensation parameter vector are input together into the first Critic network to obtain the first Q-value estimate; the current state vector and the compensation parameter vector are input together into the second Critic network to obtain the second Q-value estimate. The smaller value between the first Q-value estimate and the second Q-value estimate is selected as the target Q-value, which is used for the delay policy update of the reinforcement learning network.

4. The terahertz band nonlinear distortion compensation method according to claim 3, characterized in that, The number of nodes in each fully connected layer of the Actor network, the first Critic network, and the second Critic network is 128; the negative slope coefficient of the LeakyReLU activation function is 0.

2.

5. The terahertz band nonlinear distortion compensation method according to claim 1, characterized in that, The step of converting the compensation parameter vector into predistortion coefficients through a preset mapping relationship includes: Obtain the compensation parameter vector output by the reinforcement learning network; The compensation parameter vector is input into a differentiable mapping layer, which contains a trainable linear transformation matrix; A linear transformation is performed on the compensation parameter vector using the linear transformation matrix to obtain predistortion coefficients, wherein the number of predistortion coefficients is greater than the number of elements in the compensation parameter vector.

6. The terahertz band nonlinear distortion compensation method according to claim 3, characterized in that, The step of adjusting the subsequent decisions of the reinforcement learning network based on the feedback information includes: The error vector amplitude is calculated based on the acquired power amplifier output signal, and the error vector amplitude is used as feedback information. The current state vector, compensation parameter vector, feedback information, and the state vector of the next moment are combined into a data record and stored in a preset buffer pool; Data records are read from the buffer pool and sampled according to the priority of each data record. The priority is determined based on the absolute value of the difference between the output value of the first Critic network and the output value of the second Critic network corresponding to the compensation parameter vector in the data record. Gradient updates are performed on the Actor and Critic networks of the reinforcement learning network based on the sampled data records.

7. The terahertz band nonlinear distortion compensation method according to claim 1, characterized in that, The step of adjusting subsequent decisions of the reinforcement learning network based on the feedback information further includes temperature compensation triggering, the step of which includes: The temperature value of the power amplifier chip is acquired in real time, and the temperature value at the previous moment is recorded. Calculate the temperature change between the current temperature value and the previous temperature value; The temperature change is compared with a preset temperature change threshold. If the temperature change exceeds the temperature change threshold, the parameters of the reinforcement learning network are updated once without waiting for a preset timed trigger period. When performing the parameter update, the feedback information corresponding to the power amplifier output signal collected at the current moment is used as a high-priority sample and directly used for this parameter update.

8. A terahertz band nonlinear distortion compensation system, characterized in that, include: The acquisition module is used to acquire the radio frequency signal at the output of the power amplifier in real time, extract the feature parameters reflecting the nonlinear distortion state based on the radio frequency signal, and generate the state vector at the current moment based on the feature parameters. The first input module is used to input the current state vector into a preset reinforcement learning network and output the compensation parameter vector at the current time through the reinforcement learning network. The conversion module is used to convert the compensation parameter vector into predistortion coefficients based on the output result through a preset mapping relationship, and to perform digital predistortion processing on the input baseband signal using the predistortion coefficients; The second input module is used to input the pre-distorted signal into the power amplifier according to the processing result, and to acquire the corresponding output signal of the power amplifier. An adjustment module is used to generate feedback information based on the collected output signal and adjust the subsequent decisions of the reinforcement learning network according to the feedback information. The completion module is used to compensate for nonlinear distortion based on the results of the adjustment.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images