Power amplifier dynamic predistortion optimization processing method and system based on reinforcement learning

Abstract

The application discloses a power amplifier dynamic predistortion optimization processing method and system based on fusion reinforcement learning, relates to the technical field of radio frequency power amplifier signal processing, and comprises the following steps: collecting feedback signals and environmental change signs during power amplifier operation, unifying time scales and frequency band scales, generating a comprehensive reference draft of the feedback signals and the environmental change signs, and establishing a common reference for subsequent comparison. The application distinguishes interference behaviors from power amplifier operation states by constructing a comprehensive reference draft with unified scales and extracting interference rhythms and interference fingerprints, and avoids false updates of reinforcement learning during interference. By setting a freezing point, inserting a silent sampling window and a reverse reward mask, the application ensures that predistortion adjustment bypasses interference, so that parameter updating is more reliable and the linearization state of the power amplifier remains stable.

Description

Technical Field

[0001] This invention relates to the field of radio frequency power amplification signal processing technology, specifically to a method and system for optimizing dynamic predistortion processing of power amplifiers by incorporating reinforcement learning. Background Technology

[0002] Dynamic predistortion optimization processing for power amplifiers, incorporating reinforcement learning, is an intelligent signal correction method that introduces artificial intelligence decision-making mechanisms into the linearization control process of RF power amplifiers. Its core idea is to abstract the adjustment process of predistortion parameters into a decision-making problem involving continuous interaction between an intelligent agent and the power amplifier environment, based on the traditional digital predistortion principle and aided by reinforcement learning algorithms. The system collects the power amplifier's operating status and output signal characteristics in real time, calculates signal distortion and spectral purity, and feeds these back to the intelligent agent as reward signals. Through continuous trial and error and learning, the intelligent agent dynamically updates the predistortion parameters, enabling adaptive compensation of the input signal before it enters the power amplifier, thereby offsetting the nonlinear distortion generated by the power amplifier under different operating conditions. This processing can automatically optimize parameters when the power amplifier temperature, power supply voltage, and load conditions change, achieving real-time correction of signal distortion and maintenance of linearity, ultimately improving the output signal quality and energy efficiency stability of the communication system.

[0003] The existing technology has the following shortcomings: In existing technologies, power amplifier predistortion optimization largely relies on reinforcement learning agents to automatically adjust compensation parameters based on changes in the characteristics of the output signal to reduce nonlinear distortion. However, in scenarios where external radio frequency interference suddenly increases, a large number of high-frequency noise components will be mixed into the power amplifier output signal, causing abnormal changes in the signal waveform and spectrum. Reinforcement learning systems are prone to misinterpreting these interference features as nonlinear distortion of the power amplifier itself. At this time, the agent will incorrectly perform parameter update operations, improperly adjusting the amplitude and phase of the input compensation signal, thereby disrupting the original linearization balance. This incorrect adjustment can cause a sharp distortion of the output signal waveform and a sudden spread of the spectrum within a short period of time, allowing adjacent channel signals to intrude. This can easily lead to a decrease in communication quality, signal crosstalk, or even system interference alarms, and in severe cases, it can cause abnormal output of the entire unit or communication interruption.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for optimizing dynamic predistortion of power amplifiers by incorporating reinforcement learning, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a power amplifier dynamic predistortion optimization processing method incorporating reinforcement learning, comprising the following steps: The system collects feedback signals and environmental change indicators during the operation of the power amplifier, standardizes the time scale and frequency band scale, and generates a comprehensive reference draft of feedback signals and environmental change indicators to establish a common reference for subsequent comparisons. Based on the comparison of environmental change signs and feedback signal waveform fluctuations in the comprehensive reference draft, the external interference rhythm index is extracted under the common reference, and the power amplifier stable segment in the comprehensive reference draft is organized into a comparison segment for subsequent anomaly identification. By relying on the external interference rhythm index to trace back the comparison segment, the abnormal segment dominated by external interference is located at the end of the comparison segment, generating an interference fingerprint draft, and recording the frequency band position and time span of the interference in the interference fingerprint draft for subsequent mapping analysis with reinforcement learning records. Based on the comparison of interference fingerprint drafts, the frequency band positions and time spans of interference fingerprint draft records are mapped to the decision history item by item. In the mapping results, reward mutations and decision reversals are identified to form a list of misjudgment triggers. High-risk time slots are marked in the list of misjudgment triggers for reinforcement learning adjustment strategies. The interaction rhythm of reinforcement learning is dynamically adjusted based on the misjudgment trigger list. High-risk time slots in the misjudgment trigger list are set as predistortion parameter update freeze points. A silent sampling window is inserted at the freeze point and a reverse reward mask is injected to guide subsequent predistortion adjustment to avoid external interference-dominated segments, thereby improving the accuracy of predistortion parameter updates and the linearization stability of the system.

[0007] Preferably, the steps for generating a comprehensive reference draft are as follows: The system collects feedback signals during the operation of the power amplifier and simultaneously collects signs of environmental changes. The changes in amplitude, phase, and frequency band of the feedback signals and the changes in signs of environmental changes are recorded and formed into a basic acquisition sequence according to a common acquisition cycle. Based on the basic acquisition sequence, feedback signals and signs of environmental change are uniformly labeled on the time scale and frequency band scale, and the acquisition points of feedback signals and signs of environmental change are arranged according to a common step size to form intermediate material; In the intermediate material, the changes in the waveform of the feedback signal and the changes in the signs of environmental change are arranged together, and the two are combined and presented in a unified time scale and a unified frequency band scale to form a comprehensive reference draft; The time and frequency scales of the comprehensive reference draft are continuously read and extended to the entire operating cycle. In the subsequent comparison stage, the comprehensive reference draft is used as a common reference to carry out correlation analysis between feedback signals and signs of environmental changes.

[0008] The preferred method for extracting the external interference rhythm index is as follows: Under the condition that the time scale and frequency band scale of the comprehensive reference draft are unified, the fluctuations of the feedback signal waveform and the signs of environmental changes are analyzed segment by segment to form potential interference-related candidate segments. Based on the candidate fragments associated with potential interference, the preceding and subsequent change trajectories are read in the comprehensive reference manuscript. The changes in the prelude of interference, the derivative changes of interference, and the decay changes of interference are extracted and organized into an external interference rhythm index. After the external interference rhythm index is formed, extract the segments with continuous fluctuations in the feedback signal waveform and smooth signs of environmental change from the comprehensive reference manuscript and organize them into control segments so that the control segments maintain the scale of the comprehensive reference manuscript. The external interference rhythm index and the control segment are aligned and expanded in the comprehensive reference manuscript to enable anomaly identification of interference behavior and natural running behavior under common reference conditions.

[0009] Preferably, during the alignment and expansion process of the external interference rhythm index and the control segment, the interference prelude changes and interference derivative changes in the external interference rhythm index are arranged continuously, and the control segments are kept in order in the same time scale and frequency band scale, so as to enhance the ability to distinguish between interference behavior and natural operation behavior under common reference conditions.

[0010] Preferably, the steps for generating the interference fingerprint draft are as follows: By relying on the external interference rhythm index to backtrack the control segment, the fluctuations of the feedback signal waveform of the control segment and the signs of environmental changes are read according to the time scale to form a continuous expression; Based on the peak value of the change in the external interference rhythm index, the interference change trajectory is projected to the end of the control segment, and the abnormal segment dominated by external interference is located in the time scale and frequency band scale. Record the start and end positions of the abnormal segments on the time axis as the interference time span, and record the frequency band positions covered by the abnormal segments as the interference frequency band positions, and organize them into an interference fingerprint draft. The interference fingerprint draft is aligned with the time and frequency scale of the comprehensive reference draft, so that the interference fingerprint draft can be continuously matched with the time span and frequency position of the feedback signal waveform fluctuations for subsequent mapping analysis.

[0011] Preferably, during the generation of the interference fingerprint draft, the time span and frequency band position of the abnormal segment are compared with the continuous change trajectory of the external interference rhythm index. Under the condition of consistent comparison, the boundary of the interference time span is defined, and the coverage range of the interference frequency band position is defined, so as to enhance the accuracy of the interference fingerprint draft in expressing the abnormal segment dominated by external interference.

[0012] Preferably, the steps for forming the false alarm trigger list are as follows: After the interference fingerprint draft is formed, the history of reinforcement learning decisions is read along the interference time span as the main line, and the records of reinforcement learning in the time scale are matched with the interference time span segment by segment. Based on the interference time span and interference frequency band location, the decision records in the reinforcement learning decision history are mapped item by item on the time scale and frequency band scale to form a continuous behavioral trajectory of interference entry, interference duration and interference decay. Identify reward mutations and decision reversals that occur during interference in reinforcement learning within the mapped trajectory, and record and compile these changes into a list of misjudgment triggers. High-risk time slots are marked on the time scale based on the misjudgment trigger list. The reward mutation position and the decision reversal position are mapped to the interference time span, so that the high-risk time slots are marked in the reinforcement learning adjustment strategy for subsequent adjustment.

[0013] Preferably, when marking high-risk time slots, the reward mutation position and the decision reversal position are simultaneously marked on the frequency band scale, with the interference frequency band position in the interference fingerprint draft as a reference. The simultaneously marked position is then associated with the corresponding time period in the interference time span, so that the high-risk time slots form a joint location on the time scale and the frequency band scale for reinforcement learning adjustment strategies.

[0014] The preferred steps for dynamically adjusting the interaction rhythm of reinforcement learning are as follows: Arrange high-risk time slots according to the list of misjudgments, and map the high-risk time slots to the parameter update nodes of the reinforcement learning interaction rhythm on the time scale to form continuous markings; Set a pre-distortion parameter update freeze point at the node corresponding to the high-risk time slot, and extend the coverage of the freeze point along the time scale to the end of the high-risk time slot to pause parameter updates. A silent sampling window is inserted inside the freezing point. The silent sampling window is expanded according to the starting position of the freezing point and the change trajectory of the feedback signal during the freezing point is recorded. Based on the feedback signal change trajectory formed by the silent sampling window, an inverse reward mask is injected inside the freeze point to mark the record inside the freeze point as a period that cannot be used for predistortion parameter update, so as to guide subsequent predistortion adjustment to avoid external interference dominating the segment.

[0015] The power amplifier dynamic predistortion optimization processing system integrating reinforcement learning includes a reference construction module, an interference rhythm extraction module, an interference fingerprint generation module, a misjudgment trigger recognition module, and a dynamic adjustment execution module. By referring to the construction module, feedback signals and environmental change signs during the operation of the power amplifier are collected, time scale and frequency band scale are unified, and a comprehensive reference draft of feedback signals and environmental change signs is generated to establish a common reference for subsequent comparisons. The interference rhythm extraction module compares environmental change signs and feedback signal waveform fluctuations with the comprehensive reference draft, extracts the external interference rhythm index under the common reference, and organizes the power amplifier stable segment in the comprehensive reference draft into a comparison segment for subsequent anomaly identification. The interference fingerprint generation module relies on the external interference rhythm index to trace back the comparison segment, locates the abnormal segment dominated by external interference at the end of the comparison segment, generates an interference fingerprint draft, and records the frequency band position and time span of the interference in the interference fingerprint draft for subsequent mapping analysis with reinforcement learning records. The misjudgment trigger identification module, based on the comparison of interference fingerprint drafts with the reinforcement learning decision history, maps the frequency band position and time span of the interference fingerprint draft records to the decision history item by item. In the mapping results, it identifies reward mutations and decision reversals, forms a misjudgment trigger list, and marks high-risk time slots in the misjudgment trigger list for reinforcement learning adjustment strategies. The dynamic adjustment execution module dynamically adjusts the reinforcement learning interaction rhythm based on the misjudgment trigger list. High-risk time slots in the misjudgment trigger list are set as predistortion parameter update freeze points. A silent sampling window is inserted at the freeze point and a reverse reward mask is injected to guide subsequent predistortion adjustment to avoid external interference-dominated segments, thereby improving the accuracy of predistortion parameter updates and the linearization stability of the system.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a comprehensive reference model with unified time and frequency scales before predistortion adjustment. Based on this model, it extracts an external interference rhythm index and constructs an interference fingerprint draft. Then, it combines this with reinforcement learning decision history to identify misjudged trigger behaviors, enabling the interference factors and the actual operating state of the power amplifier to be distinguished temporally. This allows reinforcement learning to identify the source of external disturbances when processing feedback information. Therefore, reinforcement learning will not enter an erroneous update process during periods of dominant interference, and the predistortion parameters can remain continuous, ensuring the stability of predistortion adjustment even in dynamic environments.

[0017] This invention sets a predistortion parameter update freeze point based on the misjudgment trigger list, and inserts a silent sampling window and a reverse reward mask at the freeze point. This allows reinforcement learning to pause policy adjustments and independently record the interference trajectory during the interference period, enabling subsequent parameter updates to avoid the interference-affected time period. Through this method, the predistortion parameter update path is more reliable, and the power amplifier's linearization state remains balanced under interference conditions, thus improving output signal quality and linearization stability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the power amplifier dynamic predistortion optimization processing method that integrates reinforcement learning according to the present invention.

[0020] Figure 2 This is a schematic diagram of the power amplifier dynamic predistortion optimization processing system that integrates reinforcement learning according to the present invention. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0022] This invention provides, for example Figure 1 The power amplifier dynamic predistortion optimization method fused with reinforcement learning, as shown, includes the following steps: The system collects feedback signals and environmental change indicators during the operation of the power amplifier, standardizes the time scale and frequency band scale, and generates a comprehensive reference draft of feedback signals and environmental change indicators to establish a common reference for subsequent comparisons. To achieve coordinated analysis of feedback signals and environmental change indicators during pre-distortion adjustment, relevant information can be organized, correlated, and uniformly processed according to the following procedure to form a comprehensive reference draft that can support subsequent comparison processes. Throughout the process, the expression of feedback signals and environmental change indicators remains consistent so that the same semantic system can be continuously used for comparison, judgment, and localization in subsequent steps. The specific implementation steps are as follows: Feedback signals from the power amplifier during operation are continuously acquired, and environmental change indicators within the same time period are simultaneously obtained. The feedback signal encompasses amplitude and phase changes over time in the power amplifier output, as well as spectral variations within the frequency band. Environmental change indicators include external factors affecting the waveform changes of the feedback signal, such as external interference fluctuations, changes in external radiation sources, temperature field variations, and load condition changes within the frequency band. To ensure a correlation between the feedback signal and environmental change indicators in subsequent processes, each acquisition segment is acquired according to a uniform acquisition cycle. After acquisition, the time and frequency band records of each segment are compiled, forming two basic acquisition sequences with different sources and content but sharing a common time and frequency band scale. During this process, each change point in the feedback signal and each change record in the environmental change indicators are arranged side-by-side according to the acquisition sequence, allowing the change trajectories of the feedback signal and environmental change indicators to be accurately located on the same timeline and within the same frequency band scale, thus forming preliminary synchronization material suitable for further processing.

[0023] After obtaining the initial synchronized data, the feedback signal and environmental change indicators undergo time and frequency scale unification. To ensure that the feedback signal and environmental change indicators can be overlaid and analyzed within the same reference system during subsequent comparisons, the data from both sources needs to be re-labeled according to a unified time scale. During this process, the feedback signal acquisition points and environmental change indicator acquisition points are arranged equidistantly with the same time step, ensuring a one-to-one correspondence between all acquisition points on the time axis. Simultaneously, to ensure that the frequency band information of the feedback signal can be linked with that of the environmental change indicators, the frequency band range in the feedback signal is expanded, allowing the frequency band positions in the external interference information to map to the positions of spectral changes in the feedback signal within the same frequency band range. After these processes, the feedback signal and environmental change indicators are unified in both time and frequency scale dimensions, enabling continuous comparison of the two types of data within a common scale system in subsequent processing stages. After completing the calibration unification, the waveform state and spectrum state of the feedback signal at each acquisition time point are combined with the change state of environmental change signs at the corresponding time point to form intermediate material containing time scale information, frequency band scale information and two types of signal content.

[0024] After completing the intermediate data, the feedback signal and environmental change indicators are fused to generate a complete comprehensive reference draft. In this process, the waveform and spectral changes of the feedback signal at different time periods are jointly arranged with the amplitude and trajectory of the environmental change indicators, and presented side-by-side in the same descriptive unit according to a unified time and frequency scale. By synchronously describing the positions of feedback signal changes and environmental change indicator changes, the comprehensive reference draft can fully express the correspondence between the feedback signal and environmental change indicators during amplifier operation. Furthermore, the waveform change characteristics of the feedback signal in each time period are correlated with the jump characteristics of the environmental change indicators in the same time period, ensuring that the comprehensive reference draft not only includes the original records of the feedback signal and environmental change indicators but also information on whether there are synchronous, adjacent, or sequential changes between them. The comprehensive reference draft generated in this way presents the overall correspondence between the feedback signal and environmental change indicators during amplifier operation, making it a common reference material that can support subsequent comparison work. When creating the comprehensive reference draft, the intermediate materials are sorted according to the continuity of time so that the entire reference draft can be continuously read in chronological order, while maintaining the consistency of the frequency band scale. This allows for the direct location of the corresponding and different points between changes in feedback signals and changes in environmental change signs in subsequent steps based on the comprehensive reference draft.

[0025] After the comprehensive reference draft is formed, it serves as a common reference for subsequent comparison processes. By continuously reading the time and frequency band scales in the comprehensive reference draft, the correlation trajectory between the feedback signal and environmental change indicators is extended, forming a foundational data that can be directly used in subsequent comparison steps. This processing allows subsequent steps to directly utilize the unified scale system in the comprehensive reference draft when comparing the fluctuations in the feedback signal waveform with the trajectory of environmental change indicators. This avoids repeatedly performing scale adjustments at different time periods or frequency bands and ensures a consistent reference system for continuous analysis of interference change characteristics, feedback signal waveform change characteristics, and the correspondence between the two in subsequent steps. Therefore, the comprehensive reference draft is not merely a collection and recording of results, but provides coherent, unified, and directly applicable foundational material for subsequent steps such as interference rhythm extraction, control segment generation, and interference segment localization. This ensures that the entire pre-distortion adjustment process obtains a complete and consistent reference system from the initial stage, guaranteeing that each subsequent step can be carried out based on the same framework.

[0026] Based on the comparison of environmental change signs and feedback signal waveform fluctuations in the comprehensive reference draft, the external interference rhythm index is extracted under the common reference, and the power amplifier stable segment in the comprehensive reference draft is organized into a comparison segment for subsequent anomaly identification. Based on the comprehensive reference draft, environmental change indicators and feedback signal waveform fluctuations are compared, and an external interference rhythm index is extracted under the common reference. Simultaneously, stable sections of the power amplifier are identified from the comprehensive reference draft and compiled into comparison segments to lay the foundation for subsequent anomaly identification. To ensure a consistent expression method, a unified reference system, and continuity between each stage of the above process, the specific implementation steps are as follows: Based on the established and unified time and frequency scales of the comprehensive reference draft, a segment-by-segment analysis of the feedback signal waveform fluctuations and corresponding environmental change indicators in the comprehensive reference draft is performed to extract the synchronous change trajectory between the two. To this end, during the reading of the comprehensive reference draft, each record is treated as a composite expression unit composed of feedback signal waveform fluctuation information and environmental change indicator information, and unfolded sequentially in chronological order. During the unfolding process, for each composite expression unit, the continuity characteristics of amplitude, phase, and frequency band changes in the feedback signal waveform fluctuations within that time period are first identified, followed by the combination of the rhythm, direction, and intensity of changes in the environmental change indicators within that time period. When two change sequences exhibit correlated trends within the same time period—for example, the feedback signal waveform fluctuations show continuous fluctuations while the environmental change indicators exhibit jumps, disturbances, or other forms of disturbance within the same time period—this synchronous change phenomenon is marked as a potential interference-related candidate segment. After completing the segment-by-segment analysis of all composite expression units, these potential interference-related candidate segments are recorded collectively and rearranged in chronological order to provide continuous basic material for further extraction of the synchronous trajectory. In this way, the potential interference correlation candidate fragments derived from the comprehensive reference draft can accurately present the correlation between environmental change signs and feedback signal waveform fluctuations, laying the foundation for the next stage of refining the external interference rhythm index.

[0027] After compiling the candidate segments for potential interference, these segments are further analyzed using the common reference system in the comprehensive reference draft to extract an index of external interference rhythm. This process involves reading environmental change signs and feedback signal waveform fluctuations before and after the candidate segments in the comprehensive reference draft to further identify the continuous change patterns of interference in the time and frequency band dimensions. During the reading process, each candidate segment is considered as the starting point where interference may enter the feedback signal. The trajectory of environmental change signs is traced backward from the candidate segment to identify the prelude changes in the interference construction process, such as the gradual accumulation of interference or the migration trend of interference within the frequency band. At the same time, the trajectory of feedback signal waveform fluctuations is extended backward from the candidate segment to identify the derivative changes after the interference is triggered, such as the continuous extension of the feedback signal waveform phase change or the continuous disturbances in the frequency band of the feedback signal. Through the joint analysis of the prelude changes and derivative changes of interference, the changes in the three stages of interference appearance, development, and decay can be continuously expressed, so that the behavioral logic of interference in the entire comprehensive reference draft can be fully presented. Based on this logic, the rhythm of interference changes is refined into a unified expression with time span, frequency band span, and change sequence characteristics. This expression is then organized into an external interference rhythm index, which is used to continue participating in subsequent stages of interference localization, anomaly identification, and misjudgment avoidance.

[0028] Building upon the existing external interference rhythm index, and relying on the common reference system in the comprehensive reference draft, stable operating segments of the power amplifier during periods unaffected by interference are identified and compiled into reference segments for subsequent anomaly identification. In this process, segments in the comprehensive reference draft where feedback signal waveform fluctuations are regular and continuous, and environmental changes remain smooth and stable, are first read. These segments are typically unaffected by the interference change trajectories identified in the interference rhythm index and can therefore be considered stable signal segments exhibited by the power amplifier during natural operation. When identifying these segments, the time and frequency scales in the comprehensive reference draft are retained in their original state, ensuring that the reference segments maintain a consistent time and frequency labeling system with the interference rhythm index. Subsequently, these segments are organized chronologically to fully represent the power amplifier's operating characteristics in an interference-free environment. The compilation of reference segments includes not only recording the regularity of signal waveform fluctuations but also recording the smoothness of environmental changes, enabling the reference segments to serve as a benchmark for interference judgment in the subsequent anomaly identification stage. In this way, the resulting control fragments and interference rhythm indexes are both derived from the comprehensive reference manuscript and maintain a common reference system, enabling subsequent comparisons to be conducted directly within the same expressive framework.

[0029] With both the external interference rhythm index and the control segment already established, they are jointly incorporated into the extended expression of the comprehensive reference draft for coherent use in subsequent anomaly identification. By establishing correspondences between the interference rhythm index and the control segment and the various parts of the comprehensive reference draft, the subsequent anomaly identification stage can directly perform a three-way comparison based on the original content of the comprehensive reference draft, the change trajectory of the interference rhythm index, and the stable trajectory of the control segment. This allows for the differentiation between interference behavior and natural operating behavior under common reference conditions. In this way, the entire pre-distortion adjustment process obtains a continuous, uniformly scaled, and logically coherent material foundation at the interference identification stage. This enables subsequent interference localization, interference fingerprint construction, and misjudgment risk identification to be carried out on a stable, continuous, and traceable material system, providing a stable interference identification foundation for the entire pre-distortion adjustment process and ensuring that subsequent steps can directly inherit the interference rhythm index and control segment formed in this step.

[0030] By relying on the external interference rhythm index to trace back the comparison segment, the abnormal segment dominated by external interference is located at the end of the comparison segment, generating an interference fingerprint draft, and recording the frequency band position and time span of the interference in the interference fingerprint draft for subsequent mapping analysis with reinforcement learning records. After the external interference rhythm index and control segments are established through the preceding steps, in order to accurately locate the interference changes in both time and frequency spans, and to provide sufficient evidence for subsequent mapping analysis based on reinforcement learning records, the change trajectory at the end of the control segments can be identified by retrospectively correlating the external interference rhythm index and control segments. This allows for the extraction of anomalous segments dominated by external interference, which are then compiled into a draft interference fingerprint. The draft interference fingerprint records the temporal coverage and frequency distribution of the interference, providing a directly usable representation of the interference trajectory for subsequent mapping of reinforcement learning behavior records. To ensure the consistency of the representation system, the entire processing flow can be carried out in the following stages to maintain the continuity, traceability, and consistency of all representations. The specific implementation steps are as follows: After the external interference rhythm index has established a continuous interference change trajectory, the control segments are backtracked and unfolded chronologically, using this trajectory as the main line. During the unfolding process, starting from the beginning of the control segment, a sequential scan is performed according to the time scale, and the continuity characteristics of the feedback signal waveform fluctuations in each control segment are read. Simultaneously, the stable state characteristics of environmental change indicators at each corresponding position are also read. By comparing the stable expression of feedback signal waveform fluctuations in the control segment with the interference change trajectory reflected in the external interference rhythm index, the signal regularity exhibited by the control segment during the period when interference has no effect can be clearly identified. During the backtracking process, the initial change position of the external interference rhythm index is used as a reference point. When the control segment enters a time range close to the change starting point of the interference rhythm index, this time range is designated as the key analysis interval, and the feedback signal waveform fluctuations and environmental change indicators within this interval are read segment by segment. This allows the backtracking process to capture the change trend of the control segment as it is about to leave the stable state under common reference. In this way, a continuous backtracking correlation is established between the external interference rhythm index and the control segment, providing sufficient preliminary information for subsequent location of abnormal segments.

[0031] After completing the retrospective analysis of the control segment and obtaining the continuous expression of the control segment near the interference range, the peak value of the external interference rhythm index was used as the localization basis to gradually identify the abnormal segment dominated by external interference at the end of the control segment. During the localization process, the expressions in the external interference rhythm index reflecting the changes in interference intensity, interference duration, and interference coverage frequency band were projected onto the continuous expression at the end of the control segment in chronological order, so that the waveform change trajectory of the feedback signal at the end of the control segment could be gradually aligned with the interference rhythm trajectory. When the two overlapped in both the time scale and frequency band scale dimensions, the overlapping position was taken as a candidate abnormal segment, and further extended to adjacent time periods to find the moment when the interference completely covered the stable structure of the control segment. When the interference began to change the originally stable waveform of the feedback signal in the control segment, this position could be regarded as the starting position of the abnormal segment dominated by external interference, and the change pattern of the feedback signal waveform during the duration of the interference could be read again, so that the entire abnormal segment had continuity and traceability. In this process, the abnormal segment not only included the first change position after the interference entered, but also the continuous fluctuations caused by the fluctuations of the feedback signal waveform during the duration of the interference, thus forming a complete abnormal trajectory that could express the interference behavior. By fully reading the anomalous segments, the interfering behaviors can be independently extracted from the control segments and used as input material for subsequent steps.

[0032] After anomalous segments dominated by external interference have been continuously identified, they are compiled into an interference fingerprint draft to serve as a direct representation for subsequent mapping analysis. During the drafting process, the start and end positions of the anomalous segments on the time axis are recorded according to the original scale, providing a temporal description of the interference occurrence. Simultaneously, the frequency bands covered by the anomalous segments are marked according to the frequency band scale, ensuring the draft accurately represents the interference's coverage area within the frequency band. To maintain consistency between the interference fingerprint draft and the comprehensive reference draft, the draft maintains the same time and frequency band scales as the comprehensive reference draft, allowing direct matching between the draft and any part of the comprehensive reference draft. Furthermore, the interference fingerprint draft not only records the distribution of interference across time and frequency bands but also describes the continuous trajectory of the feedback signal waveform fluctuations caused by the interference within the anomalous segments, fully presenting the changing patterns of the interference behavior. After processing, the interference fingerprint draft is transformed into a continuous, complete expression of interference variations that can be directly matched on the time and frequency scales.

[0033] After the interference fingerprint draft was created, to ensure its effectiveness in subsequent mapping analysis with reinforcement learning records, the draft was realigned with the comprehensive reference draft. This allowed the interference fingerprint draft to be directly located along the timeline of the comprehensive reference draft in later stages. During alignment, the time span recorded in the interference fingerprint draft was expanded according to the time scale of the comprehensive reference draft, ensuring the accurate identification of the corresponding time period of the interference in the comprehensive reference draft. Simultaneously, the frequency band positions recorded in the interference fingerprint draft were mapped according to the frequency band scale of the comprehensive reference draft, establishing a direct link between the interference fingerprint draft and the frequency range of the feedback signal waveform fluctuations in the comprehensive reference draft. In this way, the interference fingerprint draft became a representation of interference behavior that could simultaneously establish a correspondence with both the comprehensive reference draft and the reinforcement learning records. This allowed for direct comparison with reinforcement learning behavior trajectories during subsequent mapping analysis, enabling rapid resolution of the interference behavior's location, duration, and covered frequency bands during the mapping process. In this way, the interference fingerprint draft becomes the basic expression for subsequent interference identification, misjudgment trigger identification, and risk time slot marking, enabling the entire pre-distortion adjustment process to form sufficient, continuous, and consistent interference description material during the interference localization stage, providing stable support for subsequent steps.

[0034] Based on the comparison of interference fingerprint drafts, the frequency band positions and time spans of interference fingerprint draft records are mapped to the decision history item by item. In the mapping results, reward mutations and decision reversals are identified to form a list of misjudgment triggers. High-risk time slots are marked in the list of misjudgment triggers for reinforcement learning adjustment strategies. To ensure a continuous correspondence between the draft interference fingerprint and the reinforcement learning decision-making history, and to establish a comparable relationship between the trajectory of interference behavior and the behavioral records at different stages of reinforcement learning, thereby identifying reward mutations and decision reversals during the comparison process, and forming a list of misjudgment triggers and marking high-risk time slots, the entire process of comparison, mapping, and marking can be carried out around the draft interference fingerprint in the following stages. This approach allows the impact of interference behavior on the feedback signal to be located within the reinforcement learning decision-making history, providing a complete picture of the entire process of interference behavior triggering error updates, and offering a clear basis for the execution of subsequent adjustment strategies. The specific implementation steps are as follows: After the interference fingerprint draft has been formed and includes information on the interference frequency band location and interference time span, it is used as a reference to systematically read the reinforcement learning decision history. This establishes a continuous reference relationship that maps interference behavior to reinforcement learning behavior. During the reading process, the time span recorded in the interference fingerprint draft is used as the main thread. All decision records generated within that time span in the reinforcement learning decision history are organized in chronological order, ensuring that each behavior in the reinforcement learning decision history corresponds to a specific time point in the interference fingerprint draft. To maintain consistency between the expression system of the interference fingerprint draft and the expression system of the reinforcement learning decision history, the feedback signal quality index, policy adjustment direction, and reward change trajectory at each decision moment are simultaneously read during the reinforcement learning decision history reading. This ensures that the impact of interference behavior in the feedback signal can correspond parallel to the behavior expression in the reinforcement learning decision history. In this process, the records of the interference fingerprint draft in the frequency band scale also serve as a reference, providing additional frequency band location information for reading the reinforcement learning decision history. This allows the frequency band activity locations that may be affected by interference in the reinforcement learning behavior history to be directly linked to the frequency band representation of the interference fingerprint draft. In this way, a preliminary alignment between the interference fingerprint draft and the reinforcement learning decision history is achieved in both the time and frequency band dimensions, providing a continuous correlation basis for subsequent mapping processes.

[0035] After the initial alignment of the interference fingerprint draft and the reinforcement learning decision history, the interference frequency band positions and interference time spans recorded in the interference fingerprint draft are mapped item by item to the reinforcement learning decision history. This ensures that every position in the feedback signal where the interference behavior appears can be found in the reinforcement learning decision history. During this item-by-item mapping, the interference time span in the interference fingerprint draft is used as the core. Each decision record in the reinforcement learning decision history is overlapped with the interference time span according to the time scale, and all decision records in the reinforcement learning decision history that overlap with the interference time span are included in the mapping sequence. During the mapping process, starting from the beginning time point of the interference fingerprint draft, the interference frequency band positions and feedback signal waveform fluctuations are mapped to each decision node in the reinforcement learning decision history, ensuring that the policy update trajectory occurring simultaneously with the interference behavior is marked point by point. Then, using the end time point of the interference fingerprint draft as the boundary, the portion of the reinforcement learning decision history after the interference ends is distinguished from the portion during the interference period, allowing the scope of the interference behavior's influence to be fully outlined in the reinforcement learning behavior chain. Thus, the reinforcement learning decision history trajectory that can express the three stages of interference entry, interference persistence and interference decay can be obtained in the mapping sequence, so that the expression of the interference fingerprint draft can form a continuous mapping curve in the reinforcement learning decision history.

[0036] After a point-by-point mapping between the interference fingerprint draft and the reinforcement learning decision history has been established, the mapping results are continuously read to identify reward mutations and decision reversals that occur during interference. Reward mutations typically manifest as a sudden decrease or increase in reward during a decision node. These changes often do not originate from the amplifier's operating state itself, but rather from abnormal changes in the feedback signal caused by the interference time span and frequency band location recorded in the interference fingerprint draft. When reading the mapping results, the interference time span is used as the main thread to analyze reward changes point-by-point in the reinforcement learning decision history. This identifies whether reward changes jump upon the entry of interference, whether discontinuous changes occur during the duration of interference, and whether the change pattern reverts to the previous pattern after the interference subsides. This allows for accurate extraction of reward mutations from the mapping results. Simultaneously, decision reversals manifest as a reversal of direction, a shift in intensity, or a reversal of behavioral trends within the interference time span. When the policy direction in the reinforcement learning decision history changes during the interference period, and this change highly corresponds to the location of the interference frequency band and the duration of the interference, the change can be marked as a decision reversal phenomenon. After identifying reward mutation phenomena and decision reversal phenomena, these marked contents are integrated into a misjudgment trigger list, so that the trigger points of misjudgment behaviors can be continuously recorded in the list.

[0037] After the misjudgment trigger list is formed, the records in the list are reordered according to the time scale to mark high-risk time slots. To ensure that high-risk time slots accurately represent the key locations where reinforcement learning misjudged behavior occurred during the interference period, the reward mutation positions and decision reversal positions recorded in the misjudgment trigger list are mapped to specific time periods within the interference time span, using the time span of the interference fingerprint draft as a reference. This allows reinforcement learning misjudgments at the time of interference entry, interference duration, and interference decay to be marked with different categories in the list. During the marking of high-risk time slots, each time position in the misjudgment trigger list is expanded according to the change trajectory recorded in the interference fingerprint draft. This ensures that each high-risk time slot not only includes the instantaneous location of the misjudgment behavior but also the time range within which the misjudgment behavior may affect subsequent decisions, thus providing a continuous representation of high-risk time slots. In this way, high-risk time slots become direct input materials for reinforcement learning regulation strategies, enabling subsequent steps to perform operations such as freezing strategies, silent sampling window injection, and reverse reward masking based on high-risk time slots. This allows the predistortion regulation process to avoid generating erroneous policy updates in the interference-dominant segment, thereby ensuring the stability and consistency of reinforcement learning behavior and predistortion regulation behavior.

[0038] The interaction rhythm of reinforcement learning is dynamically adjusted based on the misjudgment trigger list. High-risk time slots in the misjudgment trigger list are set as predistortion parameter update freeze points. A silent sampling window is inserted at the freeze point and a reverse reward mask is injected to guide subsequent predistortion adjustment to avoid external interference-dominated segments, thereby improving the accuracy of predistortion parameter updates and the linearization stability of the system. After the misjudgment trigger list has been formed and includes the corresponding reward mutation positions and decision reversal positions, in order to enable the reinforcement learning interaction rhythm to be adjusted in a targeted manner according to the misjudgment trigger list, and to prevent high-risk time slots from being misjudged by the reinforcement learning strategy as time points that can participate in parameter updates during subsequent predistortion adjustment, the reinforcement learning interaction rhythm can be continuously intervened in. This allows the predistortion parameter update process to actively avoid external interference-dominated segments, and further improves the linear stability of the entire predistortion adjustment chain. The specific implementation steps are as follows: Based on the misjudgment trigger list, which already includes the reward mutation locations and decision reversal locations caused by interference, a preliminary analysis of the reinforcement learning interaction rhythm is conducted using the misjudgment trigger list as the primary information. This ensures that the timeline of reinforcement learning's participation in predistortion parameter updates aligns with the misjudgment trigger list. During the analysis, each high-risk time slot in the misjudgment trigger list is arranged sequentially according to the time scale, and these high-risk time slots are identified as potentially risky periods within the reinforcement learning interaction rhythm. Subsequently, using the position of the high-risk time slot in the misjudgment trigger list as a benchmark, the corresponding parameter update nodes are found within the reinforcement learning interaction rhythm, allowing the actions performed by reinforcement learning at these nodes to be marked as time points unsuitable for predistortion parameter updates. To maintain a consistent representation, the start and end positions of the high-risk time slots in the time span dimension are expressed according to the original time scale, ensuring a continuous representation of the high-risk time slots within the reinforcement learning interaction rhythm. In this way, the reinforcement learning interaction rhythm begins to establish a direct connection with the misjudgment trigger list, enabling subsequent freeze point setting operations to proceed directly from this foundation.

[0039] After the reinforcement learning interaction rhythm has been aligned with the misjudgment trigger list, high-risk time slots in the misjudgment trigger list are set as pre-distortion parameter update freeze points to intervene in the reinforcement learning behavior, causing it to pause parameter updates at these freeze points. During the freeze point setting process, the starting position of the high-risk time slot is used as the first time position of the freeze point, extending to the end position of the high-risk time slot, ensuring the freeze point covers the entire high-risk time slot in the time dimension. Within the freeze point range, reinforcement learning no longer performs any adjustments to the pre-distortion parameters, but maintains the parameter values ​​from the previous stage, preventing the dominant interference segment from affecting the current pre-distortion parameters. The freeze point setting prevents reinforcement learning from relying on feedback signals for behavioral decisions during periods of dominant external interference, thus avoiding the misinterpretation of feedback signal waveform fluctuations caused by interference as natural changes in the amplifier's operating state. After the freeze point setting is completed, the content of the high-risk time slots is fully embedded into the reinforcement learning interaction rhythm, making the freeze point a direct intervention to block misjudgment behavior, and also providing a positional basis for subsequent silent sampling window insertion.

[0040] After the freeze point has been successfully set, to prevent reinforcement learning from completely losing its reference for feedback signals during the freeze point and thus being unable to determine whether the interference has ended, a silent sampling window is inserted within the freeze point. This allows the feedback signals during the freeze point to be recorded in a non-updating manner. During the insertion of the silent sampling window, its starting position is set at the beginning of the freeze point, ensuring that the recording of the feedback signal's trajectory begins immediately upon the freeze point opening. Simultaneously, the duration of the silent sampling window is kept consistent with the duration of the freeze point, ensuring that the feedback signals throughout the entire high-risk time slot are fully recorded. The feedback signals collected by the silent sampling window do not affect the predistortion parameter update; instead, they serve as a reference for tracking the interference's changing trend within the freeze point, creating a feedback signal recording trajectory independent of the reinforcement learning parameter update mechanism. In this way, the insertion of the silent sampling window transforms the freeze point from a completely blank period into one with real-time feedback signal recording capabilities. This allows subsequent determination of whether the dominant interference segment continues based on the records within the silent sampling window, enabling the freeze point to continuously track interference behavior. After the silent sampling window is inserted, the behavior inside the frozen point is expressed more completely, providing a source of interference signal trajectories for the injection of the reverse reward mask.

[0041] After the silent sampling window is inserted into the freeze point and has continuous recording capability, to ensure that reinforcement learning parameter updates after the freeze point avoid external interference-dominant segments, a reverse reward mask is injected into the freeze point. This marks the feedback signals during the freeze point as time periods where parameter updates are not feasible in the reinforcement learning behavior record. The injection of the reverse reward mask is based on the feedback signal change trajectory recorded by the silent sampling window, marking all time positions within the freeze point as unsuitable for predistortion parameter updates. This allows reinforcement learning to automatically ignore the behavior records within the freeze point when resuming parameter updates after the freeze point, avoiding using the feedback signals within the freeze point as a basis for policy decisions. Simultaneously, the injection of the reverse reward mask creates a prohibited update trajectory in the reinforcement learning behavior record that overlaps with the interference-dominant segment. During subsequent predistortion adjustment, these masked positions are actively avoided, ensuring that the predistortion adjustment process completely avoids the interference-dominant segment. By injecting a reverse reward mask, the behavior inside the freeze point is made more recognizable in the reinforcement learning behavior sequence, enabling reinforcement learning to avoid interference segments when re-performing parameter updates, thereby improving the accuracy of predistortion parameter updates and the stability of the entire linearization adjustment link.

[0042] This invention constructs a comprehensive reference model with unified time and frequency scales before predistortion adjustment. Based on this model, it extracts an external interference rhythm index and constructs an interference fingerprint draft. Then, it combines this with reinforcement learning decision history to identify misjudged trigger behaviors, enabling the interference factors and the actual operating state of the power amplifier to be distinguished temporally. This allows reinforcement learning to identify the source of external disturbances when processing feedback information. Therefore, reinforcement learning will not enter an erroneous update process during periods of dominant interference, and the predistortion parameters can remain continuous, ensuring the stability of predistortion adjustment even in dynamic environments.

[0043] This invention sets a predistortion parameter update freeze point based on the misjudgment trigger list, and inserts a silent sampling window and a reverse reward mask at the freeze point. This allows reinforcement learning to pause policy adjustments and independently record the interference trajectory during the interference period, enabling subsequent parameter updates to avoid the interference-affected time period. Through this method, the predistortion parameter update path is more reliable, and the power amplifier's linearization state remains balanced under interference conditions, thus improving output signal quality and linearization stability.

[0044] This invention provides, for example Figure 2 The power amplifier dynamic predistortion optimization processing system fused with reinforcement learning, as shown, includes a reference construction module, an interference rhythm extraction module, an interference fingerprint generation module, a misjudgment trigger recognition module, and a dynamic adjustment execution module. By referring to the construction module, feedback signals and environmental change signs during the operation of the power amplifier are collected, time scale and frequency band scale are unified, and a comprehensive reference draft of feedback signals and environmental change signs is generated to establish a common reference for subsequent comparisons. The interference rhythm extraction module compares environmental change signs and feedback signal waveform fluctuations with the comprehensive reference draft, extracts the external interference rhythm index under the common reference, and organizes the power amplifier stable segment in the comprehensive reference draft into a comparison segment for subsequent anomaly identification. The interference fingerprint generation module relies on the external interference rhythm index to trace back the comparison segment, locates the abnormal segment dominated by external interference at the end of the comparison segment, generates an interference fingerprint draft, and records the frequency band position and time span of the interference in the interference fingerprint draft for subsequent mapping analysis with reinforcement learning records. The misjudgment trigger identification module, based on the comparison of interference fingerprint drafts with the reinforcement learning decision history, maps the frequency band position and time span of the interference fingerprint draft records to the decision history item by item. In the mapping results, it identifies reward mutations and decision reversals, forms a misjudgment trigger list, and marks high-risk time slots in the misjudgment trigger list for reinforcement learning adjustment strategies. The dynamic adjustment execution module dynamically adjusts the reinforcement learning interaction rhythm based on the misjudgment trigger list. High-risk time slots in the misjudgment trigger list are set as predistortion parameter update freeze points. A silent sampling window is inserted at the freeze point and a reverse reward mask is injected to guide subsequent predistortion adjustment to avoid external interference-dominated segments, thereby improving the accuracy of predistortion parameter updates and the linearization stability of the system.

[0045] The power amplifier dynamic predistortion optimization processing method fused with reinforcement learning provided in this embodiment of the invention is implemented through the aforementioned power amplifier dynamic predistortion optimization processing system fused with reinforcement learning. For details of the specific methods and processes of the power amplifier dynamic predistortion optimization processing system fused with reinforcement learning, please refer to the embodiments of the power amplifier dynamic predistortion optimization processing method fused with reinforcement learning, which will not be repeated here.

[0046] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A power amplifier dynamic predistortion optimization processing method integrating reinforcement learning, characterized in that, Includes the following steps: Collect feedback signals and environmental change indicators during power amplifier operation, unify time and frequency scales, and generate a comprehensive reference draft of feedback signals and environmental change indicators; Based on the comparison of environmental change signs and feedback signal waveform fluctuations in the comprehensive reference draft, the external interference rhythm index is extracted under common reference, and the power amplifier stability segment in the comprehensive reference draft is organized into a comparison segment. By relying on the external interference rhythm index to trace back the comparison segment, the abnormal segment dominated by external interference is located at the end of the comparison segment, an interference fingerprint draft is generated, and the frequency band position and time span of the interference are recorded in the interference fingerprint draft. By reinforcing the decision history of interference fingerprint draft comparison, the frequency band position and time span of interference fingerprint draft records are mapped to the decision history item by item. In the mapping results, reward mutation and decision reversal phenomena are identified to form a misjudgment trigger list and high-risk time slots are marked in the misjudgment trigger list. The interaction rhythm of reinforcement learning is dynamically adjusted based on the misjudgment trigger list. High-risk time slots in the misjudgment trigger list are set as predistortion parameter update freeze points. Silent sampling windows are inserted at the freeze points and inverse reward masks are injected to guide subsequent predistortion adjustments to avoid external interference dominating segments.

2. The power amplifier dynamic predistortion optimization method based on reinforcement learning according to claim 1, characterized in that, The steps for generating a comprehensive reference draft are as follows: The system collects feedback signals during the operation of the power amplifier and simultaneously collects signs of environmental changes. The changes in amplitude, phase, and frequency band of the feedback signals and the changes in signs of environmental changes are recorded and formed into a basic acquisition sequence according to a common acquisition cycle. Based on the basic acquisition sequence, feedback signals and signs of environmental change are uniformly labeled on the time scale and frequency band scale, and the acquisition points of feedback signals and signs of environmental change are arranged according to a common step size to form intermediate material; In the intermediate material, the changes in the waveform of the feedback signal and the changes in the signs of environmental change are arranged together, and the two are combined and presented in a unified time scale and a unified frequency band scale to form a comprehensive reference draft; The time and frequency scales of the comprehensive reference draft are continuously read and extended to the entire operating cycle. In the subsequent comparison stage, the comprehensive reference draft is used as a common reference to carry out correlation analysis between feedback signals and signs of environmental changes.

3. The power amplifier dynamic predistortion optimization method based on reinforcement learning according to claim 2, characterized in that, The process of extracting the external interference rhythm index is as follows: Under the condition that the time scale and frequency band scale of the comprehensive reference draft are unified, the fluctuations of the feedback signal waveform and the signs of environmental changes are analyzed segment by segment to form potential interference-related candidate segments. Based on the candidate fragments associated with potential interference, the preceding and subsequent change trajectories are read in the comprehensive reference manuscript. The changes in the prelude of interference, the derivative changes of interference, and the decay changes of interference are extracted and organized into an external interference rhythm index. After the external interference rhythm index is formed, extract the segments with continuous fluctuations in the feedback signal waveform and smooth signs of environmental change from the comprehensive reference manuscript and organize them into control segments so that the control segments maintain the scale of the comprehensive reference manuscript. The external interference rhythm index and the control segment are aligned and expanded in the comprehensive reference manuscript to enable anomaly identification of interference behavior and natural running behavior under common reference conditions.

4. The power amplifier dynamic predistortion optimization processing method based on reinforcement learning according to claim 3, characterized in that, During the alignment and expansion process of the external interference rhythm index and the control segment, the interference prelude changes and interference derivative changes in the external interference rhythm index are arranged continuously, and the control segments are kept in order in the same time scale and frequency band scale, so as to enhance the ability to distinguish between interference behavior and natural operation behavior under common reference conditions.

5. The power amplifier dynamic predistortion optimization method based on reinforcement learning according to claim 3, characterized in that, The steps for generating an interference fingerprint draft are as follows: By relying on the external interference rhythm index to backtrack the control segment, the fluctuations of the feedback signal waveform of the control segment and the signs of environmental changes are read according to the time scale to form a continuous expression; Based on the peak value of the change in the external interference rhythm index, the interference change trajectory is projected to the end of the control segment, and the abnormal segment dominated by external interference is located in the time scale and frequency band scale. Record the start and end positions of the abnormal segments on the time axis as the interference time span, and record the frequency band positions covered by the abnormal segments as the interference frequency band positions, and organize them into an interference fingerprint draft. The interference fingerprint draft is aligned with the time and frequency scale of the comprehensive reference draft, so that the interference fingerprint draft can be continuously matched with the time span and frequency position of the feedback signal waveform fluctuation.

6. The power amplifier dynamic predistortion optimization method based on reinforcement learning according to claim 5, characterized in that, During the generation of the interference fingerprint draft, the time span and frequency band position of the abnormal segment are compared with the continuous change trajectory of the external interference rhythm index. Under the condition of consistent comparison, the boundary of the interference time span and the coverage range of the interference frequency band position are defined to enhance the accuracy of the interference fingerprint draft in expressing the abnormal segment dominated by external interference.

7. The power amplifier dynamic predistortion optimization method based on reinforcement learning according to claim 5, characterized in that, The steps for generating the false alarm trigger list are as follows: After the interference fingerprint draft is formed, the history of reinforcement learning decisions is read along the interference time span as the main line, and the records of reinforcement learning in the time scale are matched with the interference time span segment by segment. Based on the interference time span and interference frequency band location, the decision records in the reinforcement learning decision history are mapped item by item on the time scale and frequency band scale to form a continuous behavioral trajectory of interference entry, interference duration and interference decay. Identify reward mutations and decision reversals that occur during interference in reinforcement learning within the mapped trajectory, and record and compile these changes into a list of misjudgment triggers. High-risk time slots are marked on the time scale based on the misjudgment trigger list. The reward mutation position and the decision reversal position are mapped to the interference time span, so that the high-risk time slots are marked in the reinforcement learning adjustment strategy for subsequent adjustment.

8. The power amplifier dynamic predistortion optimization method based on reinforcement learning according to claim 7, characterized in that, When marking high-risk time slots, the reward mutation position and decision reversal position are simultaneously marked on the frequency band scale, with the interference frequency band position in the interference fingerprint draft as a reference. The simultaneously marked position is then associated with the corresponding time period in the interference time span, so that the high-risk time slots form a joint location on the time scale and frequency band scale for reinforcement learning adjustment strategies.

9. The power amplifier dynamic predistortion optimization method based on reinforcement learning according to claim 7, characterized in that, The steps for dynamically adjusting the interaction rhythm of reinforcement learning are as follows: Arrange high-risk time slots according to the list of misjudgments, and map the high-risk time slots to the parameter update nodes of the reinforcement learning interaction rhythm on the time scale to form continuous markings; Set a pre-distortion parameter update freeze point at the node corresponding to the high-risk time slot, and extend the coverage of the freeze point along the time scale to the end of the high-risk time slot to pause parameter updates. A silent sampling window is inserted inside the freezing point. The silent sampling window is expanded according to the starting position of the freezing point and the change trajectory of the feedback signal during the freezing point is recorded. Based on the feedback signal change trajectory formed by the silent sampling window, an inverse reward mask is injected inside the freeze point to mark the record inside the freeze point as a period that cannot be used for predistortion parameter update, so as to guide subsequent predistortion adjustment to avoid external interference dominating the segment.

10. A power amplifier dynamic predistortion optimization processing system incorporating reinforcement learning, used to implement the power amplifier dynamic predistortion optimization processing method incorporating reinforcement learning as described in any one of claims 1-9, characterized in that, This includes a reference construction module, an interference rhythm extraction module, an interference fingerprint generation module, a false alarm trigger identification module, and a dynamic adjustment execution module. Referring to the construction module, feedback signals and environmental change signs during the operation of the power amplifier are collected, time scale and frequency band scale are unified, and a comprehensive reference draft of feedback signals and environmental change signs is generated; The interference rhythm extraction module, based on the comprehensive reference draft, compares environmental change signs and feedback signal waveform fluctuations, extracts the external interference rhythm index under common reference, and organizes the power amplifier stable segment in the comprehensive reference draft into a comparison segment. The interference fingerprint generation module relies on the external interference rhythm index to trace back the comparison segment, locates the abnormal segment dominated by external interference at the end of the comparison segment, generates an interference fingerprint draft, and records the frequency band position and time span of the interference in the interference fingerprint draft. The misjudgment trigger identification module, based on the reinforcement learning decision history of interference fingerprint draft comparison, maps the frequency band position and time span of interference fingerprint draft records to the decision history item by item. In the mapping results, it identifies reward mutation and decision reversal phenomena, forms a misjudgment trigger list, and marks high-risk time slots in the misjudgment trigger list. The dynamic adjustment execution module dynamically adjusts the interaction rhythm of reinforcement learning based on the misjudgment trigger list. High-risk time slots in the misjudgment trigger list are set as predistortion parameter update freeze points. A silent sampling window is inserted at the freeze point and a reverse reward mask is injected to guide subsequent predistortion adjustment to avoid external interference dominating the segment.

Images