An online incremental learning and adaptive update method and system

By using an online incremental learning and adaptive update system, the artificial intelligence model in the laser pointing control system is monitored and updated in real time, which solves the model drift problem and improves the system's stability and continuous operation capability.

CN123067151APending Publication Date: 2026-11-10JIANGSU STARLINK LASER TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611352469.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing laser pointing control systems, the artificial intelligence prediction model drifts due to changes in the environment and equipment, affecting the control effect. It lacks a real-time monitoring and automatic update mechanism, making it difficult to guarantee high real-time performance and stability.

Method used

An online incremental learning and adaptive update system is set up in the laser pointing control system. The system acquires state data through the main control loop, monitors model performance in real time, generates update trigger signals, and performs incremental learning and model updates while the main control loop is running continuously.

Benefits of technology

This enables real-time performance monitoring and timely updates of the model, improving the continuous operation capability and stability of the laser pointing control system and reducing the impact of model updates on the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123067151A_ABST
    Figure CN123067151A_ABST
Patent Text Reader

Abstract

This invention discloses an online incremental learning and adaptive update method and system, relating to the field of intelligent laser control technology. It includes a main control loop and an incremental learning subsystem connected to it. By collecting laser system operating status data and inputting it into an artificial intelligence prediction model, the system obtains prediction results of laser pointing changes, driving the laser pointing adjustment mechanism to make adjustments. Simultaneously, it caches prediction data, actual output data, and error data. Based on model performance evaluation parameters, it detects model drift and, when drift occurs, calls upon the operating status data for online incremental learning. After performance verification, the model update and deployment are completed. This invention enables online monitoring, adaptive adjustment, and reliable updating of the artificial intelligence prediction model without affecting the normal operation of the control system, improving the prediction accuracy and system stability during long-term model operation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent laser control technology, and in particular to an online incremental learning and adaptive update method and system. Background Technology

[0002] With the development of high-precision laser processing, laser communication, laser measurement, and high-energy laser systems, the pointing stability of the laser beam has become a crucial factor affecting system performance. In practical applications, lasers and their optical transmission paths are easily affected by factors such as mechanical vibration, ambient temperature changes, device thermal drift, and external disturbances, causing dynamic deviations in the laser beam. Therefore, existing laser pointing control systems typically employ sensors to collect laser state information in real time and combine this with control algorithms to adjust the actuators to achieve stable control of the laser beam direction.

[0003] With the development of artificial intelligence technology, laser pointing control methods based on AI predictive models are gradually being applied to the aforementioned systems. These methods typically utilize multi-source sensors to collect state data during the laser system's operation, and then use AI predictive models to predict the laser pointing trend. Based on the prediction results, compensation control commands are generated to improve the response speed and control accuracy of laser pointing control.

[0004] However, existing laser pointing control systems based on artificial intelligence prediction models typically require pre-training of the model using historical operating data and then deploying the trained model into the control system. During long-term actual operation, the laser system's operating environment, equipment status, and external disturbances constantly change, potentially causing discrepancies between the actual operating data and the data distribution corresponding to the model training phase. This leads to model drift in the artificial intelligence prediction model, causing the model's predictions to gradually deviate from the actual operating conditions, thus affecting the laser pointing control performance.

[0005] In existing technologies, restoring the predictive performance of an artificial intelligence model typically requires re-collecting data and performing offline training before replacing the retrained model in the control system. However, this approach usually necessitates stopping or pausing the original control process, impacting the continuous operation capability of the laser system and making it difficult to respond promptly to changes in the operating environment. Furthermore, due to the lack of a real-time monitoring mechanism for the model's operational status, existing technologies struggle to determine promptly whether the AI ​​prediction model has experienced performance degradation, and cannot automatically trigger model updates based on the model's actual operational status.

[0006] On the other hand, although some artificial intelligence systems use online learning or incremental learning to update models, they usually focus on maintaining model performance in general data processing scenarios and do not analyze the correlation between predicted values, actual output values ​​and prediction errors in the laser pointing control process. Therefore, they are difficult to apply directly to laser pointing control systems that require high real-time performance and high stability.

[0007] Therefore, existing technologies lack a systematic approach to solving the drift problem of AI laser pointing stability control models, and in particular, lack a complete technical solution that can monitor model prediction performance in real time, automatically detect model drift, and trigger incremental learning updates as needed. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, the present invention provides an online incremental learning and adaptive update method and system to solve the problems existing in the background art.

[0009] This invention provides the following technical solution: an online incremental learning and adaptive update method and system, comprising: During the operation of the laser pointing control system, the laser system operating status data is acquired through the main control loop, and the operating status data is input into the artificial intelligence prediction model to obtain the laser pointing change prediction result. Based on the laser pointing change prediction result, a control command is generated and the laser pointing adjustment mechanism is driven to perform pointing adjustment. At the same time, the output status of the laser system is acquired as feedback data. During the operation of the main control loop, the predicted data output by the artificial intelligence prediction model, the real data corresponding to the output state of the laser system, and the error data between the predicted data and the real data are cached. The performance of the artificial intelligence prediction model is monitored based on the cached error data to obtain model performance evaluation parameters. The model performance evaluation parameters are compared with the baseline performance parameters corresponding to the stable operation state of the artificial intelligence prediction model. When the model performance evaluation parameters meet the preset drift conditions, it is determined that the artificial intelligence prediction model has drifted, and a model update trigger signal is generated. In response to the model update trigger signal, while the main control loop is running continuously, cached running status data is used as incremental learning samples to perform online incremental learning on the artificial intelligence prediction model; The performance of the artificial intelligence prediction model after incremental learning is verified. When the verification result meets the preset update conditions, the updated artificial intelligence prediction model is deployed to the main control loop.

[0010] Preferably, the main control loop includes a multi-source sensor, an artificial intelligence prediction model, a laser pointing adjustment mechanism, and a laser system output module; The multi-source sensor is used to collect state data during the laser pointing control process, the artificial intelligence prediction model is used to predict the laser pointing change trend based on the state data, the laser pointing adjustment mechanism is used to adjust the laser propagation direction based on the prediction result, and the laser system output module is used to output the adjusted laser pointing state.

[0011] Preferably, the status data collected by the multi-source sensor includes at least one of laser beam position data, vibration status data, ambient temperature data, and laser intensity data.

[0012] Preferably, the artificial intelligence prediction model predicts the trend of laser pointing change within a future time period based on continuous time series state data; The artificial intelligence prediction model includes at least one of the following: neural network model, deep learning model, and time series prediction model.

[0013] Preferably, the artificial intelligence prediction model includes an input processing unit, a feature extraction unit, and a prediction output unit; The input processing unit is used to divide the continuously acquired state data into time segments, the feature extraction unit is used to extract state features of different time segments, and the prediction output unit is used to output the laser pointing prediction result.

[0014] Preferably, the model performance evaluation parameters include at least one of the root mean square value of prediction error, the average value of prediction error, the standard deviation of prediction error, and the maximum value of prediction error; the performance change state of the artificial intelligence prediction model is determined based on the difference between the model performance evaluation parameters and the preset benchmark performance parameters.

[0015] Preferably, the determination of whether the artificial intelligence prediction model has experienced model drift is further based on the analysis of disturbance state data during the operation of the laser system, obtaining the spectral characteristics corresponding to the disturbance state data, and using the changes in the spectral characteristics to assist in determining whether the artificial intelligence prediction model has experienced model drift.

[0016] Preferably, the online incremental learning uses the model parameters of the current artificial intelligence prediction model as initialization parameters, and adjusts the parameters of the artificial intelligence prediction model based on the cached new operating status data, so that the updated artificial intelligence prediction model adapts to the current operating status of the laser pointing control system.

[0017] Preferably, the performance verification evaluates the predictive performance of the AI ​​prediction model before and after incremental learning based on verification data, and determines whether the AI ​​prediction model after incremental learning meets the deployment conditions based on the prediction performance evaluation results; the AI ​​prediction model performs model update when it meets the deployment conditions, and retains the AI ​​prediction model before incremental learning when it does not meet the deployment conditions.

[0018] An online incremental learning and adaptive update system includes a main control loop and an incremental learning subsystem connected to the main control loop; The main control loop includes: A multi-source sensing module is used to collect state data during the laser pointing control process; An artificial intelligence prediction model module is connected to the multi-source sensing module and is used to output a laser pointing change prediction result based on the state data. A laser pointing adjustment module, connected to the artificial intelligence prediction model module, is used to adjust the laser propagation direction according to the prediction result; The laser system output module is used to output the adjusted laser pointing state; The incremental learning subsystem includes: A data caching module, connected to the artificial intelligence prediction model module and the laser system output module, is used to cache prediction data, real data, and error data between prediction data and real data. The performance monitoring module is used to calculate the performance evaluation parameters of the artificial intelligence prediction model based on the error data; The drift detection module is used to determine whether the artificial intelligence prediction model has drifted based on the performance evaluation parameters, and to output a model update trigger signal. The incremental learning module is used to call cached data according to the model update trigger signal and perform online incremental learning on the artificial intelligence prediction model. The model management module is used to validate, deploy, and roll back the model version in case of anomalies in the incremental learning of the artificial intelligence prediction model.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention sets up an online incremental learning and adaptive update subsystem on the main control loop of the laser pointing control system, so that the artificial intelligence prediction model can perform state monitoring and adaptive optimization by using the prediction data, actual output data and prediction error data generated during operation while performing laser pointing control tasks. The data caching module continuously collects the operating status data of the main control loop, and the performance monitoring module analyzes the prediction performance of the artificial intelligence prediction model, thereby realizing real-time perception of the model's operating status. 2. This invention establishes an artificial intelligence prediction model drift detection mechanism by comparing the model performance evaluation parameters with the benchmark performance parameters under stable operating conditions. It can promptly identify the model degradation state when the model prediction performance declines but has not yet seriously affected the control effect, and generate a model update trigger signal. Compared with the method of relying on manual periodic detection or offline retraining, it improves the timeliness of model maintenance. 3. In the continuous operation of the main control loop, this invention uses data cached during system operation as incremental learning samples to perform online incremental learning on the artificial intelligence prediction model. After verification, the updated model is redeployed to the main control loop, enabling the model to adapt to changes in the laser system's operating environment and equipment status. This avoids the problem of needing to pause the control system during traditional model updates and improves the continuous operation capability of the laser pointing control system. 4. This invention evaluates the performance of the AI ​​prediction model after incremental learning by setting up model verification and model management mechanisms, and performs model version rollback when update anomalies occur, making the AI ​​prediction model update process controllable and safe, and reducing the impact of model update failure on the operational stability of the laser pointing control system. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the online incremental learning and adaptive update method provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the online incremental learning and adaptive update system structure provided in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the performance monitoring and drift detection process provided in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the incremental learning and model update process provided in an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of the model drift detection spectrum analysis provided in an embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram comparing the prediction error before and after incremental learning, provided in an embodiment of the present invention. Detailed Implementation

[0026] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0027] This invention provides an online incremental learning and adaptive update method, which is applied to a laser pointing control system to solve the problem that the prediction performance of existing laser pointing control systems based on artificial intelligence prediction models deteriorates during long-term operation due to changes in the distribution of model input data caused by environmental changes, equipment status changes, and external disturbances.

[0028] An online incremental learning and adaptive update method, such as Figure 1 As shown, it includes the following steps: S1: During the operation of the laser pointing control system, the laser system operating status data is acquired through the main control loop, and the operating status data is input into the artificial intelligence prediction model to obtain the laser pointing change prediction result.

[0029] The main control loop includes a multi-source sensor, an artificial intelligence prediction model, a laser pointing adjustment mechanism, and a laser system output module.

[0030] Multi-source sensors are used to collect state data during the laser pointing control process. The state data includes at least one of laser beam position data, vibration state data, ambient temperature data, and laser intensity data.

[0031] Specifically, beam position data is used to reflect the current spatial position changes of the laser beam; vibration state data is used to reflect the impact of mechanical disturbances on the laser, optical components, and mounting structure; ambient temperature data is used to reflect the impact of temperature changes on the operating state of the laser system; and laser intensity data is used to reflect changes in the laser output state.

[0032] By collecting various types of operational status data, the artificial intelligence prediction model can comprehensively consider the laser system's own status as well as changes in the external environment, thereby improving the completeness of the input data.

[0033] Furthermore, the artificial intelligence prediction model predicts the trend of laser pointing changes in future time periods based on continuous time series state data.

[0034] In this embodiment, the artificial intelligence prediction model can be at least one of a neural network model, a deep learning model, and a time series prediction model.

[0035] The artificial intelligence prediction model includes an input processing unit, a feature extraction unit, and a prediction output unit.

[0036] The input processing unit is used to divide the continuously acquired state data into time segments and convert the continuous running state data into an input sequence with time correlation; the feature extraction unit is used to extract the change features corresponding to the state data in different time segments; and the prediction output unit is used to output the laser pointing change prediction result based on the extracted state features.

[0037] S2: Generate control commands based on the laser pointing change prediction results, drive the laser pointing adjustment mechanism to perform pointing adjustment, and simultaneously acquire the laser system output status as feedback data.

[0038] Specifically, based on the prediction results output by the artificial intelligence prediction model, the control unit generates corresponding control commands and sends them to the laser pointing adjustment mechanism, so that the actuator adjusts the laser propagation direction.

[0039] In one embodiment, the laser pointing adjustment mechanism includes a control unit and an execution unit.

[0040] The control unit generates control commands based on the prediction results output by the artificial intelligence prediction model; the execution unit adjusts the propagation direction of the laser beam according to the control commands.

[0041] Furthermore, the control unit includes a feedback control module, which generates a feedback adjustment amount based on the deviation between the laser system output state and the target state.

[0042] Among them, the artificial intelligence prediction model is used to predict future trends based on historical operating patterns, and the feedback control module is used to correct errors based on the actual output state, thereby improving the stability of the laser pointing control process.

[0043] In one embodiment, the execution unit includes a piezoelectric ceramic driven fast deflection mirror, which generates micro-displacement through the piezoelectric ceramic driving structure to achieve rapid adjustment of the laser beam direction.

[0044] S3: During the operation of the main control loop, the predicted data output by the artificial intelligence prediction model, the real data corresponding to the output state of the laser system, and the error data between the predicted data and the real data are cached.

[0045] Specifically, the prediction results output by the artificial intelligence prediction model are compared with the actual output state of the laser system to obtain prediction error data.

[0046] The prediction error data can reflect the current operating performance of the artificial intelligence prediction model.

[0047] By continuously caching predicted data, real data, and error data, a data foundation is provided for subsequent model performance evaluation, model drift detection, and incremental learning.

[0048] S4: Monitor the performance of the AI ​​prediction model based on cached error data to obtain model performance evaluation parameters.

[0049] The model performance evaluation parameters include at least one of the following: root mean square prediction error, average prediction error, standard deviation of prediction error, and maximum prediction error.

[0050] The performance monitoring module determines the changes in the current predictive ability of the artificial intelligence prediction model based on the above evaluation parameters.

[0051] For example, when the prediction error of an artificial intelligence prediction model gradually increases during long-term operation, it indicates that the current model may not be able to accurately describe the new operating state.

[0052] S5: Compare the model performance evaluation parameters with the baseline performance parameters corresponding to the stable operation state of the artificial intelligence prediction model. When the model performance evaluation parameters meet the preset drift conditions, it is determined that the artificial intelligence prediction model has drifted, and a model update trigger signal is generated. The overall process of performance monitoring and drift detection is as follows: Figure 3 As shown.

[0053] Specifically, the baseline performance parameters of the artificial intelligence prediction model under stable operating conditions are obtained in advance, and the model performance evaluation parameters calculated in real time are compared with the baseline performance parameters.

[0054] When the difference between the two exceeds the preset range, it indicates that the current running state data distribution has deviated from the data distribution during the model training stage, and the artificial intelligence prediction model experiences performance degradation, i.e., model drift occurs.

[0055] Furthermore, to improve the accuracy of model drift detection, the disturbance state data during the operation of the laser pointing control system can be used for auxiliary judgment.

[0056] Specifically, by acquiring the spectral characteristics corresponding to disturbance data such as vibration state data and temperature change data, and judging whether the external operating environment has changed based on the changes in spectral characteristics, the process of spectral analysis is as follows: Figure 5 As shown.

[0057] S6: In response to the model update trigger signal, while the main control loop is continuously running, cached runtime data is used as incremental learning samples to perform online incremental learning on the artificial intelligence prediction model. Figure 4 As shown.

[0058] Specifically, incremental learning uses the model parameters of the current artificial intelligence prediction model as initialization parameters and adjusts the model parameters using cached new running status data.

[0059] The incremental learning samples include state data collected by multiple sensors, predicted data output by the artificial intelligence prediction model, actual output data of the laser system, and error data between the predicted data and the actual data.

[0060] Compared to re-collecting all historical data for model training, this embodiment adopts an incremental learning approach based on newly added operational status data, which only optimizes the current model, reducing the computational load of model updates and avoiding impacting the normal operation of the laser pointing control system.

[0061] S7: Perform performance verification on the AI ​​prediction model after incremental learning. When the verification results meet the preset update conditions, deploy the updated AI prediction model to the main control loop.

[0062] Specifically, performance verification evaluates the predictive performance of the AI ​​prediction model before and after incremental learning based on verification data, and determines whether the updated AI prediction model meets the deployment conditions based on the prediction performance evaluation results.

[0063] When the updated AI prediction model meets the deployment conditions, the updated model is loaded into the main control loop, allowing it to replace the original model in performing the prediction task.

[0064] If the updated AI prediction model does not meet the deployment conditions, the AI ​​prediction model before incremental learning will be retained to avoid abnormal model updates affecting the operational stability of the laser pointing control system.

[0065] The present invention also provides an online incremental learning and adaptive update system for implementing the above method.

[0066] like Figure 2 As shown, the system includes a main control loop and an incremental learning subsystem connected to the main control loop.

[0067] The main control loop includes a multi-source sensing module, an artificial intelligence prediction model module, a laser pointing adjustment module, and a laser system output module. It should be noted that the multi-source sensing module is the aforementioned multi-source sensor, and the laser pointing adjustment module is the aforementioned laser pointing adjustment mechanism.

[0068] The multi-source sensing module is used to collect state data during the laser pointing control process; The artificial intelligence prediction model module is connected to the multi-source sensing module and is used to output the prediction result of laser pointing change based on the state data; The laser pointing adjustment module is connected to the artificial intelligence prediction model module and is used to adjust the laser propagation direction according to the prediction results; The laser system output module is used to output the adjusted laser pointing state.

[0069] The incremental learning subsystem includes a data caching module, a performance monitoring module, a drift detection module, an incremental learning module, and a model management module.

[0070] The data caching module is connected to the artificial intelligence prediction model module and the laser system output module to cache predicted data, real data, and error data between predicted data and real data.

[0071] The performance monitoring module is used to calculate the performance evaluation parameters of the artificial intelligence prediction model based on the cached error data.

[0072] The drift detection module is used to determine whether the artificial intelligence prediction model has drifted based on performance evaluation parameters and outputs a model update trigger signal.

[0073] The incremental learning module is used to retrieve cached data based on model update trigger signals to perform online incremental learning on the artificial intelligence prediction model.

[0074] The model management module is used to validate, deploy, and roll back the model version in case of anomalies in the incremental learning of the artificial intelligence prediction model.

[0075] Furthermore, the multi-source sensing module includes a beam position detection unit, a vibration detection unit, a temperature detection unit, and a light intensity detection unit.

[0076] The laser pointing adjustment module includes a control unit and an execution unit. The control unit is used to generate control commands based on the output results of the artificial intelligence prediction model, and the execution unit is used to adjust the propagation direction of the laser beam according to the control commands.

[0077] The execution unit can use piezoelectric ceramics to drive a fast deflecting mirror.

[0078] The model management module includes a model validation unit, a model deployment unit, and a model rollback unit.

[0079] The model verification unit is used to determine whether the updated artificial intelligence prediction model meets the operational requirements; the model deployment unit is used to load the verified artificial intelligence prediction model into the main control loop; and the model rollback unit is used to restore the historical model version in case of model anomalies.

[0080] With the above structure, this embodiment can achieve online drift detection, adaptive updating, and secure deployment of the artificial intelligence prediction model while maintaining the continuous operation of the laser pointing control system, thereby improving the stability and reliability of the artificial intelligence prediction model during long-term operation.

[0081] Example 1

[0082] This embodiment takes a 1Hz high-power petawatt-level laser system that uses a Patch-MLP prediction model for beam pointing stability control as an example to describe in detail the specific implementation process of the present invention.

[0083] (a) Initial state of the system The system has deployed a pre-trained Patch-MLP prediction model with an input window length of 2000 sampling points (corresponding to approximately 2 seconds) and a prediction offset of 5 sampling points (corresponding to approximately 5 ms). The model runs continuously at a frequency of 1 kHz, outputting the predicted beam direction at future moments to drive a piezoelectric ceramic fast deflection mirror for feedforward compensation. The root mean square error of the prediction during normal operation is approximately 0.5 microradians.

[0084] (ii) Data caching After system startup, the data caching module continuously caches runtime status data. The cached data includes: the actual beam pointing position (x, y) at each sampling moment, the model prediction value (x̂, ŷ), the prediction error (e_x, e_y), and the corresponding timestamp. The circular buffer is set to store 10 minutes of data (i.e., 600,000 sampling points). When the buffer is full, new data overwrites the oldest data.

[0085] (III) Performance monitoring and benchmark establishment During the stable operation phase (first 30 minutes) immediately after model training, the performance monitoring module calculates statistical indicators of prediction error to establish a performance benchmark. Specifically, with a sliding window length of 1 minute, the root mean square (RMS) value of the prediction error within each window is calculated, followed by the average and standard deviation of the RMS values ​​across all windows. The calculated benchmark RMS value is μ0 = 0.50 microradians, and the standard deviation σ0 = 0.05 microradians. Simultaneously, a Fast Fourier Transform (FFT) is performed on the beam pointing data to extract the dominant frequency components and their amplitudes, serving as the spectral benchmark.

[0086] (iv) Drift detection After the system ran for several hours, a newly installed vacuum pump in the laboratory started working, introducing a new vibrational frequency component of approximately 67.7 Hz. This new vibration caused increased beam pointing jitter, and the prediction error of the AI ​​prediction model gradually increased.

[0087] The drift detection module continuously calculates the current root mean square (RMS) value of the prediction error in a 1-minute sliding window. When the RMS value for three consecutive windows exceeds 1.3 times the baseline RMS value (i.e., 0.65 microradians), a first warning is triggered. Simultaneously, the spectrum analysis function detects a new significant frequency component at 67.7 Hz (a frequency not present in the baseline spectrum), triggering a second warning. When both detection mechanisms trigger warnings, the drift detection module generates an updated trigger signal.

[0088] (v) Incremental learning After receiving the update trigger signal, the incremental learning module extracts the most recent 5 minutes of data (300,000 sampling points) from the data cache module as the incremental training set. The data is then divided into a training subset (4 minutes) and a validation subset (1 minute) in an 8:2 ratio.

[0089] Using the currently deployed model weights as initial values, 10% of the initial training learning rate (0.0001) (i.e., 0.00001) was used as the incremental learning rate for 5 epochs of training. At the end of each epoch, the loss on the validation subset was calculated. After training, the validation set loss decreased by approximately 15% compared to before training, indicating that the incremental learning was successful.

[0090] (vi) Model Deployment The model management module deploys the new model after incremental learning as the current running version, while saving the old model as a rollback version. After the new model is deployed, the root mean square error of the prediction error recovers from 0.65 microradians to approximately 0.52 microradians, approaching the baseline level. The comparison of prediction errors before and after incremental learning is as follows: Figure 6 As shown. The system continuously monitors the performance of the new model during subsequent operation. If no anomalies are found within 1 hour, the update is confirmed to be successful.

[0091] The entire update process (from drift detection to model deployment) was completed in about 3 minutes, during which the system's main control loop continued to operate without any control interruption.

[0092] Example 2

[0093] This embodiment illustrates the rollback mechanism of the present invention when model update fails.

[0094] In another run, the system detected an increase in prediction error (possibly due to an unexpected performance mutation in the optical components), triggering incremental learning. However, because the statistical characteristics of the new data differed significantly from the original data, the validation set loss after incremental learning actually increased compared to before training (i.e., the model was unable to effectively fit the new data while retaining old knowledge).

[0095] At this point, the incremental learning module determines that the update has failed, while the model management module retains the current model without deployment. Simultaneously, the system sends a maintenance notification to the operator, suggesting checking for hardware failure in the optical system or whether more data collection for retraining is needed. The system continues to use the existing model for control; although prediction accuracy decreases, basic control functions remain normal, avoiding the risk of system loss of control due to the deployment of an immature model.

[0096] Example 3

[0097] This embodiment illustrates the continuous monitoring and secondary updates of the present invention after successful incremental learning.

[0098] After the initial incremental learning deployment was successful, the system continued to run. The performance monitoring module continuously monitored the prediction error of the new model. After several hours of operation, the model performance remained stable. However, due to the diurnal variation in laboratory temperature, the slowly varying drift pattern of the beam direction gradually changed. Approximately 8 hours later, the root mean square value of the prediction error slowly rose again to more than 1.3 times the baseline value, triggering a second update.

[0099] The incremental learning module executes the update process again, using the model weights from the first update as initial values ​​and fine-tuning them with the latest 5-minute data. After the update, the model recovers to good predictive performance. This process can be repeated continuously, allowing the model to adapt to slow changes in the environment and system state over long-term operation.

[0100] Experiments show that, during several weeks of continuous operation, the present invention can automatically detect and adapt to multiple changes in the environment and system state, including new vibration sources, changes in temperature drift patterns, and slow degradation of optical components, and always maintain the AI ​​prediction model at a near-optimal performance level.

[0101] Several points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection" and "linkage" should be interpreted broadly, and can be mechanical or electrical connection, or internal connection between two components, or direct connection. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationship. When the absolute position of the described object changes, the relative positional relationship may change.

[0102] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. An online incremental learning and adaptive update method, characterized in that, include: During the operation of the laser pointing control system, the laser system operating status data is acquired through the main control loop, and the operating status data is input into the artificial intelligence prediction model to obtain the laser pointing change prediction result. Based on the laser pointing change prediction result, a control command is generated and the laser pointing adjustment mechanism is driven to perform pointing adjustment. At the same time, the output status of the laser system is acquired as feedback data. During the operation of the main control loop, the predicted data output by the artificial intelligence prediction model, the real data corresponding to the output state of the laser system, and the error data between the predicted data and the real data are cached. The performance of the artificial intelligence prediction model is monitored based on the cached error data to obtain model performance evaluation parameters. The model performance evaluation parameters are compared with the baseline performance parameters corresponding to the stable operation state of the artificial intelligence prediction model. When the model performance evaluation parameters meet the preset drift conditions, it is determined that the artificial intelligence prediction model has drifted, and a model update trigger signal is generated. In response to the model update trigger signal, while the main control loop is running continuously, cached running status data is used as incremental learning samples to perform online incremental learning on the artificial intelligence prediction model; The performance of the artificial intelligence prediction model after incremental learning is verified. When the verification result meets the preset update conditions, the updated artificial intelligence prediction model is deployed to the main control loop.

2. The online incremental learning and adaptive update method according to claim 1, characterized in that: The main control loop includes a multi-source sensor, an artificial intelligence prediction model, a laser pointing adjustment mechanism, and a laser system output module; The multi-source sensor is used to collect state data during the laser pointing control process, the artificial intelligence prediction model is used to predict the laser pointing change trend based on the state data, the laser pointing adjustment mechanism is used to adjust the laser propagation direction based on the prediction result, and the laser system output module is used to output the adjusted laser pointing state.

3. The online incremental learning and adaptive update method according to claim 2, characterized in that: The status data collected by the multi-source sensor includes at least one of the following: laser beam position data, vibration status data, ambient temperature data, and laser intensity data.

4. The online incremental learning and adaptive update method according to claim 1, characterized in that: The artificial intelligence prediction model predicts the trend of laser pointing change within a future time period based on continuous time series state data. The artificial intelligence prediction model includes at least one of the following: neural network model, deep learning model, and time series prediction model.

5. The online incremental learning and adaptive update method according to claim 4, characterized in that: The artificial intelligence prediction model includes an input processing unit, a feature extraction unit, and a prediction output unit; The input processing unit is used to divide the continuously acquired state data into time segments, the feature extraction unit is used to extract state features of different time segments, and the prediction output unit is used to output the laser pointing prediction result.

6. The online incremental learning and adaptive update method according to claim 1, characterized in that: The model performance evaluation parameters include at least one of the following: root mean square prediction error, average prediction error, standard deviation of prediction error, and maximum prediction error; the performance change state of the artificial intelligence prediction model is determined based on the difference between the model performance evaluation parameters and the preset benchmark performance parameters.

7. The online incremental learning and adaptive update method according to claim 1, characterized in that: The determination of whether the artificial intelligence prediction model has drifted is also based on the analysis of the disturbance state data during the operation of the laser system, obtaining the spectral characteristics corresponding to the disturbance state data, and using the changes in the spectral characteristics to help determine whether the artificial intelligence prediction model has drifted.

8. The online incremental learning and adaptive update method according to claim 1, characterized in that: The online incremental learning uses the model parameters of the current artificial intelligence prediction model as initialization parameters, and adjusts the parameters of the artificial intelligence prediction model based on the cached new operating status data, so that the updated artificial intelligence prediction model can adapt to the current operating status of the laser pointing control system.

9. The online incremental learning and adaptive update method according to claim 1, characterized in that: The performance verification evaluates the predictive performance of the AI ​​prediction model before and after incremental learning based on the verification data, and determines whether the AI ​​prediction model after incremental learning meets the deployment conditions based on the prediction performance evaluation results; the AI ​​prediction model performs a model update when it meets the deployment conditions, and retains the AI ​​prediction model before incremental learning when it does not meet the deployment conditions.

10. An online incremental learning and adaptive update system, employing the online incremental learning and adaptive update method as described in any one of claims 1 to 9, characterized in that: The system includes a main control loop and an incremental learning subsystem connected to the main control loop; The main control loop includes: A multi-source sensing module is used to collect state data during the laser pointing control process; An artificial intelligence prediction model module is connected to the multi-source sensing module and is used to output a laser pointing change prediction result based on the state data. A laser pointing adjustment module, connected to the artificial intelligence prediction model module, is used to adjust the laser propagation direction according to the prediction result; The laser system output module is used to output the adjusted laser pointing state; The incremental learning subsystem includes: A data caching module, connected to the artificial intelligence prediction model module and the laser system output module, is used to cache prediction data, real data, and error data between prediction data and real data. The performance monitoring module is used to calculate the performance evaluation parameters of the artificial intelligence prediction model based on the error data; The drift detection module is used to determine whether the artificial intelligence prediction model has drifted based on the performance evaluation parameters, and to output a model update trigger signal. The incremental learning module is used to call cached data according to the model update trigger signal and perform online incremental learning on the artificial intelligence prediction model. The model management module is used to validate, deploy, and roll back the model version in case of anomalies in the incremental learning of the artificial intelligence prediction model.