Multi-stage pid parameter regulation method for laser frequency locking system based on deep learning

Abstract

The application relates to the technical field of laser system control, and discloses a multi-stage PID parameter regulation and control method for a laser frequency locking system based on deep learning, which comprises the following steps: acquiring an error signal time sequence of a PDH laser frequency locking system; outputting a frequency locking state score and fast-slow double-loop PID parameters through a deep learning model; judging whether the frequency locking state score and the fast-slow double-loop PID parameters both satisfy corresponding preset conditions; if the frequency locking state score and the fast-slow double-loop PID parameters do not both satisfy the corresponding preset conditions, iteratively optimizing the deep learning model until the frequency locking state score and the fast-slow double-loop PID parameters both satisfy the corresponding preset conditions; and outputting the fast-slow double-loop PID parameters of the optimized deep learning model. The application can realize online self-adaptive reasoning and closed-loop optimization of fast-slow double-loop multi-stage PID control parameters, and improve the intelligent level and the ability to adapt to complex working conditions of the PDH laser frequency locking system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser system control, and in particular to a multi-stage PID parameter regulation method for a laser frequency locking system based on deep learning. BACKGROUND

[0002] As a mainstream scheme for high-precision laser frequency stabilization control, the Pound-Drever-Hall (PDH) frequency locking technology is widely used in high-resolution spectroscopy, precision measurement, quantum physics and other fields. Traditional PDH frequency locking systems mostly use analog or digital proportional-integral-derivative (PID) controllers with fixed parameters, and the parameter setting mainly depends on experience, semi-empirical formula or manual trial and error.

[0003] However, the actual operating environment of the laser frequency locking system is complex, and is often affected by factors such as temperature drift, vibration, nonlinear effects and device aging, resulting in time-varying and nonlinear dynamic characteristics of the system. The PID control method with fixed parameters is difficult to adapt to the changes in the system state, and the locking performance and robustness have obvious bottlenecks. In related technologies, adaptive or intelligent algorithms (such as fuzzy control, genetic algorithm) are used for parameter control, but they have problems such as insufficient real-time performance and limited parameter adjustment range, which cannot meet the application requirements of high dynamic, high bandwidth and unattended laser frequency locking, and cannot adapt to the complex working conditions of the PDH laser frequency locking system. SUMMARY

[0004] Therefore, the present application provides a multi-stage PID parameter regulation method for a laser frequency locking system based on deep learning to solve the technical problem that the prior art cannot adapt to the complex working conditions of the PDH laser frequency locking system.

[0005] In a first aspect, the present application provides a multi-stage PID parameter regulation method for a laser frequency locking system based on deep learning, comprising:

[0006] obtaining an error signal time sequence of a PDH laser frequency locking system;

[0007] inputting the error signal time sequence into a trained deep learning model to obtain a frequency locking state score and fast-slow double-loop PID parameters output by the deep learning model;

[0008] determining whether the frequency locking state score and the fast-slow double-loop PID parameters both satisfy corresponding preset conditions;

[0009] If the frequency locking state score and the fast-slow double loop PID parameter do not both satisfy the corresponding preset condition, the current error signal time sequence is continuously acquired, and the deep learning model is iteratively optimized based on the current error signal time sequence until the frequency locking state score and the fast-slow double loop PID parameter both satisfy the corresponding preset condition.

[0010] The fast-slow double loop PID parameter output by the optimized deep learning model is output to the fast-slow double loop digital PID controller, so that the fast-slow double loop digital PID controller regulates and controls the PDH laser frequency locking system based on the input fast-slow double loop PID parameter.

[0011] The deep learning-based laser frequency locking system multi-level PID parameter regulation method of the application can infer the frequency locking state score and the fast-slow double loop PID parameter based on the error signal time sequence through the deep learning model, continuously acquire the current error signal time sequence if the frequency locking state score and the fast-slow double loop PID parameter do not both satisfy the corresponding preset condition, and iteratively optimize the deep learning model based on the current error signal time sequence until the frequency locking state score and the fast-slow double loop PID parameter both satisfy the corresponding preset condition, so that the fast-slow double loop digital PID controller regulates and controls the PDH laser frequency locking system based on the input fast-slow double loop PID parameter, realizes online adaptive inference and closed-loop optimization of the fast-slow double loop multi-level PID control parameter, and improves the intelligent level and the ability to adapt to complex working conditions of the PDH laser frequency locking system.

[0012] In some optional embodiments, the deep learning model adopts a GRU, an LSTM or a time sequence neural network structure fused with an attention mechanism.

[0013] In this way, the long-term dependence and important features in the error signal time sequence can be better captured, thereby improving the inference accuracy of the model.

[0014] In some optional embodiments, the deep learning model adopts a time sequence neural network architecture fused with an attention mechanism, and the time sequence neural network architecture includes a time sequence encoder, a frequency locking state score output end and a fast-slow double loop PID parameter output end.

[0015] The time sequence encoder is configured to extract features of the model input to obtain a feature vector, wherein, in the inference process of the deep learning model, the model input includes the error signal time sequence and a multi-channel auxiliary quantity, the multi-channel auxiliary quantity includes one or more combinations of a laser current, a temperature control signal, an environmental temperature and humidity, a vibration noise and a historical control quantity, and all the model inputs are input to the time sequence encoder after being aligned in time.

[0016] The frequency locking state score output end is configured to map the feature vector to the frequency locking state score.

[0017] The fast and slow double-loop PID parameter output end is configured to map the feature vector into fast and slow double-loop PID parameters.

[0018] In this way, after introducing the multi-channel auxiliary input, the model can adaptively adjust the parameter mapping relationship under different working points and different disturbance environments, thereby maintaining the optimal control performance under multi-source disturbances.

[0019] In some optional embodiments, the deep learning model is trained according to the following steps:

[0020] The training data set is divided into a training set, a test set and a validation set, and the deep learning model is supervised learning trained by using the training set, wherein the training data set includes training PDH error signal time series collected under multiple disturbances and different working conditions, each training PDH error signal time series is labeled with a target frequency locking state score and corresponding optimal fast and slow double-loop PID parameter group;

[0021] The deep learning model calculates the value of the joint loss function based on the input training set, and optimizes the model parameters of the deep learning model with the minimum value of the joint loss function as the target, wherein the joint loss function includes a frequency locking score regression loss, a PID parameter regression loss, a regularization loss of parameter change rate and a regularization loss of output smoothness constraint;

[0022] The inference performance of the deep learning model is verified based on the test set and the validation set.

[0023] In this way, by introducing a multi-dimensional joint loss function, the parameter change rate, the historical parameter offset and the control output smoothness are jointly constrained to prevent system oscillation or instability caused by parameter mutation.

[0024] In some optional embodiments, the deep learning model is iteratively optimized based on the current error signal time series, including:

[0025] Obtain the running data of the PDH laser frequency locking system, and determine the current running phase of the PDH laser frequency locking system based on the running data;

[0026] Adjust the weights of the frequency locking score regression loss, the PID parameter regression loss, the regularization loss of parameter change rate and the regularization loss of output smoothness constraint in the joint loss function based on the current running phase;

[0027] Iteratively optimize the deep learning model based on the joint loss function with the adjusted weights, the current error signal time series and the collected multi-channel auxiliary quantities.

[0028] In this mode, the current operation stage is determined according to the operation data of the PDH laser frequency locking system, and then the optimization weights of each constraint term are dynamically adjusted, so as to realize multi-objective adaptive optimization.

[0029] In some optional embodiments, the training data set further includes an extended data set, and the extended data set includes typical system noise and abnormal signals.

[0030] In this mode, the deep learning model can learn the discrimination rule under the condition of signal interference, waveform distortion or reduced signal-to-noise ratio, so as to maintain stable frequency locking determination and parameter adaptive output in actual extreme working conditions, and improve the long-term reliability and anti-disturbance ability of the overall system.

[0031] In some optional embodiments, after outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller, the method further includes:

[0032] Monitoring operation data of the PDH laser frequency locking system;

[0033] Feeding back the operation data to the deep learning model, and the deep learning model performs online fine-tuning or incremental learning based on the operation data.

[0034] In this mode, the deep learning model is retrained or fine-tuned by supplementing newly collected operation data, effectively adapting to long-term drift factors such as laser aging and cavity environment changes, improving model adaptability and system anti-disturbance ability, and realizing intelligent closed-loop maintenance of "self-learning and self-evolution".

[0035] In some optional embodiments, after outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller, the method further includes:

[0036] Monitoring system operation indexes of the PDH laser frequency locking system;

[0037] Judging whether the system operation indexes are abnormal, and when the system operation indexes are abnormal, outputting alarm information and / or obtaining fast and slow double-loop PID parameters by re-inference of the deep learning model.

[0038] In this mode, adaptive abnormal alarm and self-recovery mechanism are realized by monitoring system operation indexes, and when the system operation indexes are abnormal, fault alarm or re-inference of fast and slow double-loop PID parameters is automatically triggered, so as to ensure that the system can continuously and stably operate in complex environment and extreme working conditions.

[0039] In some optional embodiments, after outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller, the method further includes:

[0040] The fast and slow double-loop PID parameters output by the optimized deep learning model are segmented and gradually entered according to a preset maximum step size and a control amount slope limit value;

[0041] At the end of each evaluation window, if the current frequency locking state score is not lower than the frequency locking state score in the previous window or the score drop value does not exceed the set drop threshold, the gradual entering is continued, and if the score drop values of a continuous target number of windows exceed the set drop threshold or the fast and slow double-loop PID parameters output by the deep learning model reach the amplitude limiting threshold, the system is rolled back to the latest stable parameter group, and the fast and slow double-loop digital PID controller adjusts and controls the PDH laser frequency locking system based on the latest stable parameter group, wherein the stable parameter group is the fast and slow double-loop PID parameters meeting the preset condition.

[0042] In this way, by setting the amplitude limiting threshold, the preset maximum step size, the control amount slope limit value, and the gradual parameter writing, and rolling back to the latest stable parameter group when the score drops, not only is the negative impact of poor output parameters on the system avoided, but also the smoothness and safety of the parameter updating process are significantly improved, and the frequency locking instability, oscillation, or laser overdrive caused by parameter mutation is prevented. At the same time, this mechanism shortens the time for the system to recover from an abnormal state to stable locking, enhances the overall robustness and fault tolerance, and ensures that the system can still maintain high-precision and high-reliability frequency locking performance in a complex disturbance environment and during long-term operation.

[0043] In a second aspect, the application provides a multi-stage PID parameter adjustment and control system for a laser frequency locking system based on deep learning, which is applied to implement the multi-stage PID parameter adjustment and control method for a laser frequency locking system based on deep learning of the first aspect or any of the corresponding embodiments.

[0044] An error signal acquisition module is configured to acquire an error signal time sequence of the PDH laser frequency locking system.

[0045] An intelligent terminal is provided with a deep learning model, and the frequency locking state score and the fast and slow double-loop PID parameters are obtained by the deep learning model based on the error signal time sequence. If the frequency locking state score and the fast and slow double-loop PID parameters do not both meet the corresponding preset conditions, the error signal acquisition module is used to continuously acquire the current error signal time sequence, and the deep learning model is iteratively optimized based on the current error signal time sequence until the frequency locking state score and the fast and slow double-loop PID parameters both meet the corresponding preset conditions. The fast and slow double-loop PID parameters output by the optimized deep learning model are output to the fast and slow double-loop digital PID controller.

[0046] A fast and slow double-loop digital PID controller is configured to perform closed-loop adjustment on the laser and the TEC temperature control unit in the PDH laser frequency locking system according to the input fast and slow double-loop PID parameters. Attached Figure Description

[0047] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of the structure of a multi-level PID parameter control system for a laser frequency locking system based on deep learning, according to an embodiment of the present invention.

[0049] Figure 2 This is a flowchart of a multi-level PID parameter control method for a laser frequency locking system based on deep learning, according to an embodiment of the present invention.

[0050] Figure 3 This is a flowchart of another deep learning-based multi-level PID parameter control method for laser frequency locking systems according to an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] To address the challenges of real-time optimization of PID parameters in existing PDH laser frequency locking systems, which suffer from dynamic system changes, limited overall frequency locking robustness, and low automation, this paper proposes a deep learning-based method, device, and system for multi-level PID parameter control in laser frequency locking systems. By using a deep learning model to extract features and dynamically model the time series of error signals, online adaptive reasoning and closed-loop optimization of fast and slow dual-loop PID parameters are achieved, thereby improving the intelligence level and adaptability of PDH laser frequency locking systems to complex operating conditions.

[0053] like Figure 1 As shown in the figure, the deep learning-based multi-level PID parameter control system for laser frequency locking in this invention includes:

[0054] The error signal acquisition module is used to acquire the time series of error signals from the PDH laser frequency locking system.

[0055] The intelligent terminal is provided with a deep learning model, and a frequency locking state score and fast and slow double-loop PID parameters are obtained by reasoning based on an error signal time sequence through the deep learning model. If the frequency locking state score and the fast and slow double-loop PID parameters do not all meet the corresponding preset conditions, the current error signal time sequence is continuously acquired through an error signal acquisition module, and the deep learning model is iteratively optimized based on the current error signal time sequence until the frequency locking state score and the fast and slow double-loop PID parameters all meet the corresponding preset conditions. The fast and slow double-loop PID parameters output by the optimized deep learning model are output to the fast and slow double-loop digital PID controller.

[0056] The fast and slow double-loop digital PID controller is used for closed-loop adjustment of the laser and the TEC temperature control unit in the PDH laser frequency locking system according to the input fast and slow double-loop PID parameters.

[0057] Specifically, the PDH laser frequency locking system is an advanced technical system for accurately locking the frequency of a laser to the resonant frequency of an optical resonant cavity. The system realizes stable control of the laser frequency through the cooperative work of a series of precise optical and electrical modules. The PDH laser frequency locking system usually includes:

[0058] The laser is the laser light source of the PDH laser frequency locking system, which is responsible for outputting the laser with the frequency to be locked. The emission frequency is not fixed and will be significantly affected by current driving and temperature drift. Changes in the current will directly change the particle number inversion distribution of the gain medium inside the laser, thereby affecting the laser frequency; and temperature drift will cause changes in the physical size and refractive index of the laser cavity, causing the laser frequency to shift.

[0059] The acousto-optic modulator (AOM) moves the frequency of the incident laser through the acousto-optic effect to ensure that the light incident on the resonant cavity is in the appropriate locking interval. At the same time, the acousto-optic modulator can also suppress parasitic reflections, reduce unnecessary optical feedback to the system, and improve the stability and reliability of the system.

[0060] The electro-optic modulator (EOM) is used to realize phase modulation, which superimposes the modulation signal of the local oscillator (LO) on the laser.

[0061] The polarizing beam splitter (PBS) mainly separates reflected light and transmitted light.

[0062] A photodetector (PD) includes a front-end photodetector and a back-end photodetector. The front-end photodetector receives the cavity reflected light and converts it into an electrical signal, which contains information about the laser frequency and the cavity mode detuning. The back-end photodetector monitors the transmitted light and obtains feedback information about the transmitted light, which can be used for PDH laser frequency locking system state evaluation or additional diagnosis to help the operator fully understand the running status of the system.

[0063] A local oscillator source provides a reference radio frequency signal, which is mixed with the modulation signal of the front-end photodetector. Through mixing and subsequent filtering processing, a PDH error signal can be obtained. This error signal is an important basis for the system to judge whether the laser frequency is accurately locked to the cavity mode, and it directly represents the detuning of the laser frequency relative to the cavity mode. The error signal is collected by an error signal collection module to form a digitized error signal time sequence, and is input to a deep learning model to provide data support for the control and optimization of the system. In the embodiments of the present application, the error signal collection module uses a high-speed analog-to-digital converter to collect the error signal time sequence and input it to the high-speed analog-to-digital converter interface (AIN) of the deep learning model.

[0064] A thermoelectric cooler (TEC) temperature control unit is used to adjust the temperature of the laser or the resonant cavity. Since temperature changes will cause low-frequency drift of the laser frequency, precise control of the temperature by the TEC temperature control unit can compensate for low-frequency drift and improve the long-term stability of the system.

[0065] The deep learning-based laser frequency locking system multi-stage PID parameter regulation system includes an error signal collection module, a fast and slow double-loop digital PID controller, and an intelligent terminal equipped with a deep learning model.

[0066] The error signal collection module uses a high-speed analog-to-digital converter to collect the error signal time sequence and input it to the high-speed analog-to-digital converter interface of the deep learning model.

[0067] The intelligent terminal includes a computer, a server, and a single-chip microcomputer, and a deep learning model on the intelligent terminal obtains a frequency locking state score and fast-slow double-loop PID parameters through inference of an input error signal time sequence. If the frequency locking state score and the fast-slow double-loop PID parameters do not both satisfy corresponding preset conditions, the error signal acquisition module continuously acquires a current error signal time sequence, and the deep learning model is iteratively optimized based on the current error signal time sequence until the frequency locking state score and the fast-slow double-loop PID parameters both satisfy the corresponding preset conditions. The fast-slow double-loop PID parameters output by the optimized deep learning model are output to a fast-slow double-loop digital PID controller.

[0068] The fast-slow double-loop digital PID controller includes a fast-loop PID controller and a slow-loop PID controller. The fast-loop PID controller and the slow-loop PID controller respectively output a fast-loop control signal FAST_OUT and a slow-loop control signal SLOW_OUT. The fast-loop control signal FAST_OUT is output to a laser current driving unit, used for adjusting the laser current driving, capable of quickly suppressing high-frequency disturbance and transient noise, and ensuring that the laser frequency responds to the fast-changing interference in a short time. The slow-loop control signal SLOW_OUT is output to a TEC temperature control driving unit, used for controlling a TEC temperature control unit, realizing temperature drift and low-frequency drift compensation, eliminating the low-frequency drift of the laser frequency by slowly adjusting the temperature, and improving the long-term stability of the PDH laser frequency locking system.

[0069] The fast-slow double-loop digital PID controller supports parameter hot plugging and state limiting, and the fast-slow double loop is independently adjusted. The deep learning model dynamically iteratively outputs PID parameters at each level and corresponding boundary values according to real-time feedback of the system. The parameter updating process supports high-frequency online writing and does not cause interruption of the controller, ensuring that the control system always works in the optimal or suboptimal control interval.

[0070] The frequency locking state score is used to comprehensively reflect the laser frequency stability, system disturbance rejection performance, and closed-loop convergence characteristics.

[0071] It should be understood that in the present application, "multi-level" specifically refers to the parameter level rather than adding another control loop, and the fast-slow double-loop PID parameters output by the deep learning model include fast-loop proportional parameters , fast-loop integral parameters , fast-loop differential parameters , and slow-loop integral parameters .

[0072] The fast loop output end adopts P, I, and D parameters , and the slow loop adopts integral parameters . The sampling frequencies and update time scales of the two loops are different. The fast loop and the slow loop have a first sampling frequency and a second sampling frequency The slow loop only acts on low-frequency drift, and the first sampling frequency is much larger than the second sampling frequency; the fast loop mainly controls high-bandwidth disturbance. Range constraints, step / slope limits and gradual entry writing parameters are imposed on the above parameters during online hot updating, and the system is rolled back to the latest stable parameter group when the score decreases.

[0073] To avoid coupling, the error signal is divided to obtain a fast-changing error and a slow-changing error . Wherein, is extracted through a low-pass filter , representing the low-frequency component in the error signal, mainly reflecting the slow-changing trend such as temperature drift, and is used for slow loop integral adjustment; represents the high-frequency component in the error signal, and is used for fast loop to suppress transient disturbance. Wherein, is the time constant of the low-pass filter, which determines the boundary of fast / slow components; is an integral limiting parameter, which is used to prevent the integral element from being oversaturated.

[0074]

[0075] Wherein the slow loop is only based on the slow-changing error , and the fast loop is only based on the fast-changing error . The parameter vector output by the deep learning model is , and the optional extension is used for differential filtering and safety constraints against integral saturation.

[0076] The multi-stage PID parameter regulation system of the laser frequency locking system based on deep learning further comprises a state monitoring module for detecting the frequency locking state score and system operation index, and feeding back to the deep learning model in real time. The system monitoring module supports adaptive criteria based on multiple indexes, including the frequency locking state score, the error signal mean square error, and the fast and slow loop control output energy consumption. If any index is detected to be abnormal or the system stability decreases, the data acquisition and deep learning inference process are automatically re-iterated to ensure that the system can recover to the optimal or near-optimal locking state in time, and realize automatic abnormal detection and self-healing.

[0077] The embodiment of the application further provides a multi-stage PID parameter regulation method of a laser frequency locking system based on deep learning, which is applied to terminal equipment such as computers, servers and single-chip microcomputers, and runs a deep learning model through the terminal equipment. As shown in Figure 2 , the process comprises the following steps:

[0078] Step S101, acquiring the error signal time sequence of the PDH laser frequency locking system.

[0079] Specifically, the error signal time sequence of the PDH laser frequency locking system is periodically collected by the error signal collection module in the multi-stage PID parameter regulation system based on the deep learning laser frequency locking system, and then input to the high-speed analog-to-digital converter interface of the deep learning model.

[0080] In step S102, the error signal time sequence is input to the trained deep learning model to obtain the frequency locking state score and the fast and slow double-loop PID parameters output by the deep learning model.

[0081] The deep learning model adopts a GRU, LSTM or attention mechanism-fused time sequence neural network structure.

[0082] Specifically, the deep learning model can use a transfer learning method, and by means of a pre-trained model in a related signal processing field, such as a GRU or LSTM basic network structure for time sequence prediction, the model is secondarily trained and fine-tuned to improve the adaptability of the model to the features of the error signal time sequence and the parameter inference accuracy. In the transfer learning process, the parameters of the bottom layer time sequence feature extraction layer are kept, and the training and optimization are mainly performed on the high layer control parameter mapping layer, so as to shorten the model training period and improve the robustness of the model under limited samples and new working conditions.

[0083] In actual engineering deployment, the deep learning model can be implemented through a heterogeneous computing platform, such as deploying the model inference task on a high-performance GPU, FPGA or AI accelerator to ensure millisecond-level decision-making efficiency. Through parallel floating-point calculation and high-bandwidth data interface, real-time processing of high-dimensional error signals and high-frequency parameter output are realized to meet the actual needs of high-speed closed-loop control. At the same time, the model can also be flexibly trimmed according to the software and hardware conditions of the system platform, supporting multi-level deployment from embedded end to high-performance server.

[0084] In a more preferred embodiment, the deep learning model adopts a time sequence neural network architecture fused with an attention mechanism, which can automatically focus on the key time window of error signal changes to realize efficient feature extraction and parameter mapping.

[0085] Specifically, the time sequence neural network architecture includes a time sequence encoder, a frequency locking state score output end and a fast and slow double-loop PID parameter output end.

[0086] The time sequence encoder is used to extract the features of the model input to obtain a feature vector. In the inference process of the deep learning model, the model input includes the error signal time sequence and multi-channel auxiliary quantities, and the multi-channel auxiliary quantities include one or more combinations of the laser current, temperature control signal, environmental temperature and humidity, vibration noise and historical control quantities. All model inputs are time-aligned and then input to the time sequence encoder.

[0087] A frequency-locked state score output end for mapping the feature vector to a frequency-locked state score;

[0088] A fast-slow dual-loop PID parameter output end for mapping the feature vector to fast-slow dual-loop PID parameters.

[0089] After the introduction of the multi-channel auxiliary input in the embodiment of the present application, the model can adaptively adjust the parameter mapping relationship under different working points and different disturbance environments, thereby maintaining the best control performance under multi-source disturbances.

[0090] By way of example, the deep learning model adopts a bidirectional long short-term memory network (BiLSTM-Attention) architecture with a fusion attention mechanism, and is provided with a double-output head, including a frequency-locked state score output end and a fast-slow dual-loop PID parameter output end. The input of the model is a PDH error signal time sequence of a predetermined length, and the joint input includes multi-channel auxiliary quantities such as laser current, temperature control signal, environmental temperature and humidity, vibration noise, and historical control quantity. All inputs are input to the time sequence encoder after being aligned in time. The time sequence encoder is composed of a bidirectional LSTM layer and a self-attention layer, which can automatically extract the features of the error signal in the key time window and form a context representation.

[0091] At the output end of the model, the frequency-locked state score output end adopts a multi-layer perceptron structure to map the encoded context features to a frequency-locked state score. The score value is used to comprehensively reflect the system frequency stability, closed-loop convergence characteristics, and disturbance rejection performance. The frequency-locked state score includes but is not limited to the root mean square value of the system error signal, the frequency stability, the output amplitude of the fast-slow loop, and the laser-cavity mode matching degree, etc.

[0092] The fast-slow dual-loop PID parameter output end also adopts a multi-layer perceptron structure to map the context features to the proportional, integral, and derivative gains of the fast and slow loops, as well as auxiliary hyperparameters (such as integral saturation amplitude limiting, differential filter time constant, etc.). To ensure that the parameters are within the physical safe range, the parameter head output is mapped to the preset parameter interval through a bounded activation function, thereby avoiding parameter over-limiting or causing system instability.

[0093] After the introduction of the multi-channel auxiliary input, the deep learning model can adaptively adjust the parameter mapping relationship under different working points and different disturbance environments. For example, in a temperature rapid change scenario, the model can automatically reduce the fast loop gain and increase the differential filter time constant; in a vibration enhanced scenario, the model can increase the fast loop bandwidth to improve the high frequency suppression capability. Therefore, the model adjusts the input splicing and conditional modulation mechanism in structure to use the auxiliary input to adjust the dynamic range of the time sequence encoder weight and the output mapping layer, thereby maintaining the best control performance under multi-source disturbances.

[0094] Step S103, judge whether the frequency locking state score and the fast and slow double loop PID parameters meet the corresponding preset conditions.

[0095] Specifically, the frequency locking state score and the fast and slow double loop PID parameters are both provided with corresponding preset parameter ranges. When the frequency locking state score or the fast and slow double loop PID parameters are within the corresponding preset parameter ranges, it is considered that the frequency locking state score or the fast and slow double loop PID parameters meet the corresponding preset conditions.

[0096] The preset parameter ranges include two aspects:

[0097] 1. The threshold range of the frequency locking state score, which is used to determine whether the system enters the stable locking interval, for example, the upper limit or lower limit based on the error signal mean square deviation, frequency deviation, convergence time and other indicators;

[0098] 2. The allowed range of the fast and slow double loop PID parameters, including the fast loop proportional gain, integral gain, derivative gain and slow loop integral gain, and the maximum / minimum values of the super parameters (such as integral limit, differential filter time constant) thereof, which can be set by experimental identification, theoretical analysis or adaptive algorithm.

[0099] It should be understood that the preset parameter ranges can be set based on historical experiments, theoretical analysis or dynamic adaptive algorithm, and can be flexibly configured for different application scenarios.

[0100] When the frequency locking state score is outside the preset range or the inferred PID parameters exceed the preset range, the system automatically enters the parameter optimization process.

[0101] Step S104, if the frequency locking state score and the fast and slow double loop PID parameters do not meet the corresponding preset conditions, the current error signal time sequence is continuously obtained, and the deep learning model is iteratively optimized based on the current error signal time sequence until the frequency locking state score and the fast and slow double loop PID parameters meet the corresponding preset conditions.

[0102] Specifically, if the frequency locking state score and the fast and slow double loop PID parameters both meet the corresponding preset conditions, it indicates that the inferred parameters meet the set standard, and the PDH laser frequency locking system determines that the optimal or quasi-optimal locking state has been reached, the current parameters are maintained, continuous monitoring is performed, and the system enters the steady state running phase.

[0103] If the frequency locking state score or the fast and slow double loop PID parameters do not meet the corresponding preset conditions, it indicates that the system has phenomena such as frequency instability and control parameter deviation. At this time, the deep learning model enters the parameter adaptive optimization process. Based on the deep learning model and the real-time observed error signal time sequence, the fast and slow double loop PID parameters are optimized and iterated online to realize the relocking and performance recovery of the system frequency closed loop.

[0104] The latest error signal time series and multi-channel auxiliary quantities collected are input into the deep learning model again. The deep learning model outputs new fast and slow double-loop PID parameters, including proportional gain, integral gain, derivative gain, and adaptive hyperparameters such as integral saturation boundary and derivative filter time constant, in real time according to the latest dynamic characteristics of the model input.

[0105] Through continuous iteration and optimization, the frequency locking state score and the fast and slow double-loop PID parameters meet the corresponding preset conditions.

[0106] In step S105, the fast and slow double-loop PID parameters output by the optimized deep learning model are output to the fast and slow double-loop digital PID controller, so that the fast and slow double-loop digital PID controller regulates and controls the PDH laser frequency locking system based on the input fast and slow double-loop PID parameters.

[0107] Specifically, the fast and slow double-loop digital PID controller includes a fast-loop PID controller and a slow-loop PID controller, which independently regulate and control different frequency response regions of the system.

[0108] The fast-loop PID controller is mainly responsible for suppressing high-frequency disturbances and laser intrinsic noise, and its control output directly acts on the laser current driving end. The fast-loop channel has dynamic integral coefficient, integral element with amplitude limiting protection, and differential element with first-order low-pass filter under the dynamic adjustment of the deep learning model inference parameters, ensuring high bandwidth and response speed, and effectively preventing overshoot, jitter and integral saturation phenomenon.

[0109] The slow-loop PID controller is mainly used to compensate for cavity temperature drift and long-term low-frequency drift of the system, and its control output acts on the TEC temperature control unit. The slow-loop channel adopts input low-pass filtering and amplitude limiting protection measures, so that it only integrates low-frequency disturbances, avoiding false response caused by high-frequency signal interference. The slow-loop integral time constant is consistent with the physical response process of the laser and the cavity, ensuring that the slow drift error is compensated for a long time and thoroughly, and the system realizes zero steady-state error.

[0110] Further, the output signals of the fast and slow double-loop digital PID controller are processed through engineering limiting and digital smoothing, ensuring that the control signal amplitude is within the physical hardware safety range, reducing the influence of high-frequency components on downstream drivers, and preventing hardware distortion or over-limiting.

[0111] The fast ring PID controller and the slow ring PID controller respectively output a fast ring control signal FAST_OUT and a slow ring control signal SLOW_OUT, the fast ring control signal FAST_OUT is output to a laser current driving unit, is used for adjusting laser current driving, can quickly suppress high-frequency disturbance and transient noise, and ensures that the laser frequency responds to fast changing interference in a short time. The slow ring control signal SLOW_OUT is output to a TEC temperature control driving unit, so as to control the TEC temperature control unit, realize temperature drift and low-frequency drift compensation, eliminate the low-frequency drift of the laser frequency by slowly adjusting the temperature, and improve the long-term stability of the PDH laser frequency locking system.

[0112] The multi-stage PID parameter regulation method of the laser frequency locking system based on deep learning in the embodiment of the application can obtain the frequency locking state score and the fast and slow double-loop PID parameters based on the error signal time sequence through the deep learning model. If the frequency locking state score and the fast and slow double-loop PID parameters do not all meet the corresponding preset conditions, the current error signal time sequence is continuously obtained, and the deep learning model is iteratively optimized based on the current error signal time sequence until the frequency locking state score and the fast and slow double-loop PID parameters all meet the corresponding preset conditions. The fast and slow double-loop digital PID controller regulates the PDH laser frequency locking system based on the input fast and slow double-loop PID parameters, realizes online adaptive inference and closed-loop optimization of the fast and slow double-loop multi-stage PID control parameters, and improves the intelligent level and the ability to adapt to complex working conditions of the PDH laser frequency locking system.

[0113] The embodiment of the application significantly improves the response speed, locking precision and engineering automation level of the PDH laser frequency locking system, and is suitable for high-precision laser frequency stabilization and optical measurement fields.

[0114] In some embodiments, the deep learning model is trained according to the following steps:

[0115] In step S201, the training data set is divided into a training set, a test set and a validation set, and the deep learning model is supervised learning trained by using the training set. The training data set includes training PDH error signal time sequences collected under various disturbances and different working conditions, each training PDH error signal time sequence is labeled with a target frequency locking state score and corresponding optimal fast and slow double-loop PID parameter group.

[0116] In step S202, the deep learning model calculates the value of the joint loss function based on the input training set, and optimizes the model parameters of the deep learning model with the minimum value of the joint loss function as the target. The joint loss function includes a frequency locking score regression loss, a PID parameter regression loss, a regularization loss of the parameter change rate and a regularization loss of the output smoothness constraint.

[0117] Step S203, the inference performance of the deep learning model is verified based on the test set and the verification set.

[0118] A plurality of training data sets of PDH error signal time series for training containing different physical disturbances and system working conditions are obtained, all of which are provided with artificially labeled frequency-locked state scores and corresponding optimal fast and slow double-loop PID parameter sets. The training data sets are divided into a training set, a test set and a verification set, the training set is used for supervised learning training of the deep learning model, and the test set and the verification set are used for monitoring the fitting degree and generalization ability of the model, and finally a trained model is obtained.

[0119] Preferably, the training data set covers PDH error signal time series under various typical physical disturbances (such as mechanical vibration, temperature drift, laser aging, etc.) and system working conditions. The training data set can be obtained by combining experimental acquisition, theoretical simulation, numerical disturbance enhancement, etc., and the label system is improved by artificial labeling, expert scoring and automatic generation, so as to ensure that the model has sufficient generalization and migration ability.

[0120] Further, the training data set further includes an extended data set, and the extended data set includes typical system noise and abnormal signals.

[0121] Specifically, the extended data set includes the following types:

[0122] Electronic noise data: superimposing white noise and 1 / f noise generated by a front-end photodetector, an amplifier and a high-speed analog-to-digital converter on the error signal time series acquisition process. This type of noise is generated by experimental measurement and simulation superposition, and is used to train the model to correctly evaluate the frequency-locked state and output reasonable PID parameters in an environment with low electronic signal-to-noise ratio.

[0123] Temperature drift data: applying a step or ramp temperature disturbance to the laser or cavity temperature controller to obtain a slowly varying and large amplitude error signal sequence. This type of data is used to train the model to adaptively adjust the slow loop PID integral time constant under long-term temperature drift conditions to ensure that the system has no steady-state error.

[0124] Optical abnormal signal: including non-typical error signal waveforms caused by mode jump, sudden drop in optical coupling efficiency or reflected light interference. This type of signal is obtained by deliberately mismatching or rapidly scanning the cavity length in the experiment, and is used to train the model to timely identify the frequency-locked failure and trigger parameter reset under abnormal conditions.

[0125] By introducing the extended data set containing typical system noise and abnormal signals in the training process, the discrimination ability and robustness of the deep learning model under extreme working conditions are improved.

[0126] In the training strategy, the data set containing noise and abnormal signal is mixed with the normal working condition data set according to a certain proportion and input into the deep learning model, and the frequency locking state score and PID parameter output are optimized simultaneously through supervised learning. In order to avoid overfitting of the model to abnormal data, a weight adjustment mechanism is introduced in the loss function, so that the model aims to optimize accuracy when processing normal data, and aims to robustness and correct discrimination when processing noise / abnormal data. In this way, the model can learn the discrimination rule under the condition of signal interference, waveform distortion or low signal-to-noise ratio, so as to maintain stable frequency locking determination and parameter adaptive output in actual extreme working conditions, and improve the long-term reliability and anti-disturbance ability of the overall system.

[0127] In this way, the deep learning model can learn the discrimination rule under the condition of signal interference, waveform distortion or low signal-to-noise ratio, so as to maintain stable frequency locking determination and parameter adaptive output in actual extreme working conditions, and improve the long-term reliability and anti-disturbance ability of the overall system.

[0128] The training process adopts a joint loss function, including frequency locking score regression loss, PID parameter regression loss, parameter change rate regularization loss and output smoothness constraint regularization loss, to avoid sudden changes or high-frequency oscillations in the parameter adjustment process. Through the division of training set, validation set and test set, the accuracy and generalization ability of the model under different working conditions are ensured.

[0129] In actual operation, the deep learning model receives the latest PDH error signal time series and multi-channel auxiliary quantities in real time in a sliding time window, and outputs the current frequency locking state score and fast and slow double-loop PID parameters.

[0130] As an additional input feature, if the frequency locking state score or the fast and slow double-loop PID parameters both meet the corresponding preset conditions, i.e. the frequency locking state score is higher than the preset threshold and the output PID parameters fall within the set range, the PDH laser frequency locking system maintains the current fast and slow double-loop PID parameters and enters a stable running state. If the frequency locking state score or the fast and slow double-loop PID parameters do not both meet the corresponding preset conditions, i.e. the frequency locking state score is lower than the threshold or the fast and slow double-loop PID parameters deviate from the preset range, the parameter hot update mechanism is triggered, the deep learning model re-reasons, and a new set of PID parameters are written into the register of the digital PID controller through the high-speed bus online, without interrupting the closed-loop control. The fast loop control signal is used to drive the laser current to suppress high-frequency disturbances and transient noise; the slow loop control signal acts on the TEC temperature control unit for long-term drift compensation.

[0131] In this way, by introducing a multi-dimensional joint loss function, the parameter change rate, historical parameter deviation and control output smoothness are jointly constrained to prevent system oscillation or instability caused by parameter mutation.

[0132] Through the improved deep learning model, even in the presence of environmental disturbances, device aging and other complex situations, the frequency locking state can be accurately evaluated and the optimal PID parameter can be inferred, so that high-precision and high-robustness adaptive frequency locking is realized.

[0133] In addition, for different engineering application scenarios, the model hyperparameters, input features, criterion thresholds, etc. can be flexibly adjusted. For example, in the case of ultra-high stability requirement, the frequency locking score threshold is appropriately increased or the high-frequency disturbance discrimination ability is strengthened; in the case of large bandwidth application, the timing window length and input dimension are optimized to obtain better comprehensive performance.

[0134] In some embodiments, in step S104, the deep learning model is iteratively optimized based on the current error signal time series, comprising:

[0135] In step S1041, the running data of the PDH laser frequency locking system is obtained, and the current running stage of the PDH laser frequency locking system is determined based on the running data;

[0136] In step S1042, the weights of the frequency locking score regression loss, the PID parameter regression loss, the regularization loss of the parameter change rate and the regularization loss of the output smoothness constraint in the joint loss function are adjusted based on the current running stage;

[0137] In step S1043, the deep learning model is iteratively optimized based on the joint loss function with the adjusted weights, the current error signal time series and the collected multi-channel auxiliary quantities.

[0138] Specifically, to ensure the stability and safety of parameter optimization, the present application introduces a multi-dimensional auxiliary loss function in the inference process of the deep learning model, which jointly constrains the parameter change rate, the historical parameter offset and the control output smoothness to prevent system oscillation or instability caused by parameter mutation. The joint loss function is:

[0139]

[0140] In the formula, is the main loss function, representing the frequency locking score regression loss, is the three auxiliary loss functions, respectively representing the PID parameter regression loss, the regularization loss of the parameter change rate and the regularization loss of the output smoothness constraint, representing the corresponding weight coefficients. Through the PID parameter regression loss, it is ensured that the new parameter will not deviate greatly from the historical stable interval; through the regularization loss of the parameter change rate, the amplitude of adjacent two parameter updates is limited to avoid rapid parameter tuning; through the regularization loss of the output smoothness constraint, the variance of the controller output is reduced to prevent the driving signal from appearing sharp fluctuations.

[0141] The deep learning model can also determine the current operation stage according to the operation data of the PDH laser frequency locking system, dynamically adjust the optimization weight of each constraint term, and realize multi-objective adaptive optimization. The operation stages of the PDH laser frequency locking system include a lock building stage (Fast Convergence), a steady-state stage (Steady-State Phase) and a disturbance recovery stage (Disturbance Rejection), each operation stage corresponds to a numerical interval, and the current operation stage is obtained according to the interval in which the value of the operation data is located.

[0142] In the lock building stage, the main loss function weight dominates, and the weight is increased to ensure fast convergence; in the steady-state stage, the weights of the regularization loss of the parameter change rate and the regularization loss of the output smoothness constraint are increased to ensure that the parameter update is gentle and the system is stable in the long term; in the disturbance recovery stage, the proportion of the main loss function and the PID parameter regression loss is appropriately increased to ensure that the system quickly returns to a stable state under external disturbance.

[0143] To ensure the effectiveness and safety of parameter optimization, the deep learning model can introduce an auxiliary loss function in the parameter inference process, jointly constrain the parameter change rate, historical parameter trend and control output smoothness, avoid parameter drastic changes and high-frequency oscillation of the control loop, ensure the stability and safety of long-term operation of the system, and prevent system shock or instability caused by parameter mutation. At the same time, the model can dynamically adjust the optimization weight according to the system historical feedback and the information output by the monitoring module, and realize multi-objective adaptive optimization.

[0144] The inferred new set of fast and slow double-loop PID parameters are written into the parameter register of the fast and slow double-loop digital PID controller in real time and online through a high-speed interface or bus. Preferably, the system supports a parameter hot update mechanism, that is, the fast and slow double-loop digital PID controller can seamlessly receive new parameters and take effect immediately without stopping the system, losing the loop, and interrupting the condition, ensuring the continuity and real-time of the frequency locking process. The output signals of the fast and slow double-loop digital PID controller are subjected to amplitude limiting, digital smoothing or filtering and other engineering processing, and then act on the laser current modulation channel and the TEC temperature control channel respectively. The change trend of the output signal and the physical response result of the system are collected and fed back to the deep learning model in real time and synchronously, for new round of parameter optimization and model fine-tuning.

[0145] In actual engineering deployment, the system can set safety thresholds such as maximum adjustment step number, maximum allowed change range or lock retry number of the parameter adaptive process, to prevent dead loop or out-of-control situation in extreme working conditions or fault state. All adaptive adjustment processes and results will be automatically recorded by the state monitoring module for subsequent analysis and operation and maintenance.

[0146] Preferably, in the adaptive adjustment process, the system can also switch the fast and slow double-loop PID parameter adaptive optimization strategy according to different working scenes. For example, in a high-noise environment, the fast loop bandwidth is preferentially optimized to strengthen the high-frequency disturbance suppression capability; in long-term steady-state application, the slow loop integral limit is strengthened to improve the robustness of temperature drift compensation. All strategies and thresholds of the system parameter adaptive optimization can be flexibly configured according to actual needs to meet the needs of different engineering scenes.

[0147] If the frequency locking state score and the fast and slow double-loop PID parameters reach or exceed the preset conditions after adaptive parameter adjustment, the system automatically enters the next stable control stage.

[0148] If the adaptive optimization still does not reach the target, the system can choose to continue the loop adjustment, or manually intervene and fault alarm according to the system strategy.

[0149] The embodiment of the application realizes multi-objective adaptive optimization by determining the current running stage according to the running data of the PDH laser frequency locking system, and then dynamically adjusting the optimization weight of each constraint term.

[0150] In some optional embodiments, after outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller, the following steps are further included:

[0151] Step S1061, monitoring the system running index of the PDH laser frequency locking system;

[0152] Step S1062, judging whether the system running index is abnormal, and outputting an alarm information and / or obtaining the fast and slow double-loop PID parameters by re-reasoning through the deep learning model when the system running index is abnormal.

[0153] Specifically, in the control link execution process, the state monitoring module monitors the system running index such as the frequency locking state score, the error signal dynamic, the control output change, etc. in real time. If it is detected that the frequency locking score is stably improved, the error signal amplitude is significantly reduced, and the system has no oscillation or overshoot, it is automatically determined that the frequency locking process enters the stable stage, the current closed-loop control parameter and the system state are recorded, and the closed-loop monitoring and adaptive maintenance stage is entered, realizing the long-term stability guarantee of the laser frequency locking.

[0154] The closed-loop control parameter can specifically include:

[0155] Fast loop PID parameters: proportional coefficient Kp_fast, integral coefficient Ki_fast, differential coefficient Kd_fast, and integral upper limit value, integral saturation prevention threshold, differential filter time constant, etc.

[0156] Slow loop PID parameters: integral coefficient Ki_slow, and corresponding integral limit, output limit threshold, etc.

[0157] Control output state: steady-state amplitude of fast loop current modulation signal, steady-state value of slow loop TEC temperature control driving signal, and their output change rate and average energy consumption level.

[0158] Parameter boundary condition: including the current adopted parameter upper and lower limit range, maximum allowed update step, parameter change rate constraint, and other safety constraint conditions.

[0159] System criterion threshold: frequency lock state score threshold for determining stable state, upper limit of error signal mean square deviation, lower limit of convergence time, and other set parameters.

[0160] System state includes key operating indicators, which can specifically include: error signal root mean square value (RMS), spectral characteristics and instantaneous deviation; frequency lock state score (reflecting the overall locking performance of the system); closed-loop convergence time and steady-state holding time; frequency stability index (such as Allan variance, phase noise level); controller output signal amplitude, energy consumption level, and fluctuation range; temperature control state of the laser and the cavity (set temperature, actual temperature, drift rate); environmental monitoring data (temperature and humidity, vibration intensity, power supply voltage, etc.); abnormal event record (number of lockouts, mode jump, output power anomaly, etc.).

[0161] The system state monitoring module continuously detects and analyzes the frequency lock state score, error signal fluctuation, output control signal change, and environmental parameters. Fixed thresholds can be set, or thresholds can be dynamically adjusted through adaptive algorithms to achieve sensitive discrimination under different working conditions. If the monitoring results show that the frequency lock state score is continuously higher than the preset threshold, the system error signal is in a low fluctuation and zero drift steady state interval, and the fast and slow double loop PID parameters are stable and the output signal is normal, it is determined that the system enters a long-term stable working state, and enters the maintenance monitoring phase. At this time, the system records the current closed-loop control parameters and key operating indicators, and can upload related data to the background or data center to realize health management and traceability analysis throughout the life cycle.

[0162] Preferably, the system supports a closed-loop adaptive fine-tuning mechanism. During the frequency locking process, if external disturbances, hardware drifts, or laser mode mutations are monitored, the fine-tuning process can be automatically triggered, and the PID parameters can be fine-tuned by the deep learning model according to real-time feedback to achieve full-process self-healing, self-adaptation, and long-term high-stability locking. The fine-tuning process can adapt to the long-term aging of laser and resonant cavity hardware components and changes in the environment, and through online learning or training with a small amount of incremental samples, the intelligent adaptive level of the control system is further improved, and high-reliability laser frequency locking is realized throughout the life cycle.

[0163] If abnormality or performance degradation is monitored in the closed-loop control process, the parameter collection and adaptive optimization process is restarted to ensure the intelligent closed loop and stable operation of the system.

[0164] In this way, adaptive abnormality alarm and self-healing mechanism are realized by monitoring system operation indicators. When the system operation indicators are abnormal, fault alarm or re-reasoning of fast and slow double-loop PID parameters is automatically triggered to ensure that the system can continuously and stably operate in complex environments and extreme working conditions.

[0165] In some embodiments, after outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller in step S105, the method further includes:

[0166] Step S1071, monitoring operation data of the PDH laser frequency locking system;

[0167] Step S1072, feeding back the operation data to the deep learning model, and the deep learning model performs online fine-tuning or incremental learning based on the operation data.

[0168] Specifically, during long-term operation of the system, the state monitoring module can also periodically collect operation data and feed it back to the deep learning module to trigger online fine-tuning or incremental learning of the model.

[0169] In an example, the operation data specifically includes:

[0170] Error signal time series: real-time waveform and statistical indicators (such as RMS, mean, bandwidth) of the frequency locking error signal, used to reflect the current frequency locking accuracy and dynamic characteristics;

[0171] Controller output data: including fast / slow loop PID output, laser current driving signal, TEC temperature control amount, etc., used to depict the system response state;

[0172] Environmental and working condition data: for example, temperature drift, cavity vibration, device aging, laser power attenuation, and other typical external interference factors;

[0173] Historical feedback data: frequency locking state score, historical optimal parameter distribution, control output smoothness, etc., used to provide context reference for the model.

[0174] In the application process, after standardization, denoising, and feature alignment, etc. preprocessing, these operation data are input into the deep learning model together with the typical samples collected historically, to realize the following two types of adaptive update methods:

[0175] Fine-tuning: When the running environment slightly drifts (e.g., small temperature fluctuations, slow device aging), only a small number of parameters of the model are updated to quickly adapt to the current state and avoid system performance degradation.

[0176] Incremental Learning: When the system encounters new operating conditions or typical disturbance scenarios (e.g., strong vibration, large temperature drift), the newly collected operating data is added to the training set as additional samples, and the model is retrained or structure fine-tuned to expand the model's adaptability to new scenarios.

[0177] Through the above mechanisms, the model can continuously absorb new data and gradually adapt to long-term factors such as laser aging and cavity environment drift, achieving a closed-loop maintenance capability of "self-learning and self-evolution" to ensure that the system can still output optimal PID parameters in complex environments.

[0178] In some embodiments, after outputting the fast and slow dual-loop PID parameters output by the optimized deep learning model to the fast and slow dual-loop digital PID controller in step S105, the method further includes:

[0179] segmenting and gradually entering the fast and slow dual-loop PID parameters output by the optimized deep learning model according to a preset maximum step size and applying a control amount slope limit value;

[0180] At the end of each evaluation window, if the current frequency locking state score is not lower than the frequency locking state score in the previous window or the score drop value does not exceed the set drop threshold, the gradual entry continues. If the score drop value of the continuous target number of windows exceeds the set drop threshold or the fast and slow dual-loop PID parameters output by the deep learning model reach the amplitude limiting threshold, the method reverts to the latest stable parameter set. The fast and slow dual-loop digital PID controller adjusts and controls the PDH laser frequency locking system based on the latest stable parameter set, wherein the stable parameter set is the fast and slow dual-loop PID parameters that meet the preset conditions.

[0181] Specifically, the fast and slow dual-loop PID parameters output by the optimized deep learning model are used as candidate parameters, and a heuristic gradual entry method is adopted:

[0182] the difference between the candidate parameters and the current parameters is written in segments according to a preset maximum step size and a control amount slope limit value is applied to the control amount . After completing a segment, the frequency locking state score and the amplitude limiting hit rate are calculated in the window . When and are met, the next segment is entered; otherwise, the method immediately reverts to the latest stable parameter set.

[0183] In this way, by setting a limiting threshold, preset maximum step size and control quantity slope limit value for the fast and slow double-loop PID parameters, and gradually writing the parameters, and rolling back to the latest stable parameter group when the score decreases, not only is the negative impact of poor output parameters on the system avoided, but also the smoothness and safety of the parameter updating process are significantly improved, preventing frequency lock instability, oscillation or laser overdrive caused by parameter mutation. At the same time, this mechanism shortens the time for the system to recover from an abnormal state to stable locking, enhances the overall robustness and fault tolerance, and ensures that the system can still maintain high precision and high reliability in frequency locking performance in a complex disturbance environment and during long-term operation.

[0184] In some embodiments, as shown in Figure 3 The multi-stage PID parameter regulation method of the deep learning-based laser frequency locking system according to the embodiments of the present application includes the following steps:

[0185] S1, an error signal time sequence of a PDH laser frequency locking system is obtained, and the error signal time sequence is input into a deep learning model. The deep learning model is used to automatically evaluate the system frequency locking state score according to the error signal time sequence, and to infer and output the frequency locking state score and the fast and slow double-loop PID parameters. If the frequency locking state score is lower than a preset threshold, S2 is performed; if the frequency locking state score is higher than the preset threshold, the system maintains the existing parameters and enters a stable locking state.

[0186] S2, the deep learning model performs self-adaptive iterative optimization on the fast and slow double-loop PID parameters, and automatically writes the new parameter set into the fast and slow double-loop digital PID controller. Based on the latest error signal time sequence input, the deep learning model performs multi-objective optimization on the fast and slow double-loop PID parameters through time sequence feature extraction and system dynamic identification, and outputs parameters including proportional, integral and derivative gains, as well as adaptive hyperparameters (such as integral saturation boundary and differential filter time constant). The PID parameter writing adopts an online hot update mechanism and does not need to interrupt the system control.

[0187] S3, the fast and slow double-loop digital PID controller adjusts the laser current and TEC temperature control signal in real time according to the latest parameters, forms a closed-loop control, and continuously collects new error signal feedback from the system monitoring module to the deep learning model. The fast loop mainly suppresses high-frequency disturbances, and the slow loop compensates for low-frequency drift. All output signals are subjected to amplitude limiting and digital smoothing before being applied to the execution unit. The system forms a real-time self-adaptive closed-loop process of "collection - inference - parameter adjustment - control - feedback".

[0188] S4, S1 to S3 are repeatedly performed until the frequency locking state score is higher than the preset threshold and the system remains stable for a long time. If the system is disturbed or the parameters are abnormal during the closed-loop process, it automatically returns to S1, the deep learning model reiteratively optimizes the parameters, and continuously adjusts adaptively until the score is optimal and the system is stable.

[0189] In addition, the system can set an adaptive abnormality alarm and a self-recovery mechanism. When detecting abnormal conditions such as external strong disturbance, sudden change of hardware state, and sharp decline of frequency locking score, the system automatically triggers a fault alarm or re-collects data and enters a parameter adaptive optimization process, thereby ensuring that the system can continuously and stably operate in complex environments and extreme working conditions.

[0190] Through the above closed-loop monitoring, adaptive maintenance, and intelligent decision-making in the whole process, the application can significantly improve the long-term operation stability, robustness, and engineering usability of the PDH laser frequency locking system, and meet the harsh demands of modern high-precision time-frequency systems and optical measurement fields.

[0191] Although the embodiments of the application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the application, and such modifications and changes fall within the scope of protection.

Claims

1. A multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system, characterized in that, The method comprises the following steps: acquiring an error signal time sequence of a PDH laser frequency locking system; inputting the error signal time sequence into a trained deep learning model to obtain a frequency locking state score and a fast-slow double loop PID parameter output by the deep learning model; determining whether the frequency locking state score and the fast-slow double loop PID parameter both satisfy corresponding preset conditions; if the frequency locking state score and the fast-slow double loop PID parameter do not both satisfy the corresponding preset conditions, continuously acquiring the current error signal time sequence, and iteratively optimizing the deep learning model based on the current error signal time sequence until the frequency locking state score and the fast-slow double loop PID parameter both satisfy the corresponding preset conditions; outputting the fast-slow double loop PID parameter output by the optimized deep learning model to a fast-slow double loop digital PID controller, so that the fast-slow double loop digital PID controller regulates and controls the PDH laser frequency locking system based on the input fast-slow double loop PID parameter; the deep learning model adopts a time series neural network architecture, and the time series neural network architecture comprises a time series encoder, a frequency locking state score output end and a fast-slow double loop PID parameter output end; the time series encoder is used to extract features of model input to obtain a feature vector, wherein, in an inference process of the deep learning model, the model input comprises the error signal time sequence and a multi-channel auxiliary quantity, the multi-channel auxiliary quantity comprises a combination of one or more of a laser current, a temperature control signal, an environmental temperature and humidity, a vibration noise and a historical control quantity, and all model inputs are input to the time series encoder after being aligned in time; the frequency locking state score output end is used to map the feature vector into a frequency locking state score; the fast-slow double loop PID parameter output end is used to map the feature vector into the fast-slow double loop PID parameter.

2. The multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system according to claim 1, characterized in that, The deep learning model adopts a GRU, LSTM or time series neural network structure fused with an attention mechanism.

3. The multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system according to claim 1, characterized in that, The deep learning model is trained according to the following steps: divide a training data set into a training set, a test set and a validation set, and use the training set to supervise learning and train the deep learning model, wherein the training data set comprises training PDH error signal time sequences collected under multiple disturbances and different working conditions, each training PDH error signal time sequence is labeled with a target frequency locking state score and corresponding optimal fast-slow double loop PID parameter group; the deep learning model calculates a value of a joint loss function based on the input training set, and optimizes model parameters of the deep learning model with the minimum value of the joint loss function as the target, wherein the joint loss function comprises a frequency locking score regression loss, a PID parameter regression loss, a regularization loss of a parameter change rate and a regularization loss of output smoothness constraint; verify the inference performance of the deep learning model based on the test set and the validation set.

4. The multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system according to claim 3, characterized in that, iteratively optimizing the deep learning model based on the current error signal time sequence comprises: acquiring operation data of a PDH laser frequency locking system, determining a current operation stage of the PDH laser frequency locking system based on the operation data; adjusting weights of a frequency locking score regression loss, a PID parameter regression loss, a regularization loss of a parameter change rate, and a regularization loss of output smoothness constraint in the joint loss function based on the current operation stage; iteratively optimizing the deep learning model based on the joint loss function after weight adjustment, the current error signal time sequence, and the acquired multi-channel auxiliary quantity.

5. The multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system according to claim 3, characterized in that, The training data set further includes an extended data set, and the extended data set includes typical system noise and abnormal signals.

6. The multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system according to claim 1, characterized in that, After outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller, the method further includes: monitoring operation data of a PDH laser frequency locking system; feeding back the operation data to the deep learning model, and performing online fine-tuning or incremental learning of the deep learning model based on the operation data.

7. The multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system according to claim 1, characterized in that, After outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller, the method further includes: monitoring system operation indicators of the PDH laser frequency locking system; determining whether the system operation indicators are abnormal, and when the system operation indicators are abnormal, outputting alarm information and / or re-reasoning the fast and slow double-loop PID parameters by using the deep learning model.

8. The multi-stage PID parameter regulation method for a deep learning-based laser frequency locking system according to claim 1, characterized in that, After outputting the fast and slow double-loop PID parameters output by the optimized deep learning model to the fast and slow double-loop digital PID controller, the method further includes: segmenting the fast and slow double-loop PID parameters output by the optimized deep learning model into segments according to a preset maximum step size and applying a control quantity slope limit value; at the end of each evaluation window, if the current frequency locking state score is not lower than the frequency locking state score in the previous window or the score drop value does not exceed a set drop threshold, the gradual entry is continued, and if the score drop value of a continuous target number of windows exceeds the set drop threshold or the fast and slow double-loop PID parameters output by the deep learning model reach a limiting threshold, the method reverts to the latest stable parameter group, and the fast and slow double-loop digital PID controller regulates and controls the PDH laser frequency locking system based on the latest stable parameter group, wherein the stable parameter group is the fast and slow double-loop PID parameters that meet the preset conditions.

9. A multi-stage PID parameter regulation system for a deep learning-based laser frequency locking system, applied to implement the multi-stage PID parameter regulation method for the deep learning-based laser frequency locking system according to any one of claims 1 to 8, characterized in that, The method includes: an error signal acquisition module configured to acquire an error signal time sequence of a PDH laser frequency locking system; The intelligent terminal is provided with a deep learning model, and a frequency locking state score and fast-slow double-loop PID parameters are obtained by reasoning based on the error signal time sequence through the deep learning model. If the frequency locking state score and the fast-slow double-loop PID parameters do not both satisfy the corresponding preset conditions, the error signal time sequence is continuously acquired through the error signal acquisition module, and the deep learning model is iteratively optimized based on the current error signal time sequence until the frequency locking state score and the fast-slow double-loop PID parameters both satisfy the corresponding preset conditions. The fast-slow double-loop PID parameters output by the optimized deep learning model are output to a fast-slow double-loop digital PID controller. The fast-slow double-loop digital PID controller is used for closed-loop adjustment of a laser and a TEC temperature control unit in a PDH laser frequency locking system according to the input fast-slow double-loop PID parameters.

Images