Optical element adjustment control method and system based on deep learning

Abstract

The application relates to the technical field of optical element control, and discloses an optical element adjustment control method and system based on deep learning. An optical data acquisition module of the system is provided with an optical sensor array on the surface of a target optical element, and real-time deformation and optical path offset data are collected; a deep learning feature extraction module constructs a neural network model, and outputs a multi-dimensional feature vector after inputting data; a dynamic baseline modeling module establishes a dynamic behavior baseline model and updates parameters accordingly; an initial adjustment signal generation module compares the feature vector with the baseline model, and generates an initial control signal; a primary adjustment feedback module collects deviation data after primary adjustment; an error correction module analyzes the deviation data to generate a correction coefficient and a final control signal; and a final adjustment execution module prompts an adjuster to execute final adjustment and output optical path stability evaluation data, so that accurate adjustment of the optical element can be realized, and the optical path stability is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of optical element control technology, specifically to an optical element adjustment and control method and system based on deep learning. Background Technology

[0002] During the operation of an optical system, the stability of optical components directly affects the overall optical path performance. Optical components are susceptible to environmental interference in their application environments. For example, temperature changes can cause thermal expansion and contraction, leading to deformation; vibration can cause component displacement, resulting in optical path deviation. These changes directly affect the accuracy and stability of the optical path, leading to a decrease in the output precision of the optical system and failing to meet the requirements of high-precision optical applications.

[0003] Control methods for adjusting optical components often rely on traditional preset parameter adjustment or simple feedback control. Traditional preset parameter adjustment is based on experience-based, fixed adjustment parameters, which cannot be adjusted according to real-time changes in the optical component's state. When the optical component experiences unexpected deformation or optical path misalignment, the preset parameters are difficult to adapt to the actual situation, resulting in poor adjustment performance. Simple feedback control, on the other hand, typically adjusts based on only a single real-time data point, lacking in-depth data analysis and processing. For example, generating control signals directly from acquired optical path misalignment data fails to consider the multi-dimensional characteristics contained in the data and cannot establish a benchmark model related to the dynamic behavior of the optical component, leading to lag and limitations in the adjustment process.

[0004] Traditional control methods struggle to achieve precise error correction when faced with complex changes in the state of optical components. After the initial adjustment is performed, simply calculating and correcting the deviation data fails to fully uncover the underlying causes, potentially leading to errors in the final adjustment and compromising the long-term stability of the optical path. Furthermore, traditional control methods lack a systematic evaluation mechanism for the stability of the adjusted optical path, making it difficult to comprehensively understand the adjustment effect and hindering the continuous optimization and stable operation of the optical system. These problems are particularly pronounced in fields with high requirements for optical path stability, such as high-precision optical inspection and laser communication. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for adjusting and controlling optical elements based on deep learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a deep learning-based optical element adjustment and control system, the system comprising:

[0007] An optical data acquisition module is used to set an optical sensor array on the surface of a target optical element to acquire real-time deformation data and optical path offset data of the optical element.

[0008] The deep learning feature extraction module is used to construct an optical feature extraction neural network model. The real-time deformation data and optical path offset data are input into the optical feature extraction neural network model, and the multi-dimensional feature vector of the optical element is output.

[0009] The dynamic baseline modeling module establishes a dynamic behavior baseline model for optical element adjustment based on the multi-dimensional feature vectors, and updates the parameters of the dynamic behavior baseline model in real time through cross-level feature correlation analysis.

[0010] The initial adjustment signal generation module compares the multi-dimensional feature vector with the dynamic behavior baseline model to generate the initial control signal for the optical element adjuster.

[0011] The primary adjustment feedback module is used to collect primary deviation data between the actual optical path parameters and the expected optical path parameters after the optical element adjuster performs a primary adjustment action based on the initial control signal.

[0012] The error correction module analyzes the primary deviation data through a deep learning error correction model, generates correction coefficients for the initial control signal, and generates the final control signal for the optical element adjuster based on the correction coefficients.

[0013] The final adjustment execution module is used by the optical element adjuster to perform final adjustment actions based on the final control signal and output optical path stability evaluation data after final adjustment.

[0014] Preferably, the optical data acquisition module includes:

[0015] An optical sensor array deployment unit is used to deploy high-precision optical sensors in the critical deformation areas of target optical elements;

[0016] The synchronous acquisition unit synchronously acquires real-time deformation data and optical path offset data of optical elements according to a preset time interval;

[0017] The data preprocessing unit is used to perform time alignment and noise filtering on the real-time deformation data and optical path offset data.

[0018] Preferably, the deep learning feature extraction module includes:

[0019] The convolutional feature extraction unit extracts the spatial distribution features of the real-time deformation data through a three-dimensional convolutional neural network;

[0020] The temporal feature extraction unit extracts the temporal series features of the optical path offset data through a long short-term memory network;

[0021] The feature fusion unit is used to perform cross-modal fusion of the spatial distribution features and time series features to generate the multi-dimensional feature vector.

[0022] Preferably, the dynamic baseline modeling module includes:

[0023] The baseline initialization unit establishes the initial parameters for the dynamic behavior baseline model based on historical optical element adjustment data;

[0024] The online update unit is used to dynamically adjust the weight distribution of the dynamic behavior baseline model through the correlation analysis between the multi-dimensional feature vector and real-time environmental parameters.

[0025] Preferably, the initial adjustment signal generation module includes:

[0026] The deviation calculation unit is used to calculate the Euclidean distance between the multi-dimensional feature vector and the corresponding features of the dynamic behavior baseline model;

[0027] The signal mapping unit maps the Euclidean distance into the initial control signal of the optical element regulator through a fully connected neural network.

[0028] Preferably, the error correction module includes:

[0029] An error coding unit is used to perform dimensionality reduction coding on the primary deviation data through an autoencoder network;

[0030] The correction coefficient generation unit generates the correction coefficients of the initial control signal based on the encoded error characteristics and historical correction records.

[0031] Preferably, the final adjustment execution module includes:

[0032] The execution feedback unit is used to record the drive parameters when the optical element adjuster performs the final adjustment action;

[0033] The stability assessment unit analyzes the frequency domain characteristics of the optical path stability assessment data after final adjustment using wavelet transform.

[0034] Preferably, the system further includes: an anomaly detection module, used to perform multi-scale anomaly pattern recognition on the final adjusted optical path stability evaluation data, and to mark the abnormal optical path fluctuation range, including:

[0035] A multi-scale decomposition unit is used to decompose the optical path stability evaluation data into intrinsic mode components of different scales through empirical mode decomposition.

[0036] Anomaly marking unit identifies anomalous optical path fluctuation ranges based on the energy distribution characteristics of each intrinsic mode component.

[0037] Preferably, the system further includes: a source tracing analysis module, which combines the abnormal optical path fluctuation range with historical optical element adjustment data, and uses a deep learning source tracing network to trace the root cause path of the abnormal optical path fluctuation, including:

[0038] The graph neural network construction unit is used to establish a topological relationship between the adjustment parameters of optical components and optical path fluctuations.

[0039] The path backtracking unit is used to locate the root cause node of abnormal optical path fluctuations in the topological graph through a graph attention network.

[0040] Preferably, the present invention also includes a deep learning-based optical element adjustment and control method, which includes all the modules and method flow of the above-mentioned deep learning-based optical element adjustment and control system.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This deep learning-based optical component adjustment and control system, through the inclusion of an optical data acquisition module and an array of optical sensors on the surface of the target optical component, simultaneously acquires real-time deformation and optical path offset data. This enables comprehensive perception of the component's state, avoiding the incomplete component state monitoring issues caused by traditional single-data acquisition methods. This module can accurately capture the component's state changes under different environments, providing comprehensive data input for subsequent adjustments and making the adjustment process more closely aligned with the component's actual condition.

[0043] The deep learning feature extraction module constructs an optical feature extraction neural network model, inputting real-time deformation data and optical path offset data into the model and outputting multi-dimensional feature vectors. This process can deeply mine the spatial, temporal, and other features contained in the data, breaking through the limitations of traditional data processing methods that can only extract surface data information. This makes the acquired features more reflective of the true state of the optical components, providing more valuable reference for subsequent dynamic baseline modeling and adjustment signal generation.

[0044] The dynamic baseline modeling module establishes a dynamic behavior baseline model based on multi-dimensional feature vectors and updates the model parameters in real time through cross-level feature correlation analysis. Compared with the traditional fixed-reference adjustment method, this dynamic baseline model can adjust with the changes in the state of the optical components, always maintaining adaptability to the current dynamic behavior of the components, making the generation of subsequent adjustment signals more targeted and avoiding adjustment deviations caused by a fixed reference.

[0045] The initial adjustment signal generation module compares the multi-dimensional feature vector with the dynamic behavior baseline model to generate the initial control signal, so that the initial adjustment action can be carried out based on the difference between the real-time features of the component and the dynamic baseline, ensuring that the initial adjustment direction and force meet the current state requirements of the component and reducing the ineffective operation caused by traditional blind adjustment.

[0046] After the optical element adjuster performs the primary adjustment action, the primary adjustment feedback module collects the primary deviation data between the actual optical path parameters and the expected optical path parameters, providing specific deviation information for subsequent error correction, so that the error correction process has a clear target and avoids blind correction.

[0047] The error correction module analyzes the primary deviation data using a deep learning error correction model, generates correction coefficients, and obtains the final control signal. This process leverages the deep learning model's in-depth analysis of the deviation data to fully uncover the intrinsic factors contributing to the deviation, resulting in more accurate correction coefficients. The final control signal obtained based on this accuracy can effectively correct deviations in the primary regulation, improving regulation precision.

[0048] After the optical element adjuster performs the final adjustment action, the final adjustment execution module outputs the optical path stability evaluation data after the final adjustment. It can comprehensively present the optical path status after adjustment, making it easier for staff to understand the adjustment effect and providing intuitive reference information for the subsequent operation and maintenance of the optical system. This makes the entire adjustment process form a complete closed loop and ensures the long-term stable operation of the optical system. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the working principle of the deep learning-based optical element adjustment and control system described in this invention.

[0050] Figure 2 This is a schematic diagram of the working principle of the optical data acquisition module;

[0051] Figure 3 A schematic diagram illustrating the working principle of the dynamic baseline modeling module. Detailed Implementation

[0052] 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, and 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.

[0053] Please see Figure 1This invention provides a method and system for adjusting and controlling optical components based on deep learning. The system integrates optical sensing, deep learning modeling, and feedback control to achieve precise adjustment of optical component deformation and optical path offset. An optical data acquisition module deploys an array of optical sensors on the surface of the target optical component to collect deformation and optical path offset data in real time. This data, after time alignment and noise filtering preprocessing, is input to a deep learning feature extraction module. The deep learning feature extraction module constructs an optical feature extraction neural network model. This model extracts the spatial distribution features of deformation data through a three-dimensional convolutional neural network and extracts the temporal series features of optical path offset data through a long short-term memory network, ultimately fusing them to generate a multi-dimensional feature vector. A dynamic baseline modeling module establishes a dynamic behavior baseline model for optical component adjustment based on the multi-dimensional feature vector. The model parameters are updated in real time through cross-level feature correlation analysis to adapt to environmental changes. An initial adjustment signal generation module compares the multi-dimensional feature vector with the dynamic behavior baseline model, calculates the feature Euclidean distance, and generates the initial control signal for the optical component adjuster through a fully connected neural network mapping. The primary adjustment feedback module collects primary deviation data between the actual and expected optical path parameters after the optical element regulator performs primary adjustment. The error correction module analyzes the primary deviation data using a deep learning error correction model, performs dimensionality reduction encoding using an autoencoder network, and generates correction coefficients for the initial control signal by combining historical correction records, thus generating the final control signal. The final adjustment execution module drives the optical element regulator to perform final adjustment actions based on the final control signal, records the driving parameters, and analyzes the frequency domain characteristics of the optical path stability evaluation data after final adjustment using wavelet transform. The system also includes an anomaly detection module and a source analysis module, which perform multi-scale anomaly pattern recognition on the optical path stability evaluation data and trace the root causes of anomalies using a graph neural network, thereby forming closed-loop control.

[0054] Example 1: See Figure 2The implementation of the optical data acquisition module involves laying out a high-density sensor network on the surface of optical elements. The optical sensor array deployment unit is responsible for precisely installing high-precision optical sensors in the key deformation areas of the target optical element. These key areas are usually predetermined through finite element analysis, including the edge of the element, near support points, and areas with significant changes in surface curvature. The sensors are of the fiber Bragg grating type, with the grating length customized according to the measurement accuracy requirements, typically on the order of millimeters. Each sensor is adhered to the surface of the element with UV-curable adhesive and is encapsulated with temperature compensation to eliminate the influence of environmental thermal noise. The array arrangement is designed according to the shape characteristics of the optical element. For circular mirrors, a radial and circumferential interwoven grid layout is used, while for square mirrors, an orthogonal matrix layout is used. The sensor spacing is optimized to capture deformation signals of the main bending and torsional modes. The leads of all sensors converge into a centralized data acquisition chassis. The synchronous acquisition unit triggers all sensors to perform synchronous measurements at fixed time intervals. This time interval is uniformly controlled by the master clock module. The clock signal is distributed to each acquisition channel through cables of equal length to eliminate transmission delay differences. The acquisition frequency is set according to the bandwidth requirements of the optical system. For example, a sampling rate of 2000 times per second is used in high-speed adjustment applications. Each sampling point contains deformation, temperature reading, and timestamp information.

[0055] The acquisition of optical path offset data is performed in parallel. A four-quadrant detector or wavefront sensor is used to obtain the beam center position or wavefront phase distribution. The sampling time is strictly aligned with the deformation data acquisition time via a hardware trigger signal, with synchronization errors controlled within microseconds. The data stream is transmitted to the data processing unit in real time via a high-speed serial bus. The data preprocessing unit receives the raw deformation and optical path data streams and first performs time alignment. Due to slight differences in the response times of different sensors, the algorithm uses cubic spline interpolation to resample the data from all channels onto a unified time grid. Next, noise filtering is performed. For high-frequency mechanical vibration noise commonly found in deformation data, a wavelet transform-based thresholding method is used. The Db4 wavelet basis is selected for multi-scale decomposition, and soft thresholding is applied to the detail coefficients before reconstructing the signal. For random photoelectric noise in the optical path data, an adaptive Kalman filter is used for smoothing. The process noise and observation noise covariance matrix of the filter are updated online based on historical data statistics. The preprocessed data is organized into a time-series dataset, with each time point containing a complete set of deformation field distributions and corresponding optical path parameters.

[0056] The convolutional feature extraction unit of the deep learning feature extraction module receives preprocessed 3D deformation data, which is organized into a displacement value matrix of spatial grid points. The convolutional neural network adopts a 3D architecture, and its input layer size corresponds to the dimension of the sensor grid. For example, 20×20×10 represents the number of sampling points in the length, width, and time dimensions. The network structure contains four convolutional layers. The first layer uses a 5×5×3 convolutional kernel for feature mapping, with a stride of 1 and "same" padding to maintain the feature map size. Each convolutional layer is followed by a 3D batch normalization layer and a ReLU activation function. The pooling layer uses 2×2×2 max pooling to gradually reduce the spatial resolution while preserving salient features. The final flattened feature vector contains hierarchical information from local deformation patterns to global bending features. The temporal feature extraction unit processes the optical path offset time series data in parallel. The long short-term memory network is designed as a three-layer structure with 128 hidden units per layer. The network input is the sequence of optical path parameters within a past time window, and the window length is set according to the response characteristics of the optical system. Each LSTM unit contains an input gate, a forget gate, and an output gate. Information flow is controlled by the sigmoid and tanh functions. The network unfolds along the temporal dimension, gradually learning long-term dependencies in dynamic patterns such as optical path drift and periodic jitter. The hidden state at the last time step is extracted as a temporal feature vector representing the overall behavior of the sequence. The feature fusion unit performs cross-modal fusion between the spatial feature vector output by the convolutional network and the temporal feature vector output by the LSTM network. Before fusion, the two vectors are projected onto the same dimension using a fully connected layer. Then, a multi-head attention mechanism is used to calculate the interaction weights between the spatial and temporal features. The attention score is calculated using a query-key-value pair mechanism, where the spatial features serve as the query vector and the temporal features serve as the key-value vector. The weighted summed features are then residually concatenated with the original features.

[0057] The final generated multi-dimensional feature vector has a fixed dimension of 512. This vector simultaneously encodes the instantaneous deformation state of the optical element and the dynamic trend of optical path evolution, providing an information-rich representation for subsequent baseline modeling. Calibration of the optical sensor array is completed before actual deployment. Each sensor is calibrated in a controlled environment by applying known deformation through a precision displacement platform, recording the correspondence between sensor output and actual displacement, and generating individual calibration coefficients. These coefficients are stored in non-volatile memory and are used during data preprocessing to correct the raw readings, eliminating gain differences and zero-point drift between sensors. The clock module of the synchronous acquisition unit uses a temperature-compensated crystal oscillator as the frequency reference and periodically synchronizes with the GPS signal to ensure the stability of the time reference during long-term acquisition. The training data for the convolutional neural network comes from historically accumulated optical element deformation datasets. Mean squared error is used as the loss function during training, the Adam algorithm is selected as the optimizer, the initial learning rate is set to 0.001, and an exponential decay strategy is employed.

[0058] Network parameters are adjusted using the backpropagation algorithm, and early stopping is employed during training to prevent overfitting. The Long Short-Term Memory (LSTM) network is pre-trained using a time-series prediction task, learning the dynamic characteristics of the sequence by predicting optical path parameters at future moments. The attention mechanism parameters in the feature fusion stage are jointly fine-tuned with the feature extraction network, optimizing the entire feature extraction process end-to-end. The generation process of multi-dimensional feature vectors features an interpretable design. Visualization of the convolutional network shows that its shallow convolutional kernels primarily respond to local curvature changes, while deeper kernels are sensitive to overall bending patterns. Forget gate state analysis of the LSM network indicates that its memory period is shorter for sudden optical path perturbations and longer for slow drifts. This design ensures that the feature vectors contain detailed spatial deformation information and encompass the temporal context of optical path evolution, providing a comprehensive and robust feature representation for dynamic behavior baseline modeling. Each dimension of the feature vectors is normalized to adapt its numerical range to the input requirements of downstream modules.

[0059] Example 2: See Figure 3The implementation of the dynamic baseline modeling module focuses on establishing a reference model that can characterize the normal adjustment behavior of optical components and possess adaptive capabilities. The module's operation begins with the baseline initialization unit's in-depth analysis of historical optical component adjustment data. This historical data originates from a database accumulated over long-term operation of the optical platform, containing multi-dimensional feature vector sequences recorded during tens of thousands of successful adjustment cycles, corresponding environmental parameter records, and regulator status logs. The baseline initialization unit employs a Gaussian mixture model as the mathematical framework for the dynamic behavior baseline model because it effectively characterizes the multimodal probability distribution of multi-dimensional feature vectors under normal conditions. Model parameters are learned using an expectation-maximization algorithm on the historical dataset. The algorithm iteratively optimizes the mean vector, covariance matrix, and mixing coefficients of each Gaussian component until the likelihood function converges. The resulting model describes the feature vector distribution of the optical component under standard operating conditions. Any new feature vector can be evaluated for its deviation from the baseline by calculating its log-likelihood under this model.

[0060] The online update unit is responsible for adapting the baseline model to the slow drift of optical component performance and changes in environmental conditions. This unit continuously receives real-time multi-dimensional feature vectors from the feature extraction module and readings from environmental monitoring sensors, including bench temperature, cooling water flow rate, laboratory air humidity, and foundation vibration spectrum data. The update process employs a recurrent neural network-based correlation analysis mechanism. This network takes the current environmental parameter sequence and historical feature vectors as input, and its output is used to predict the expected range of the feature vectors at the next time step. The observed feature vectors are compared with the predicted values, and the magnitude and direction of the residuals are used as the basis for model updates. The weight distribution update of the dynamic behavior baseline model is achieved through a sliding window maximum likelihood estimation method. The system maintains a fixed-length recent data window, and the data within the window is considered representative of the current operating status. Every certain time interval or when new data arrives, the model parameters are re-estimated based on all the data within the window. This mechanism allows the model to gradually forget outdated historical patterns while reinforcing recently emerging new normal patterns.

[0061] The model update triggering condition does not rely entirely on a fixed schedule. Instead, it employs an adaptive rule based on uncertainty. When the Mahalanobis distance between the real-time feature vector and the current baseline model prediction exceeds a preset threshold for multiple consecutive periods, the system determines that a significant change in the operating status has occurred and immediately initiates a model parameter re-estimation process. The adjustment of weight distribution is specifically reflected in the mixing coefficients of each component of the Gaussian mixture model. Components with high consistency with recent data have their mixing coefficients increased, while those with low consistency are decreased. The covariance matrix is ​​also scaled according to the dispersion of the new data, allowing the model to more tightly encompass the current data distribution. To handle transient anomalous data caused by sudden environmental disturbances, the online update unit also integrates an outlier robustness mechanism. Before including new data points in the update window, its local outlier factor relative to existing data points within the window is calculated. If a data point is determined to be significantly anomalous, its weight is reduced during the parameter update process, or it may even be temporarily excluded from the window to prevent it from contaminating the model. The entire online update process takes place in an isolated computing sandbox. Any parameter update must undergo a consistency check, meaning that the updated model should not show a significant decrease in its ability to describe historical normal data. Only after passing the check will the new model parameters be submitted and activated, replacing the original baseline model.

[0062] The dynamic baseline modeling module maintains close data interaction with other parts of the system. It provides the initial adjustment signal generation module with the latest baseline model parameters to calculate the deviation of real-time eigenvectors. Simultaneously, it receives optical path stability evaluation results from the final adjustment execution module as a monitoring signal. If the optical path stability remains good after adjustment, it confirms the effectiveness of the current baseline model; conversely, if stability decreases, it may indicate that the baseline model needs more aggressive adjustment. The module has an internal model version management function. Each parameter update saves a snapshot of the old model, recording the reason for the update and related data. This allows for rollback to a previous stable version when necessary, providing a safety net mechanism for the system. The status of the baseline model, including the number of currently active Gaussian components, the mean and covariance norm of each component, etc., is displayed in real-time on the system monitoring interface, allowing operators to assess the long-term operational health of the optical components. The model initialization process is not a one-time event. When optical components undergo major maintenance, replacement, or the system undergoes a major upgrade, it is necessary to re-collect basic operational data for a period of time and execute a complete baseline initialization process to establish a starting point that matches the new state. The learning rate parameter of the online update unit can be configured. A smaller learning rate can be used in a relatively stable laboratory environment to make the model change more smoothly, while a larger learning rate can be used on a test platform with greater environmental fluctuations to enable the model to adapt to changes more quickly.

[0063] Example 3: The core of the initial adjustment signal generation module lies in transforming abstract optical state characteristic differences into specific, executable physical adjustment commands. This module receives multi-dimensional feature vectors from the upstream feature extraction module and the current effective baseline model parameters provided by the dynamic baseline modeling module. The deviation calculation unit starts working first, and its task is to quantify the degree of deviation between the current real-time state of the optical element and the normal state baseline perceived by the system. This unit does not simply compare the real-time feature vector with a single center point in the baseline model, but considers the probability distribution characteristics described by the entire baseline model. The calculation process involves evaluating the anomaly score of the real-time feature vector relative to the dynamic behavior baseline represented by the Gaussian mixture model. A common method is to calculate the negative log-likelihood value of the vector to the model distribution. The larger this value, the more the current state deviates from the normal operating range.

[0064] The deviation value can be calculated as follows:

[0065]

[0066] Where: symbol This represents the calculated anomaly score scalar value; a higher value indicates a greater deviation between the real-time state and the baseline. (Symbol) This represents the current multi-dimensional feature vector fed from the feature extraction module; it is an n-dimensional column vector. (Symbol) This represents the total number of Gaussian components included in the dynamic behavior baseline model. (Symbol) Representing the The mixing coefficients of the Gaussian components satisfy the following conditions: It reflects the importance weight of this component when describing the normal state. (Symbol) and Representing the first The mean vector and covariance matrix of each Gaussian component together define the position and shape of that component in the feature space. (Symbol) Indicates that at a given mean Covariance Under the condition that vector The probability density function value under this multivariate Gaussian distribution. The calculation process traverses all values ​​in the baseline model. Each Gaussian component is used to calculate its probability density value. The product of each component's probability density value is multiplied by its weight, and the sum of these products is taken. Finally, the negative logarithm of this sum is calculated to obtain a comprehensive deviation measure. This scalar deviation value It captures an overall impression of the degree of abnormality in the current system state, but it does not contain any directional information on how to correct this deviation. For example, it cannot distinguish whether the voltage of a certain actuator should be increased or decreased.

[0067] The task of the signal mapping unit is precisely to solve this directionality problem; it maps the deviation scalar. This is mapped to specific, multi-dimensional initial control signals for the optical element regulator. This mapping is implemented using a deep feedforward neural network (fully connected neural network). The input layer of this network has only one neuron, used to receive the deviation value. The network typically contains three hidden layers, with the number of neurons in each layer depending on the complexity of the controlled optical element. For example, a typical architecture might include three hidden layers with 128, 64, and 32 neurons respectively. The hidden layers use the ReLU activation function to introduce nonlinear transformation capabilities. The number of neurons in the output layer strictly corresponds to the control degrees of freedom of the optical element's regulator. For example, a mirror system that needs to control six actuators (three for translation and three for tilt) will output a 6-dimensional control signal vector. Training this neural network requires a large amount of historical data; the training dataset consists of triplets from historical records: {historical deviation value} The actual regulator control signal vector used at that time Evaluation of optical path stability after adjustment The learning objective of the network is to find a mapping function. Such that for a given input Its output control signal It tends to produce good stability results. The loss function typically combines mean squared error and a regularization term, and the optimization process employs the backpropagation algorithm and the Adam optimizer. After the network training is complete, the real-time bias value... When input to the network, the signal undergoes linear transformations and nonlinear activations at each layer, ultimately generating a preliminary control signal vector at the output layer. This vector may require a post-processing step, such as scaling according to the actual drive range of each actuator, before it can become the initial control signal that can be directly executed by the regulator.

[0068] The operation of the initial conditioning signal generation module is highly dependent on the dynamic baseline model; online updates of the model parameters directly change the deviation value. The calculation method affects the generated initial control signal. The signal generated by this module is labeled "initial" or "primary" because it is based on a static, real-time state assessment, without considering the dynamic response characteristics of the actuator or potential secondary effects after adjustment. This initial control signal is sent to the regulator to execute the primary action, and the signal and its corresponding input deviation value are recorded, thus forming a closed-loop control circuit from perception to action. The module typically includes output limiting logic to prevent extreme control commands that could damage hardware due to momentary model anomalies or excessive input deviations.

[0069] Example 4: The error encoding unit of the error correction module receives primary bias data from the primary adjustment feedback module. This data is a multi-dimensional time series, which may include, for example, the offset of the beam center in the X and Y directions, and the sequence of the first 15 Zernike coefficients of wavefront aberrations changing over time. This data is massive and redundant, so the error encoding unit uses a sparse autoencoder network to reduce its dimensionality. The encoder part of the autoencoder consists of fully connected layers, with the input dimension being the same as the feature dimension of the primary bias data (e.g., 17 feature points multiplied by 10 time steps, totaling 170 dimensions). It is progressively compressed through three hidden layers (128, 64, and 32 nodes respectively), ultimately producing a 32-dimensional latent vector representation at the bottleneck layer. This latent vector is considered to capture the most essential and compact pattern features of the primary bias, such as whether it is mainly a combination of defocus aberration and slight tilt, or a pattern mainly consisting of a mixture of coma and astigmatism.

[0070] The correction coefficient generation unit generates correction coefficients based on this 32-dimensional encoded error feature. Internally, this unit maintains a historical correction record database, storing a large number of past adjustment cases' encoded error features, applied correction coefficients, and final adjustment effect scores. The process of generating correction coefficients can be analogous to a combination of retrieval and optimization. The unit calculates the cosine similarity between the current encoded error feature and the features of historical cases in the database, identifying the K most similar historical cases. A lightweight neural network (e.g., a two-layer fully connected network) takes the current encoded error feature and the mean of the correction coefficients of these K most similar cases as input, outputting a final set of correction coefficient vectors. These coefficients are used to correct the initial control signal, typically through element-wise multiplication: final control signal = initial control signal × correction coefficient. For example, if the initial control signal indicates that an actuator needs to extend by 10 micrometers, and the corresponding generated correction coefficient is 1.05, the final control signal will correct it to 10.5 micrometers. To illustrate the data flow more specifically, consider a simplified scenario: a system for adjusting the tilt angle of a reflector, which still has beam pointing deviation after initial adjustment, see Table 1.

[0071] Table 1: Primary Deviation Data and Generated Correction Factors

[0072]

[0073] Table 1 shows the residual X and Y axis deviations after primary adjustment over several consecutive time steps. The error encoding unit encodes the sequences of these two dimensions into a low-dimensional feature vector (only the values ​​of the first two dimensions are shown in the table). Based on this feature vector, the correction coefficient generation unit retrieves the most similar case from the historical database (e.g., Case_1024, whose feature might be "negative Y-axis deviation exists and is decaying") and generates a set of correction coefficients. It can be seen that when the deviation is large (t1), the system tends to apply a stronger overshoot correction (X-axis coefficient 1.12) and a reverse compensation (Y-axis coefficient 0.95); when the deviation is close to zero (t4), the correction coefficient also tends to 1.0 to avoid introducing new disturbances. The execution feedback unit of the final adjustment execution module is responsible for sending the final control signal to the optical element regulator and accurately recording the execution process. The regulator is usually a precision driver such as a piezoelectric ceramic actuator or a voice coil motor. The execution feedback unit records the actual displacement curve, drive voltage / current waveform, and time delay from issuing the command to completing the action of each actuator at a high sampling rate after receiving the final control signal. These driving parameters, along with timestamps from a high-precision clock, are stored in a log, forming a complete "execution profile" for each adjustment action. For example, the log might show that the instruction required actuator A to move 12.5 nanometers within 5 milliseconds, but the actual movement took 5.1 milliseconds with slight overshoot oscillations. These profiles are crucial for tracing the reasons for poor adjustment performance. For instance, if stability remains unsatisfactory after final adjustment, analyzing the execution profile might reveal a nonlinear delay in the response of a particular actuator.

[0074] The stability assessment unit is activated after the final adjustment action is completed and the optical system has briefly stabilized. It acquires a new segment of optical path data, such as wavefront sensor readings or beam position detector signals lasting several hundred milliseconds. The assessment is no longer limited to simple averages or standard deviations, but instead employs wavelet transform for frequency domain feature analysis. This unit selects a set of wavelet basis functions (such as the Daubechies wavelet family) to decompose the acquired optical path stability assessment data into different frequency sub-bands. The signal energy in each sub-band is calculated, thus obtaining the energy spectrum of optical path fluctuations. An ideal stable optical path should have its fluctuation energy primarily concentrated in the extremely low-frequency region (corresponding to slow thermal drift), while the energy is very low in the mid-to-high frequency region. The assessment unit calculates several key indicators, such as total energy, the energy percentage of a specific frequency band (e.g., 1-100Hz), and the main fluctuation frequency components. These frequency domain features contain more information than the standard deviation in the time domain, enabling the differentiation of different types of disturbances; for example, high-frequency vibrations manifest as energy concentration in the high-frequency band, while mechanical creep manifests as slow changes in energy at extremely low frequencies. A rapid internal closed loop is formed between the error correction module and the final adjustment execution module. The stability evaluation data obtained after the final adjustment, especially its frequency domain characteristics, is sent back to the historical database of the error correction module as immediate feedback. If the final adjustment is successful (manifested as a significant reduction in optical path fluctuation energy and a pure spectrum), the "encoded error feature-correction coefficient" combination will be stored as a successful positive sample in the historical database and may be associated with previously retrieved similar cases, increasing the weight of those cases. Conversely, if the effect is unsatisfactory, the combination may be flagged and its priority reduced in subsequent similarity searches, or used to trigger fine-tuning of the correction coefficient generation network. Through this continuous learning and accumulation, the decision accuracy of the error correction module will gradually improve with the increase of system runtime, enabling it to predict and compensate for the remaining error after each primary adjustment more accurately. The entire process embodies the progressive control concept from coarse adjustment to fine correction, with the ultimate goal of making the optical system output a highly stable optical path.

[0075] Example 5: After final adjustment, the ability to continuously monitor and deeply analyze optical path stability is achieved. The anomaly detection module begins scanning the optical path stability evaluation data output by the final adjustment execution module. This data typically consists of high-precision wavefront errors or beam position signals acquired continuously over a long period. The multi-scale decomposition unit processes this data using the empirical mode decomposition method. This method is adaptive and can decompose complex non-stationary signals into a series of intrinsic mode functions (EMFs) arranged from high to low frequency. The decomposition process is an iterative sieving process. First, all local extrema in the original signal are identified. Spline curves are used to connect all maxima and minima to form upper and lower envelopes. The mean of the upper and lower envelopes is calculated as the first reference line. The original signal is subtracted from this reference line to obtain an intermediate signal. This operation is repeated until the intermediate signal satisfies the conditions of the EMF regarding the number of extrema and zero-crossing points. This yields the first and highest-frequency EMF component. Then, the original signal is subtracted from this first component, and the remaining signal is used as the new original signal. The above process is repeated to extract the second, third, and finally the last monotonic trend component.

[0076] For a typical optical path signal containing sudden jitter and slow drift, empirical mode decomposition may produce 6 to 8 physically meaningful modal components. The anomaly labeling unit then independently analyzes each intrinsic modal component, calculating the energy value of each component, typically defined as the average of the sum of squares of the data points for that component. Under normal operating conditions of the optical system, the energy distribution of each modal component exhibits a certain stable pattern; for example, high-frequency components have lower energy, while low-frequency components have relatively higher energy but change more gradually. The anomaly labeling unit internally stores the energy level range thresholds for each modal component under the system's baseline stable condition. When the energy of a specific modal component significantly exceeds its historical threshold within multiple consecutive analysis time windows—for example, exceeding three standard deviations of the threshold and persisting for several seconds—the unit marks this time interval as an anomalous optical path fluctuation interval. For example, if the normally weak third modal component (corresponding to the 20-50Hz frequency range) suddenly experiences an energy spike, the system will mark the start and end times of this spike, indicating that the optical platform may have been disturbed by mid-frequency mechanical vibration. The source analysis module starts working after the anomaly detection module marks a specific time interval, and its goal is to locate the root cause of the abnormal fluctuation.

[0077] The graph neural network construction unit first extracts all data related to the abnormal time period from the system database. These data nodes are diverse, including environmental parameter nodes (such as temperature readings, airflow velocity, and ground vibration acceleration at different locations within the laboratory), optical element regulator nodes (such as historical drive signals, real-time displacement feedback, and current consumption of various actuators), optical feature nodes (such as key dimension values ​​in the multi-dimensional feature vectors output by the feature extraction module), and optical path performance nodes (such as the RMS value of wavefront error and beam pointing stability). These nodes are not isolated; they are connected by edges. The construction of these edges is based on statistical relationships calculated from historical data. For example, the cross-correlation between the time series of two nodes exceeds a certain threshold, or there is a known physical causal relationship between them (such as a change in the current of an actuator directly causing a change in the mirror surface shape), thus forming a large-scale heterogeneous topological relationship graph. The path tracing unit then uses a graph attention network to perform source tracing analysis within this vast relationship graph. The working principle of the graph attention network is to calculate an attention weight for each node in the graph, which represents the importance of the node's association with the abnormal optical path fluctuation node. The analysis typically begins with nodes flagged as having anomalous optical path performance. The graph attention network propagates backward along the edges, iteratively calculating the attention scores of neighboring nodes. At each layer, a node aggregates information from its neighbors and updates its representation based on the strength of the edges (e.g., correlation coefficients) and the importance of its neighbors. After multiple iterations, a small subset of nodes in the graph acquires significantly higher attention weights than others. These high-weight nodes are considered candidate nodes for the root cause of the anomalous fluctuations.

[0078] The analysis results may show that the attention weights are highly concentrated on a vibration-isolation air leg pressure sensor node located on the west side of the optical platform, and an actuator current feedback node that controls the curvature of a certain area of ​​the mirror, while the weights of ambient temperature nodes are generally very low. This indicates that the anomaly likely originated from the instantaneous failure of the vibration isolation system on the west side, leading to changes in the mirror support stress, which in turn affected the optical path through that specific actuator. The entire analysis process generates a source tracing report, which lists the top nodes with the highest attention weights and their weight values, and graphically displays the optimal path from the anomalous optical path node to these root cause nodes. This report provides maintenance personnel with a clear diagnostic direction, enabling them to quickly locate the source of the problem, rather than blindly checking the entire optical system. The system archives these source tracing analysis cases and their conclusions in a knowledge base. When similar anomaly patterns occur in the future, historical cases can be retrieved first, thereby accelerating the diagnostic process. The addition of anomaly detection and source tracing analysis functions enables the system not only to perform real-time adjustment and control, but also to have preliminary fault diagnosis and predictive maintenance capabilities, improving the reliability and maintainability of the optical system.

[0079] Example 6: In the scenario of exhibition hall lighting adjustment, the target optical components are the optical lens group and reflector assembly responsible for light projection in the exhibition hall. The core requirement is to realize the automatic adaptation and adjustment of the light as visitors walk in or away, so as to ensure that visitors can clearly observe the exhibits from different viewing positions, while avoiding glare caused by direct light.

[0080] The optical sensor array deployment unit of the optical data acquisition module deploys two types of high-precision optical sensors in key areas around the exhibits: one is an infrared optical sensor for detecting visitor positions, and the other is a fiber optic sensor for monitoring the status of optical components and the optical path. The fiber optic sensors are installed at the light-emitting ports of the optical lens group and the edges of the reflective surfaces of the mirror assembly. The synchronous acquisition unit collects three types of data at preset time intervals: real-time deformation data of the optical lens group caused by temperature changes in the exhibition hall, optical path offset data caused by the offset of the mirror assembly installation position, and real-time location data of visitors captured by the infrared optical sensor (such as the distance between the visitor and the exhibit and their location). The data preprocessing unit performs time alignment on the above three types of data to ensure accurate matching of data from different sources in the time dimension. At the same time, it performs noise filtering to eliminate the influence of other light sources and vibrations from people walking in the exhibition hall on the data.

[0081] The optical feature extraction neural network model constructed by the deep learning feature extraction module inputs three types of preprocessed data into the model: the convolutional feature extraction unit extracts the spatial distribution features of real-time deformation data of optical components (such as deformation differences in different areas of the lens) and the spatial distribution features of visitor position data (such as the spatial coordinates of visitors relative to exhibits) through a three-dimensional convolutional neural network; the temporal feature extraction unit extracts the time-series features of optical path offset data (such as the trend of optical path offset over time) and the time-series features of visitor position data (such as the speed and direction of visitor movement) through a long short-term memory network; the feature fusion unit performs cross-modal fusion of the above spatial distribution features and time-series features to generate a multi-dimensional feature vector containing optical component state, optical path state, and visitor position information.

[0082] The dynamic baseline modeling module establishes a dynamic behavior baseline model based on multi-dimensional feature vectors: The baseline initialization unit combines historical operating data of the exhibition hall (such as the appropriate lighting adjustment parameters and optical component status data under different time periods and visitor traffic) to determine the initial parameters of the dynamic behavior baseline model. The initial parameters include the optimal lighting projection range, brightness, and optical component deformation threshold corresponding to different visitor positions; The online update unit dynamically adjusts the model weight distribution through correlation analysis between multi-dimensional feature vectors and real-time environmental parameters of the exhibition hall (such as real-time temperature and humidity in the exhibition hall). For example, when the temperature of the exhibition hall rises and causes the deformation trend of the optical components to change, the temperature-related weights in the model are adjusted in a timely manner to ensure that the baseline model always adapts to the dynamic behavior of the optical components under the current environment.

[0083] The initial adjustment signal generation module compares the multi-dimensional feature vector with the dynamic behavior baseline model: the deviation calculation unit calculates the Euclidean distance between the corresponding features of the two to determine the degree of fit between the current light path state and the visitor's position (such as whether the current light projection range matches the visitor's position); the signal mapping unit maps the Euclidean distance into an initial control signal through a fully connected neural network. The signal contains parameters for adjusting the focal length of the optical lens group and the angle of the reflector. When the visitor walks into the exhibit, an initial signal is generated to reduce the light projection range and appropriately reduce the brightness; when the visitor moves away from the exhibit, an initial signal is generated to expand the light projection range and appropriately increase the brightness.

[0084] After the optical element adjuster performs primary adjustment, the primary adjustment feedback module collects primary deviation data between the actual optical path parameters (such as the actual light projection range and brightness) and the expected optical path parameters (adaptation parameters determined based on the baseline model), such as the difference between the actual light brightness and the expected brightness, and the deviation between the projection range and the expected range.

[0085] The error correction module analyzes the initial deviation data through a deep learning error correction model: the error encoding unit performs dimensionality reduction encoding on the deviation data through an autoencoder network to extract the core features of the deviation; the correction coefficient generation unit combines the encoded error features with historical correction records (such as correction parameters of similar deviations in the past) to generate the correction coefficients of the initial control signal. For example, if the light brightness is still too high after the initial adjustment, a correction coefficient to reduce the brightness adjustment range is generated, and the final control signal is generated based on this.

[0086] The final adjustment execution module drives the optical element adjuster to perform final adjustment, precisely adjusting the lens focal length and reflector angle to ensure the light projection state perfectly matches the current visitor's position. Simultaneously, it outputs final adjustment optical path stability assessment data (such as light brightness stability and projection direction deviation range). The anomaly detection module performs multi-scale anomaly identification on the optical path stability assessment data. Through empirical mode decomposition, it decomposes the data into intrinsic mode components of different scales and marks abnormal intervals based on the energy distribution characteristics of each component (such as periods of sudden light brightness changes caused by sensor obstruction). The source analysis module constructs a topological relationship graph of optical element adjustment parameters, visitor position data, and optical path fluctuations using a graph neural network. It then uses a graph attention network to locate the root cause nodes of anomalies (such as position data acquisition errors caused by sensor obstruction), providing a basis for troubleshooting and optimization.

[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based optical element adjustment and control system, characterized in that, The system includes: An optical data acquisition module is used to set an optical sensor array on the surface of a target optical element to acquire real-time deformation data and optical path offset data of the optical element. The deep learning feature extraction module is used to construct an optical feature extraction neural network model. The real-time deformation data and optical path offset data are input into the optical feature extraction neural network model, and the multi-dimensional feature vector of the optical element is output. The dynamic baseline modeling module establishes a dynamic behavior baseline model for optical element adjustment based on the multi-dimensional feature vectors, and updates the parameters of the dynamic behavior baseline model in real time through cross-level feature correlation analysis. The initial adjustment signal generation module compares the multi-dimensional feature vector with the dynamic behavior baseline model to generate the initial control signal for the optical element adjuster. The primary adjustment feedback module is used to collect primary deviation data between the actual optical path parameters and the expected optical path parameters after the optical element adjuster performs a primary adjustment action based on the initial control signal. The error correction module analyzes the primary deviation data through a deep learning error correction model, generates correction coefficients for the initial control signal, and generates the final control signal for the optical element adjuster based on the correction coefficients. The final adjustment execution module is used for the optical element adjuster to perform final adjustment actions based on the final control signal, and to output optical path stability evaluation data after final adjustment; The system further includes: an anomaly detection module, used to perform multi-scale anomaly pattern recognition on the final adjusted optical path stability evaluation data, and to mark the abnormal optical path fluctuation range, including: A multi-scale decomposition unit is used to decompose the optical path stability evaluation data into intrinsic mode components of different scales through empirical mode decomposition. Anomaly marking unit identifies anomalous optical path fluctuation intervals based on the energy distribution characteristics of each intrinsic mode component; The system also includes: a source tracing analysis module, which combines the abnormal optical path fluctuation range with historical optical element adjustment data, and uses a deep learning source tracing network to trace the root cause path of the abnormal optical path fluctuation, including: The graph neural network construction unit is used to establish a topological relationship between the adjustment parameters of optical components and optical path fluctuations. The path backtracking unit is used to locate the root cause node of abnormal optical path fluctuations in the topological graph through a graph attention network.

2. The deep learning-based optical element adjustment and control system according to claim 1, characterized in that, The optical data acquisition module includes: An optical sensor array deployment unit is used to deploy high-precision optical sensors in the critical deformation areas of target optical elements; The synchronous acquisition unit synchronously acquires real-time deformation data and optical path offset data of optical elements according to a preset time interval; The data preprocessing unit is used to perform time alignment and noise filtering on the real-time deformation data and optical path offset data.

3. The deep learning-based optical element adjustment and control system according to claim 1, characterized in that, The deep learning feature extraction module includes: The convolutional feature extraction unit extracts the spatial distribution features of the real-time deformation data through a three-dimensional convolutional neural network; The temporal feature extraction unit extracts the temporal series features of the optical path offset data through a long short-term memory network; The feature fusion unit is used to perform cross-modal fusion of the spatial distribution features and time series features to generate the multi-dimensional feature vector.

4. The deep learning-based optical element adjustment and control system according to claim 1, characterized in that, The dynamic baseline modeling module includes: The baseline initialization unit establishes the initial parameters for the dynamic behavior baseline model based on historical optical element adjustment data; The online update unit is used to dynamically adjust the weight distribution of the dynamic behavior baseline model through the correlation analysis between the multi-dimensional feature vector and real-time environmental parameters.

5. The deep learning-based optical element adjustment and control system according to claim 1, characterized in that, The initial adjustment signal generation module includes: The deviation calculation unit is used to calculate the Euclidean distance between the multi-dimensional feature vector and the corresponding features of the dynamic behavior baseline model; The signal mapping unit maps the Euclidean distance into the initial control signal of the optical element regulator through a fully connected neural network.

6. The optical element adjustment and control system based on deep learning according to claim 1, characterized in that, The error correction module includes: An error coding unit is used to perform dimensionality reduction coding on the primary deviation data through an autoencoder network; The correction coefficient generation unit generates the correction coefficients of the initial control signal based on the encoded error characteristics and historical correction records.

7. The optical element adjustment and control system based on deep learning according to claim 1, characterized in that, The terminal adjustment execution module includes: The execution feedback unit is used to record the drive parameters when the optical element adjuster performs the final adjustment action; The stability assessment unit analyzes the frequency domain characteristics of the optical path stability assessment data after final adjustment using wavelet transform.

8. A method for adjusting and controlling optical components based on deep learning, characterized in that, It includes all modules and method flows of the deep learning-based optical element adjustment and control system as described in any one of claims 1 to 7.

Images