Intelligent double-control adjustable semiconductor laser and control method

Abstract

This invention discloses an intelligent dual-control tunable semiconductor laser and its control method, belonging to the field of semiconductor laser technology. The laser includes a semiconductor laser chip, a temperature sensing unit, a spot detection unit, a tunable optical path module, a switchable circuit drive module, a precision drive module, an AI control center, and a multi-modal control module. The multi-modal control module receives control commands and parses them into control intentions. The AI ​​control center extracts target parameters and calculates the deviation with temperature signals and spot quality parameters. Based on the deviation, it generates a coordinated control command containing drive circuit selection information and optical element displacement information, synchronously controlling the drive circuit to activate and drive the optical element displacement. After execution, the spot quality parameters are re-acquired to verify whether the deviation has been eliminated; if not, the execution is repeated. This invention has advantages such as coordinated control of drive mode and optical path, full closed-loop compensation for temperature drift, and multi-modal interaction.

Description

Technical Field

[0001] This invention relates to the field of semiconductor laser technology, and more specifically to an intelligent dual-control tunable semiconductor laser and its control method. Background Technology

[0002] Semiconductor lasers are widely used in communications, medical applications, industrial processing, and lidar due to their advantages such as small size, high efficiency, and wide wavelength range. However, existing semiconductor lasers face the following technical challenges in practical applications:

[0003] The separation of beam quality and driving mode: The driving circuit and optical system of traditional semiconductor lasers are independent of each other, and the driving mode and optical parameters cannot be dynamically matched according to real-time operating conditions. As a result, it is impossible to simultaneously achieve high peak power and high beam quality under complex process requirements.

[0004] Poor environmental adaptability: Temperature changes cause wavelength drift and divergence angle changes in laser chips. Traditional solutions rely on passive heat dissipation or simple PID temperature control, which cannot compensate for the position of optical components in real time, resulting in a decline in the quality of the laser spot during long-term operation.

[0005] Limited interaction methods: Existing devices mostly use manual knobs or PC software control, which is inconvenient for precise operation or special environments, and cannot achieve proactive perception of the operator's status.

[0006] Fixed optical path system: Optical components such as fast-axis collimating mirror, slow-axis collimating mirror, and focusing mirror are all fixed in a certain position in traditional semiconductor lasers. It is difficult to dynamically adjust them according to the light source requirements of different application scenarios, which limits the application of lasers in multiple scenarios.

[0007] Therefore, there is an urgent need in this field for a semiconductor laser and its control method that can achieve coordinated control of driving mode and optical path parameters, has active temperature drift compensation capability, supports multimodal interaction, and has dynamically adjustable optical path. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide an intelligent dual-controlled tunable semiconductor laser that overcomes or at least partially solves the above problems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A smart dual-control tunable semiconductor laser, comprising: Semiconductor laser chips; A temperature sensing unit is used to monitor the operating temperature of the semiconductor laser chip in real time and output a temperature signal; The spot detection unit is used to acquire the spot image output by the semiconductor laser in real time and extract the current spot quality parameters. An adjustable optical path module, located on the light-emitting side of the semiconductor laser chip, includes multiple optical elements; A switchable circuit drive module includes multiple drive circuits and a switch group for selectively connecting any drive circuit to the semiconductor laser chip; A precision drive module, mechanically connected to the plurality of optical elements, is used to drive the optical elements to generate displacement; The AI ​​control center, with a built-in neural network model, is connected to the temperature sensing unit, the spot detection unit, the switchable circuit drive module, and the precision drive module. It receives control commands from the operator and parses them into control intentions. It extracts target parameters from these intentions, compares them with the temperature signal and the current spot quality parameters to calculate the deviation, and generates a coordinated control command based on the deviation. This coordinated control command includes drive circuit selection information and optical element displacement information. According to the coordinated control command, it synchronously controls the switch group to connect the corresponding drive circuit and controls the precision drive module to drive the corresponding optical element to produce displacement. After the displacement is performed, the deviation is verified by re-collecting the spot quality parameters by the spot detection unit. If the deviation is not eliminated, the steps of generating the cooperative control command and driving the corresponding optical element to generate displacement are repeated.

[0010] Preferably, the control intent includes at least one of target driving mode, target spot parameters, and target focal position.

[0011] Preferably, the plurality of optical elements, along the laser emission direction, sequentially include a fast-axis collimating lens, a slow-axis collimating lens, a zoom lens group, and a focusing lens; The zoom lens group includes a zoom lens group and a compensation lens group that can move independently along the optical axis.

[0012] Preferably, the spot detection unit extracts the spot contour by performing edge detection on the spot image, and calculates the current spot diameter, ellipticity and energy distribution information based on the spot contour as the current spot quality parameters.

[0013] Preferably, the neural network model generates the thermal drift compensation displacement of the fast-axis collimating lens and / or the slow-axis collimating lens based on the temperature signal and the current spot quality parameters, as part of the optical element displacement information; After the precision drive module completes the displacement, the AI ​​control center verifies whether the divergence angle has converged based on the spot quality parameters re-collected by the spot detection unit. If it has not converged, the thermal drift compensation displacement is repeatedly generated and the corresponding optical element is driven.

[0014] Preferably, the deviation includes the diameter deviation between the target spot parameters and the current spot diameter, the positional deviation between the target focal position and the current focal position, and the temperature deviation of the temperature signal relative to the reference temperature.

[0015] Preferably, the neural network model generates a first displacement of the zoom lens group and a second displacement of the compensation lens group based on the diameter deviation between the target spot parameters and the current spot diameter, and uses the first displacement and the second displacement as input to query a pre-stored optical zoom-focus compensation curve to generate the following displacement of the focusing lens.

[0016] Preferably, the neural network model includes a temperature-compensated neural network model, a spot quality assessment compensation model, and a driving mode decision model; The temperature compensation neural network model takes the temperature signal as input and outputs the thermal drift compensation displacement of the fast-axis collimating mirror and the slow-axis collimating mirror. The spot quality assessment and compensation model takes the spot image as input and outputs the target displacement of the zoom lens group and the focusing lens; The drive mode decision model takes process requirement parameters and spot quality parameters as input and outputs the drive circuit selection information of the switch group. The temperature compensation neural network model, the light spot quality assessment compensation model, and the driving mode decision model are calculated in parallel, and the output results are merged into a single collaborative control command by the AI ​​control center.

[0017] Preferably, the spot quality assessment and compensation model further includes a confidence output head, used to output the confidence level of the target displacement output by the spot quality assessment and compensation model; The AI ​​control center is also used for: When the confidence level is greater than or equal to the preset threshold, the collaborative control command is directly generated from the output of the spot quality assessment and compensation model. When the confidence level is less than the preset threshold, the algorithm switches to hill climbing and drives the focusing lens to scan step by step within a preset range. Each step involves acquiring a spot image through the spot detection unit and calculating a sharpness evaluation value. The position of the focusing lens when the sharpness evaluation value reaches its peak is taken as the optimal focal position. The target displacement of the focusing lens in the collaborative control command is updated based on the difference between the optimal focal position and the current position of the focusing lens. The AI ​​control center also packages the light spot images collected during the hill-climbing search process with the corresponding optimal focal position and current temperature signal as training samples, and feeds back to update the light spot quality assessment and compensation model.

[0018] Based on the same inventive concept, this invention also discloses a control method for an intelligent dual-controlled tunable semiconductor laser, applicable to any of the aforementioned intelligent dual-controlled tunable semiconductor lasers, comprising: S1. Real-time monitoring of the operating temperature of the semiconductor laser chip and output of temperature signal, while real-time acquisition of spot image and extraction of current spot quality parameters; S2. Receive control commands and parse them into control intentions, and extract target parameters from the control intentions; S3. The target parameter is compared with the temperature signal and the current spot quality parameter to calculate the deviation, and a cooperative control command is generated based on the deviation using a pre-trained neural network model. The cooperative control command includes drive circuit gating information and optical element displacement information. S4. Execute the following actions synchronously according to the aforementioned coordinated control instructions: The corresponding driving circuit is controlled to be turned on by the driving circuit selection information, and the corresponding optical element is driven to generate displacement by the optical element displacement information. S5. After step S4 is completed, the spot image is re-acquired and the compensated spot quality parameters are extracted. S6. Calculate the residual deviation between the compensated spot quality parameters and the control intention. If the residual deviation exceeds a preset threshold, jump to step S3 to regenerate the cooperative control command and execute steps S4 to S5 again until the residual deviation converges to within the preset threshold.

[0019] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: (i) The AI ​​control center generates a coordinated control command that includes both drive circuit gating information and optical element displacement information, and synchronously controls the switch group and precision drive module to execute it. This achieves real-time coordination between drive mode and optical path parameters, eliminates the intermediate mismatch state of optical path lag after circuit switching, and avoids spot jitter and energy loss.

[0020] (ii) The chip temperature is monitored in real time by the temperature sensing unit. The AI ​​control center drives the fast axis collimator and the slow axis collimator to generate micron-level displacement compensation based on the temperature deviation and divergence angle information. After the displacement is executed, the spot quality parameters are re-acquired to verify whether the divergence angle has converged, forming a closed-loop control of temperature sensing-displacement compensation-spot verification, so that the laser can maintain stable beam quality in a wide temperature range.

[0021] (iii) By having multiple optical elements in the adjustable optical path module driven independently by the precision drive module, the independent programmable adjustment of beam collimation, spot size and focal position is realized, enabling the laser to adapt to a variety of application scenarios.

[0022] (iv) By receiving the operator's control instructions through the multimodal control module and parsing them into control intentions, a three-dimensional interactive architecture integrating multiple interaction methods is realized, which improves the convenience and safety of operation in special work scenarios.

[0023] (v) The confidence level is output by the spot quality assessment compensation model and the physical search is switched to hill climbing method when the confidence level is lower than the threshold. The search results are fed back to update the model, forming a hybrid architecture of AI decision-making and physical fallback, which combines the high efficiency of AI and the reliability of traditional algorithms. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a system architecture block diagram of the intelligent dual-control tunable semiconductor laser provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the overall structure of the fast-axis collimating lens driving structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the internal structure of the two-dimensional deflection stage in the fast-axis collimating mirror driving structure provided in an embodiment of the present invention; Figure 4 This is a flowchart of the intelligent dual-control tunable semiconductor laser control method provided in the embodiments of the present invention. Detailed Implementation

[0026] 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.

[0027] like Figure 1As shown, the intelligent dual-control tunable semiconductor laser provided by this invention has a system architecture that, from top to bottom, includes an input layer, a decision layer, an execution layer, and a light source and optical path layer, with a sensing feedback layer forming a feedback loop. The input layer is a multi-modal control module used to receive control commands from the operator; the decision layer is an AI control center with a built-in neural network model used to generate coordinated control commands based on control commands and sensing feedback data; the execution layer includes a precision drive module and a switchable circuit drive module. The precision drive module drives optical elements to produce displacement, while the switchable circuit drive module selectively connects the corresponding drive circuit to the semiconductor laser chip; the light source and optical path layer includes a semiconductor laser chip and a tunable optical path module. The laser generated by the semiconductor laser chip is modulated by the tunable optical path module before being output as a beam. The sensing module includes a temperature sensing unit and a spot detection unit. The temperature sensing unit is located near the semiconductor laser chip to monitor the chip's operating temperature in real time and output a temperature signal. The spot detection unit is located on the output beam path to acquire spot images in real time and extract spot quality parameters. The temperature sensing unit and the spot detection unit feed the collected data back to the AI ​​control center, forming a closed-loop control system.

[0028] In one embodiment, the semiconductor laser chip uses GaAs-based or InP-based materials, and can be selected to output 808nm, 915nm, 980nm, or 1550nm wavelengths according to actual application requirements. All the above modules are integrated into a single housing made of aluminum alloy or copper alloy. An internal partition physically isolates the circuit area from the optical area to reduce electromagnetic interference. The heat sink is tightly attached to the bottom of the housing using thermally conductive silicone grease, and the bottom of the housing has a finned heat dissipation structure or a water-cooling interface. The window uses flat optical glass with an anti-reflective coating, and an airtight seal is used between the window and the housing. The power supply interface uses a military-grade circular connector, and the PC interface uses an RJ45 or fiber optic interface, supporting Gigabit Ethernet communication.

[0029] The temperature sensing unit consists of a heat sink and a temperature sensor. A semiconductor laser chip is mounted above the heat sink, and a TEC cooler is mounted below it. The temperature sensor is mounted flush against the chip, with a distance of no more than 1 mm between them. The temperature sensor uses a negative temperature coefficient thermistor, and its raw digital temperature signal is sent to the AI ​​control center. After calibration and compensation, it is converted into an actual temperature value, and the temperature deviation and rate of temperature change are calculated.

[0030] The light spot detection unit includes a photodiode and a CMOS camera. The photodiode is mounted on the side of the housing and aligned with a portion of the reflected light path for real-time monitoring of the output light power. The CMOS camera is used to acquire light spot images in real time. In one embodiment, a macro auto-zoom lens and an auto-attenuation lens are sequentially mounted in front of its lens. The auto-attenuation lens is used to adjust the laser power entering the camera to prevent strong light from damaging the CMOS camera.

[0031] The CMOS camera acquires spot images at a frame rate of no less than 30 frames per second and transmits them to the AI ​​control center via the GigE Vision interface. The AI ​​control center uses the Canny edge detection algorithm to extract the spot contour, fits an ellipse, and calculates the spot diameter, ellipticity, and energy distribution information as current spot quality parameters.

[0032] In one embodiment, the adjustable optical path module is disposed on the light-emitting side of the semiconductor laser chip. The adjustable optical path module is provided with a fast-axis collimating lens, a slow-axis collimating lens, a zoom lens group, a focusing lens, and a window in sequence along the laser emission direction.

[0033] The fast-axis collimating lens uses a cylindrical aspherical lens with a focal length of 0.5mm to 2mm and a numerical aperture of not less than 0.8. The fast-axis collimating lens is set close to the light-emitting surface of the laser chip, at a distance of less than 100μm, and is used to compress the divergence angle of 40° to 60° in the fast-axis direction to below 5mrad.

[0034] The slow-axis collimator uses a cylindrical aspherical lens with a focal length of 10mm to 50mm. It is placed after the fast-axis collimator to compress the divergence angle of 5° to 15° in the slow-axis direction to below 5mrad.

[0035] The zoom lens group consists of a fixed lens group, a zoom group, and a compensation group. The zoom group uses a negative power lens group and can move along the optical axis to change the magnification of the light beam.

[0036] The compensation group uses a positive power lens group that can move along the optical axis to maintain the focal point position during zooming. The focusing lens uses a spherical lens to converge the collimated parallel beam to the target position.

[0037] In one embodiment, the switchable circuit drive module includes a pulse drive circuit, a continuous drive circuit, and a quasi-continuous drive circuit, as well as intelligent isolation switch groups connected in series with the three drive circuits respectively.

[0038] After receiving control commands, the AI ​​control center parses them and sends them to the safety rule layer for verification. Upon passing the safety rules, the decision result is sent to the switch controller within the switchable circuit drive module. The switch controller then closes the corresponding intelligent disconnect switch, causing the corresponding drive circuit to output current. The intelligent disconnect switch group includes a first intelligent disconnect switch, a second intelligent disconnect switch, and a third intelligent disconnect switch. The second intelligent disconnect switch is connected in series between the continuous drive circuit and the semiconductor laser chip. The third intelligent disconnect switch is connected in series between the quasi-continuous drive circuit and the semiconductor laser chip. The first intelligent disconnect switch is connected in series with the pulse drive circuit to the semiconductor laser chip via a low parasitic inductance broadband line. At any given time, at most one intelligent disconnect switch is in the ON state. The low parasitic inductance broadband line is made using gold wire or aluminum strip bonding technology, with a bonding length of less than 2 mm, a bonding arc height of less than 150 μm, and parasitic inductance controlled below 1 nH.

[0039] The multimodal control module includes a PC control module and a voice control module, used to receive control commands from the operator and send them to the AI ​​control center. The PC control module adopts a three-layer software architecture: the driver layer encapsulates the underlying communication protocols of the hardware devices and provides API interfaces; the algorithm layer handles offline planning and parameter optimization tasks and does not participate in real-time closed-loop control; the application layer is an HMI graphical interface, where the operator can select and call pre-stored process recipes, manually input drive parameters, or import processing graphics. The PC control module encapsulates the above information into EtherCAT command frames and sends them to the AI ​​control center via the EtherCAT bus at a 1ms cycle. The voice control module uses an offline speech recognition scheme. Its microphone array collects ambient sound at a 16kHz sampling rate, enhances the speech in the target direction using beamforming algorithms, normalizes the speech signal to a standard amplitude using automatic gain control, and removes echo interference using an echo cancellation module. The voice control module uses a Keyword Spotting model to detect preset wake words. After detecting the wake word, it enters the command receiving state and collects voice commands within the next 3 seconds. The voice waveform is converted into text through an offline speech recognition engine, and semantic elements are extracted through a natural language understanding module. The parsing results are mapped into an internal command format and sent to the AI ​​control center through the UART interface.

[0040] In one embodiment, the precision drive module includes multiple micro-stepping motors, which are mechanically connected to the fast-axis collimating lens, the slow-axis collimating lens, the zoom lens group, and the focusing lens, respectively. Figure 2As shown, in the driving structure of the fast-axis collimating lens, a one-dimensional linear stepper motor 4 and a miniature linear guide are fixed on the substrate 1. The fast-axis collimating lens is fixed to the lens mount 7. The lens mount 7 is fixedly connected to the two-dimensional deflection stage 6, which is fixedly connected to the slider 5. The lead screw nut 2 is connected to the slide rail 3. The lead screw is driven by a flexible coupling to push the slider 5 to move along the optical axis, with a stroke of ±0.5mm and a resolution of 0.1μm. Simultaneously, combined with... Figure 3 The diagram shows the internal structure of the two-dimensional deflection stage. The two-dimensional electric tilting platform 8 integrates two one-dimensional micro stepper motors 5 or piezoelectric actuators, which control the tilt angle of the lens around the X-axis and around the Y-axis, respectively. The angle range is ±2° and the resolution is 0.001°.

[0041] The drive structure of the slow-axis collimator is the same as that of the fast-axis collimator, with a stroke of ±1 mm and a resolution of 0.5 μm.

[0042] Each lens in the zoom lens group is mounted on a different lens mount, and each lens mount is mounted on a different slider. Each slider is nested on a concentric guide rail to form a sleeve structure. Each movable lens is equipped with an independent micro stepper motor and lead screw. The travel of the zoom group and the compensation group is ±5mm, and the resolution is 1μm.

[0043] In the focusing lens's drive structure, a one-dimensional linear stepper motor and a miniature linear guide are fixed to the substrate. The focusing lens is fixed to the lens mount, which is fixedly connected to a two-dimensional electric linear stage. The two-dimensional electric linear stage is fixedly connected to the guide slider. A lead screw pushes the guide rail to move along the optical axis, with a stroke of ±5mm and a resolution of 1μm. The miniature stepper motor or piezoelectric actuator integrated inside the two-dimensional electric linear stage provides lateral movement perpendicular to the optical axis, with a stroke of ±10mm. All micro-stepper motors are equipped with optical or magnetic scales as position feedback sensors, forming a closed-loop control.

[0044] In one embodiment, the AI ​​control center is connected to a temperature sensing unit, a spot detection unit, a switchable circuit driving module, and a precision driving module, respectively, for receiving control commands from the operator and parsing them into control intentions. The control intentions include at least one of the target driving mode, target spot parameters, and target focal position. The AI ​​control center has three built-in neural network models: a temperature compensation neural network model, a spot quality assessment compensation model, and a driving mode decision model. Specifically: (i) The temperature compensation neural network model takes the temperature signal output by the temperature sensing unit as input and outputs the displacement compensation of the fast-axis collimating mirror and the slow-axis collimating mirror. The model adopts the architecture of temporal convolutional network or gated recurrent unit to capture the dynamic relationship between temperature change and structural thermal expansion, and the back end is connected to a fully connected layer to directly regress the displacement.

[0045] The training process of the model is as follows: In a constant temperature experimental platform, through precise temperature cycling, the absolute positions of the fast-axis collimating lens and the slow-axis collimating lens that achieve the best spot quality are determined at different temperature points using the hill climbing method or the spot optimization program. The temperature sequence and the corresponding displacement compensation amount are used to form training sample pairs, and supervised learning is used for training.

[0046] (II) The spot quality assessment and compensation model uses the spot image acquired by the CMOS camera as input and outputs spot quality assessment parameters and control variables. The spot quality assessment parameters include the spot diameter. d Ellipticity e and energy distribution vector p , where the energy distribution vector p As an auxiliary task output, it is used to enhance the interpretability of the model. Control variables include focusing lens displacement. and variable displacement The model also outputs a confidence score C, which ranges from 0 to 1 and represents the degree of certainty in predicting the focal location.

[0047] The architecture of the model is as follows: The front end uses a lightweight convolutional neural network to extract the spatial features of the light spot, including the light spot contour, centroid and higher-order moments; the target parameters are encoded and concatenated with the image features in the channel dimension, and then fed into a bidirectional long short-term memory network or self-attention module for temporal modeling, processing image information of multiple consecutive frames, and finally the multi-task head outputs the above parameters respectively.

[0048] The training process is as follows: Under different defocus positions and different zoom group positions, the precise optimal focus position and corresponding motor coordinates are obtained by the hill climbing method or fringe projection method. These are then paired with the spot image to form training samples, and training is carried out using an end-to-end supervised learning method.

[0049] (III) The driving mode decision model takes process requirement parameters and spot quality parameters as inputs and outputs a selection instruction for pulse driving circuit, continuous driving circuit, or quasi-continuous driving circuit. The model architecture employs a multilayer perceptron combined with gated residual connections. Input features are first processed by an embedding layer to classify process parameters, then fused with continuous spot quality features, and finally output as a selection result via a softmax layer. The training process of this model is as follows: supervised training is performed based on a process database or expert-annotated process-spot-optimal driving mode triplet. The loss function is cross-entropy loss, and label smoothing is introduced during training to prevent overfitting.

[0050] The three models described above achieve synergy between temperature compensation and optical control through a dual-network parallel and multimodal architecture. Specifically, the network architecture simultaneously performs two types of tasks: one is thermal drift compensation based on temperature and spot feedback control of the fast-axis and slow-axis collimating lenses, called the temperature compensation network; the other is zoom and focus control based on spot features and user commands coordinating the zoom group and compensation group. The front-end convolutional neural network extracts spatial features such as spot contours and energy centroids, then fuses them with time-series data such as temperature and driving modes at the feature layer, and then feeds them into subsequent networks to capture the physical process of thermal-optical coupling.

[0051] Taking the spot quality assessment and compensation model as an example, the composite loss function used in its training process is expressed as follows:

[0052] in, , , To balance the hyperparameters, the initial values ​​after normalization are set as follows: the physical prior constraint term is 1.0 to ensure that the output does not violate physical rules; the smoothing constraint term is 0.1 to achieve a good balance between suppressing jitter and maintaining response speed; and the confidence loss term is set to 0.5 to ensure that the network carefully evaluates the uncertainty, but the focus is still on accurately predicting the position.

[0053] The loss term for data fitting is the mean squared error, and its expression is:

[0054] In the formula, For the sample size, For the first The model prediction value for each sample. For the first The true label value of each sample. This loss term makes the model output approximate the target value in the training samples.

[0055] The physical prior constraint term is expressed as follows:

[0056] in, For the first The position of the ploid group in each sample. For the first The position of the focusing lens for each sample. =0 represents the zoom-focus compensation curve of the optical system. The physical prior constraint term is constructed based on the law of the zoom-focus compensation curve of the optical system, and is used to ensure that the motor compensation amount output by the spot quality assessment compensation model does not violate the laws of optical physics.

[0057] The smoothness constraint term is expressed as follows:

[0058] in, = [ , ] For the first t The output motor displacement command vector. For the first t Step focusing lens displacement Let ||·||² represent the displacement of the zoom group at step t, and ||·||² represent the square norm of the vector. The smoothness constraint term penalizes abrupt changes between adjacent control commands, ensuring smooth and continuous motion trajectories of the focusing lens and zoom group motors, thereby reducing mechanical wear and beam jitter.

[0059] The confidence loss term is expressed as follows:

[0060] in, For the first i The confidence level of the model prediction for each sample. For the first i The confidence training objective is defined as follows:

[0061] in, The focal position predicted by the model. The true focus location, The variance parameter controls the sensitivity of the confidence level to the prediction error. As shown in the equation above, when the prediction error is small, the confidence level approaches 1; the larger the prediction error, the closer the confidence level approaches 0. The confidence loss term makes the network's output confidence level approximate the ideal confidence level, thus enabling the network to learn to assess uncertainty.

[0062] The loss function of the temperature compensation neural network model is the sum of the mean square error loss and the smoothness constraint term. Its structure is similar to the composite loss function of the spot quality assessment compensation model. The difference is that the label of the temperature compensation neural network model is the displacement compensation amount of the fast axis collimator and the slow axis collimator, and it does not include physical prior constraint term and confidence loss term.

[0063] The driving mode decision model uses the cross-entropy loss function during training, and its expression is:

[0064] in, For the first One-hot encoding of the true driving pattern label of each sample This represents the probability distribution predicted by the model. Label smoothing is introduced during training to prevent overfitting.

[0065] The temperature compensation neural network model, the spot quality assessment compensation model, and the driving mode decision model are trained independently and then deployed in the AI ​​control center to run in parallel after training.

[0066] In one embodiment, during autofocus, the spot quality assessment and compensation model also outputs a confidence level C for the predicted focus position through a confidence output head. The AI ​​control center compares C with a preset threshold of 0.85. When C is greater than or equal to the preset threshold, the AI ​​control center directly generates a collaborative control command based on the output of the spot quality assessment and compensation model; when C is lower than the preset threshold, the system determines that the current input is in a sparse region of the training samples and switches to the hill-climbing search algorithm. The hill-climbing search algorithm drives the focusing lens to move in 1μm step increments. With each step, the CMOS camera acquires a spot image, and the AI ​​control center calculates the high-frequency components of the image as a sharpness evaluation function. When the evaluation function reaches its peak, it is determined to be the optimal focus position. The target displacement of the focusing lens in the collaborative control command is updated based on the difference between the optimal focus position and the current position of the focusing lens, and the search result is used as a training sample to update the spot quality assessment and compensation model.

[0067] Furthermore, the AI ​​control center employs a transfer learning strategy. Before leaving the factory, the neural network model is pre-trained based on a general laser dataset to establish a mapping relationship between temperature, light spot, and motor position. After deployment to the current laser, the neural network model is fine-tuned using temperature-light spot-motor position data calibrated by the current laser under a preset temperature cycle. During the fine-tuning phase, the weights of the physical prior constraints are greater than those in the pre-training phase to ensure that the model prioritizes satisfying the constraints of optical physics laws when running on the current device.

[0068] like Figure 4 As shown, based on the same inventive concept, this invention also discloses a control method for an intelligent dual-control tunable semiconductor laser, applicable to any type of intelligent dual-control tunable semiconductor laser.

[0069] S1. Real-time monitoring of the operating temperature of the semiconductor laser chip and output of temperature signal, while real-time acquisition of spot image and extraction of current spot quality parameters; S2. Receive control commands and parse them into control intentions, then extract target parameters from the control intentions; S3. The target parameters are compared with the temperature signal and the current spot quality parameters to calculate the deviation. Based on the deviation, a pre-trained neural network model is used to generate a cooperative control command. The cooperative control command includes drive circuit gating information and optical element displacement information. S4. Execute the following actions synchronously according to the coordinated control instructions: The corresponding drive circuit is controlled to be turned on by the drive circuit selection information, and the corresponding optical element is driven to generate displacement by the optical element displacement information. S5. After step S4 is completed, reacquire the spot image and extract the compensated spot quality parameters. S6. Calculate the residual deviation between the compensated spot quality parameters and the control intention. If the residual deviation exceeds the preset threshold, jump to step S3 to regenerate the collaborative control command and execute steps S4 to S5 again until the residual deviation converges to within the preset threshold.

[0070] The control method of this embodiment will be described in detail below with reference to the specific working process.

[0071] After the laser is powered on, the temperature sensing unit collects temperature signals in real time at a sampling rate of 100Hz and outputs them to the AI ​​control center. At the same time, the CMOS camera in the spot detection unit collects spot images in real time at a frame rate of no less than 30 frames per second. The AI ​​control center uses the Canny edge detection algorithm to extract the spot contour, fit an ellipse, and calculate the current spot diameter, ellipticity, and energy distribution information, which are then output to the AI ​​control center as the current spot quality parameters.

[0072] When the operator inputs control commands through the multimodal control module, the AI ​​control center receives the commands and parses them into control intentions. The AI ​​control center then inputs the control intention, temperature signal, and current spot quality parameters into the corresponding neural network models. The temperature compensation neural network model generates thermal drift compensation displacements for the fast-axis and slow-axis collimating lenses based on the temperature signal and the divergence angle information in the current spot quality parameters. The spot quality assessment compensation model generates target displacements for the zoom lens group and focusing lens based on the spot image and control intention, and outputs the confidence level C through the confidence output head. The drive mode decision model generates drive circuit selection information for the switch group based on process requirement parameters and spot quality parameters.

[0073] The three models are computed in parallel, and the AI ​​control center merges their outputs into a single collaborative control command. This collaborative control command includes drive circuit gating information and the target displacement of each optical element.

[0074] The AI ​​control center synchronously executes the following two actions based on the collaborative control instructions: The drive circuit selection information control switch group connects the corresponding drive circuit, and the target displacement of each optical element controls the precision drive module to drive the corresponding optical element to produce displacement. Specifically, the target displacement of the fast-axis collimating lens and the slow-axis collimating lens is generated based on temperature deviation and divergence angle information in the current spot quality parameters. The target displacement of the zoom lens group and the compensation lens group is generated based on diameter deviation and a pre-stored optical zoom-focus compensation curve. This curve is obtained through pre-shipment calibration, and the AI ​​control center can correct it online based on the actual spot quality parameters during operation. The target displacement of the focusing lens is generated based on position deviation and the optical zoom-focus compensation curve. When zooming and focusing occur simultaneously, the AI ​​control center calls the pre-stored zoom-focus compensation curve, calculates the following displacement of the focusing lens in real time while the zoom group moves, and outputs it synchronously to maintain focus during zooming. Each micro-stepping motor uses position-speed-current three-loop PID control, moving to the target position according to an S-shaped acceleration / deceleration curve, with a closed-loop grating ruler ensuring positioning accuracy.

[0075] After the displacement is executed, the AI ​​control center verifies whether the deviation has been eliminated based on the spot quality parameters re-acquired by the spot detection unit. During closed-loop verification, the CMOS camera acquires spot images using the same acquisition parameters as during normal operation. If the deviation is not eliminated, the AI ​​control center recalculates the deviation, regenerates the collaborative control command, and drives the corresponding optical element to generate displacement again until the deviation converges to within the preset threshold. For temperature compensation, after the thermal drift compensation displacement of the fast-axis collimator and slow-axis collimator is executed, the CMOS camera re-acquires spot images to verify whether the divergence angle has converged to below 5 mrad. If it has not converged, the compensation displacement is repeatedly generated and the corresponding optical element is driven.

[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A smart dual-control tunable semiconductor laser, characterized in that, include: Semiconductor laser chips; A temperature sensing unit is used to monitor the operating temperature of the semiconductor laser chip in real time and output a temperature signal; The spot detection unit is used to acquire the spot image output by the semiconductor laser in real time and extract the current spot quality parameters. An adjustable optical path module, located on the light-emitting side of the semiconductor laser chip, includes multiple optical elements; A switchable circuit drive module includes multiple drive circuits and a switch group for selectively connecting any drive circuit to the semiconductor laser chip; A precision drive module, mechanically connected to the plurality of optical elements, is used to drive the optical elements to generate displacement; The AI ​​control center, with a built-in neural network model, is connected to the temperature sensing unit, the spot detection unit, the switchable circuit drive module, and the precision drive module. It receives control commands from the operator and parses them into control intentions. It extracts target parameters from these intentions, compares them with the temperature signal and the current spot quality parameters to calculate the deviation, and generates a coordinated control command based on the deviation. This coordinated control command includes drive circuit selection information and optical element displacement information. According to the coordinated control command, it synchronously controls the switch group to connect the corresponding drive circuit and controls the precision drive module to drive the corresponding optical element to produce displacement. After the displacement is performed, the deviation is verified by re-collecting the spot quality parameters by the spot detection unit. If the deviation is not eliminated, the steps of generating the cooperative control command and driving the corresponding optical element to generate displacement are repeated.

2. The intelligent dual-control tunable semiconductor laser according to claim 1, characterized in that, The control intent includes at least one of the target drive mode, target spot parameters, and target focal position.

3. The intelligent dual-control tunable semiconductor laser according to claim 1, characterized in that, The plurality of optical elements, along the laser emission direction, include, in sequence, a fast-axis collimating lens, a slow-axis collimating lens, a zoom lens group, and a focusing lens; The zoom lens group includes a zoom lens group and a compensation lens group that can move independently along the optical axis.

4. The intelligent dual-control tunable semiconductor laser according to claim 1, characterized in that, The spot detection unit extracts the spot contour by performing edge detection on the spot image, and calculates the current spot diameter, ellipticity and energy distribution information based on the spot contour as the current spot quality parameters.

5. The intelligent dual-control tunable semiconductor laser according to claim 3, characterized in that, The neural network model generates the thermal drift compensation displacement of the fast-axis collimating lens and / or the slow-axis collimating lens based on the temperature signal and the current spot quality parameters, as part of the optical element displacement information; After the precision drive module completes the displacement, the AI ​​control center verifies whether the divergence angle has converged based on the spot quality parameters re-collected by the spot detection unit. If it has not converged, the thermal drift compensation displacement is repeatedly generated and the corresponding optical element is driven.

6. The intelligent dual-control tunable semiconductor laser according to claim 1, characterized in that, The deviation includes the diameter deviation between the target spot parameters and the current spot diameter, the positional deviation between the target focal position and the current focal position, and the temperature deviation of the temperature signal relative to the reference temperature.

7. The intelligent dual-control tunable semiconductor laser according to claim 3, characterized in that, The neural network model generates a first displacement of the zoom lens group and a second displacement of the compensation lens group based on the diameter deviation between the target spot parameters and the current spot diameter. Using the first displacement and the second displacement as input, it queries a pre-stored optical zoom-focus compensation curve to generate the following displacement of the focusing lens.

8. The intelligent dual-control tunable semiconductor laser according to claim 3, characterized in that, The neural network model includes a temperature compensation neural network model, a spot quality assessment compensation model, and a driving mode decision model. The temperature compensation neural network model takes the temperature signal as input and outputs the thermal drift compensation displacement of the fast-axis collimating mirror and the slow-axis collimating mirror. The spot quality assessment and compensation model takes the spot image as input and outputs the target displacement of the zoom lens group and the focusing lens; The drive mode decision model takes process requirement parameters and spot quality parameters as input and outputs the drive circuit selection information of the switch group. The temperature compensation neural network model, the light spot quality assessment compensation model, and the driving mode decision model are calculated in parallel, and the output results are merged into a single collaborative control command by the AI ​​control center.

9. The intelligent dual-control tunable semiconductor laser according to claim 8, characterized in that, The spot quality assessment and compensation model also includes a confidence output head, which is used to output the confidence level of the target displacement output by the spot quality assessment and compensation model; The AI ​​control center is also used for: When the confidence level is greater than or equal to the preset threshold, the collaborative control command is directly generated from the output of the spot quality assessment and compensation model. When the confidence level is less than the preset threshold, the algorithm switches to hill climbing and drives the focusing lens to scan step by step within a preset range. Each step involves acquiring a spot image through the spot detection unit and calculating a sharpness evaluation value. The position of the focusing lens when the sharpness evaluation value reaches its peak is taken as the optimal focal position. The target displacement of the focusing lens in the collaborative control command is updated based on the difference between the optimal focal position and the current position of the focusing lens. The AI ​​control center also packages the light spot images collected during the hill-climbing search process with the corresponding optimal focal position and current temperature signal as training samples, and feeds back to update the light spot quality assessment and compensation model.

10. A control method for an intelligent dual-controlled tunable semiconductor laser, applied to the intelligent dual-controlled tunable semiconductor laser as described in any one of claims 1-9, characterized in that, include: S1. Real-time monitoring of the operating temperature of the semiconductor laser chip and output of temperature signal, while real-time acquisition of spot image and extraction of current spot quality parameters; S2. Receive control commands and parse them into control intentions, and extract target parameters from the control intentions; S3. The target parameter is compared with the temperature signal and the current spot quality parameter to calculate the deviation, and a cooperative control command is generated based on the deviation using a pre-trained neural network model. The cooperative control command includes drive circuit gating information and optical element displacement information. S4. Execute the following actions synchronously according to the aforementioned coordinated control instructions: The corresponding driving circuit is controlled to be turned on by the driving circuit selection information, and the corresponding optical element is driven to generate displacement by the optical element displacement information. S5. After step S4 is completed, the spot image is re-acquired and the compensated spot quality parameters are extracted. S6. Calculate the residual deviation between the compensated spot quality parameters and the control intention. If the residual deviation exceeds a preset threshold, jump to step S3 to regenerate the cooperative control command and execute steps S4 to S5 again until the residual deviation converges to within the preset threshold.

Images