135 results about "Neural network design" patented technology

The invention discloses a transformer electromagnetic thermal field real-time prediction method based on a physical constraint embedded neural network, and belongs to the technical field of transformer monitoring. The method aims at solving the problems that a traditional finite element method is poor in real-time performance, low in precision and weak in pure data driving model generalization. The method comprises the following steps: selecting a load rate, an environment temperature and a shell convective heat transfer coefficient as key parameters, generating a sample by optimal Latin hypercube sampling, and establishing a three-dimensional electromagnetic-thermal-fluid coupling finite element model to construct a training / testing database; constructing a deep full-connection neural network of which the input is five parameters easy to measure and the output is a winding temperature nephogram, and designing total loss function training containing data / physical loss; and deploying an on-line monitoring system after verification is qualified, and collecting parameters in real time to output a winding temperature cloud picture. The method has the characteristics of high precision, strong generalization and easy deployment, and provides support for intelligent operation and maintenance and digital twinning of the transformer.

The invention relates to an intelligent control method for a high-speed aircraft, and the method comprises the steps: building an aircraft control architecture, carrying out the decision-making of a neural network model, training a decision-making neural network, and carrying out the real-time compensation of a control instruction according to the flight state of the aircraft, so as to achieve the control of a variable-configuration high-speed aircraft. An aircraft controller adopts a model + learning double-layer architecture, a decision neural network is designed to be of a Double Network structure, overload errors of pitching, yawing and rolling channels are considered in a reward function design, training of a decision neural network model is based on a double-delay deterministic strategy gradient algorithm, a training process is optimized by adopting a prior experience playback technology and a target strategy smoothing mechanism, and a decision neural network model is obtained. The control system can achieve real-time intelligent adjustment of control instructions, overload quantities and attitude angles of three channels of the aircraft and corresponding rudder deflection instructions output by the controller at the last moment are obtained from the environment, the intelligent body outputs control instructions of pitching, yawing and rolling channels, and therefore intelligent control over the variable-configuration high-speed aircraft is achieved.

An adaptive impedance-based multi-mobile-robot collaborative transportation control method. Each mobile robot estimates the actual pose and ideal pose of a reference point and the first and second derivatives of the ideal pose by means of two finite-time fully-distributed observers, respectively; and then, on the basis of the estimated poses of the reference point, the pose of an end-effector of a mechanical arm, and closed-chain constraints for collaborative transportation, an ideal trajectory of the end-effector of the mobile robot, and an estimated value of a pose deviation between the end-effector of the mobile robot and the reference point are obtained. An adaptive impedance system of each mobile robot is interconnected with a virtual energy tank, and the energy tank is used to guide the updating of impedance parameters, thereby ensuring the passivity of the entire collaborative adaptive impedance system. To process unknown system dynamics of mobile robots, an asymptotic tracking adaptive neural network controller is designed using a neural network, thereby asymptotically achieving an ideal adaptive impedance relationship. The operational accuracy of multi-robot collaborative transportation systems is improved while ensuring safe collaboration.

The invention belongs to the field of hardware neural network design, and particularly discloses a CMOS-based attractor recurrent neural network system and an implementation method thereof.A neuron calculation module receives an external input high-level signal and generates an initial synaptic high-level signal; integration and activation calculation are carried out based on the current output by the synaptic calculation module, and whether a synaptic high-level signal is generated or not is determined according to an activation state; the synaptic storage module writes the non-zero weight into a synaptic weight SRAM (Static Random Access Memory) array; the synaptic calculation module performs product operation on the synaptic high-level signal and the synaptic weight in the synaptic weight SRAM array, and outputs current; when the STDP updating module is in a neuron training mode, the STDP updating module updates the synaptic weight SRAM array according to the synaptic high-level signal; and the group state module outputs a neuron sequence number corresponding to the maximum activation frequency. A cyclic connection structure of the attractor neural network is formed, and the real-time cognitive behavior requirement of the brain is met.

The invention relates to an instant denoising method for electroencephalogram signals of students in a classroom scene. The method comprises the following steps: based on electroencephalogram acquisition equipment and a Daisy expansion board, acquiring electroencephalogram signal data of a student through a dry electrode device; and carrying out sliding mean filtering on the electroencephalogram signal data to remove baseline drift, eliminating noise by adopting a band-pass filter to obtain preprocessed electroencephalogram signal data, and inputting the data into a lightweight neural network model to carry out channel-by-channel denoising processing to obtain a denoised electroencephalogram signal. The model comprises a multi-scale feature extraction layer, a time sequence modeling module and a residual connection structure. According to the method, through lightweight neural network design such as multi-scale feature extraction, time sequence modeling and a residual connection structure, the denoising effect is improved, the real-time performance and high efficiency of electroencephalogram denoising processing are guaranteed, and an electroencephalogram data basis with a high signal-to-noise ratio is provided for subsequent teaching quality evaluation.

The present invention relates to the automatic adaptation of deep neural networks to data and / or concept changes through a method that: reveals a variety of design principles (e.g., interconnection of learning blocks, network size, etc.) of deep neural networks for a variety of learning tasks (e.g. image and language processing). The method evolves neural networks constrained by discovered design principles; trains and validates neural networks; hosts a production environment where validated neural networks can operate on production data; monitors production data and network performance; reports different signs of obsolescence; and addresses signs of obsolescence by retraining neural networks with recent production data, replacing obsolete deep neural networks with new models designed by neural architecture research, and / or discovering new design principles to refactor the entire structure, which significantly improves the robustness of machine learning operations in production environments.

The invention discloses a topological optimization method based on a physical information neural network (PINN). The topological optimization method comprises the following steps of data preparation, neural network design, loss function definition, dynamic sampling strategy and optimization process implementation. Specifically, the method comprises the following steps: firstly, acquiring the center coordinates of each unit, then inputting the unit coordinates into a sine representation network, outputting a unit density value, calculating displacement through finite element analysis, then calculating a flexibility value according to node displacement so as to obtain loss, and finally, carrying out back propagation on the loss and updating the network weight. The method has the advantages that the deep learning technology is combined with the physical information neural network, so that the structural performance is improved, the dependence on experimental data is reduced, and the optimization efficiency and precision are improved. Particularly, the structure of the neural network is different from that of a conventional neural network, and deep embedding of physical information is considered in the design of the neural network, so that the prediction accuracy and the generalization ability of the network are improved.

The invention relates to a multi-modal medical image intracranial aneurysm segmentation method based on geometric deep learning, and the method comprises the steps: 1, designing a medical diffusion network based on a deep neural network of frequency domain convolution; 2, preprocessing the three-dimensional grid data to realize the manifold of the model and calculate the characteristics of the grid model; and step 3, using a multiple loss function to reduce the influence of the unbalanced data set, and the like. According to the method, the tedious process of manual feature design and multi-stage processing in a traditional segmentation method is avoided; the MedDiffusion Net designed by the method solves the problem of intracranial aneurysm segmentation of medical images of various modalities, and experimental results on MRA, CTA and DSA data show that the method can obtain accurate intracranial aneurysm segmentation without training data of different modalities respectively; and adaptively fusing the features in the neighborhood by learning a diffusion parameter t.

The invention discloses an under-actuated unmanned ship trajectory tracking control method based on preset time error constraint. The method comprises the following steps: acquiring a three-degree-of-freedom kinematics and dynamics model of an actual underactuated unmanned ship; determining a position tracking error and an azimuth angle tracking error between the underactuated unmanned ship and the reference trajectory; designing a preset time virtual control law; setting a longitudinal speed error and a yawing angular speed error of the underactuated unmanned ship; and designing a time-varying gain auxiliary dynamic system, and designing a preset time controller by using an adaptive self-organizing neural network. According to the invention, through the asymmetric preset performance function constraint error, the self-organizing neural network dynamic approximation disturbance and the model uncertainty, the preset time filter suppression differential explosion and the auxiliary dynamic system compensation input saturation, the high-precision trajectory tracking in the preset time is realized; the method has the advantages of controllable convergence time, unconstrained initial error and high calculation efficiency, and is suitable for autonomous navigation of the underactuated unmanned ship in a complex marine environment.

Techniques and apparatuses enabling high accuracy video object detection using reduced system resource requirements (e.g., reduced computational load, shallower neural network designs, etc.) are described. For example, a search domain of an object detection scheme (e.g., a target object class, a target object size, a target object rotation angle, etc.) may be separated into subdomains (e.g., such as subdomains of object classes, subdomains of object sizes, subdomains object rotation angles, etc.). Specialized, subdomain-level object detection / segmentation tasks may then be separated across sequential video frames. As such, different subdomain-level processing techniques (e.g., via specialized neural networks) may be implemented across different frames of a video sequence. Moreover, redundancy information of consecutive video frames may be leveraged, such that specialized object detection tasks combined with visual object tracking across consecutive frames may enable more efficient (e.g., more accurate, less computationally intensive, etc.) full domain object detection and object segmentation schemes.

The invention belongs to the technical field of robot control, and discloses a visual servo intelligent robust control method for a mobile mechanical arm for a complex operation task, which comprises the following steps of: 1, acquiring a working scene image, and calculating a space coordinate of a target feature point in real time through a projection transformation model; step 2, dynamically inhibiting visual measurement noise by adopting adaptive Kalman filtering, generating a feedforward compensation signal through an integral sliding mode observer, and decoupling chassis slippage disturbance; step 3, inputting the pose signal into a radial RBF neural network, designing an RBF gain scheduler, and adjusting an output proportion-integral gain in real time to suppress external time-varying disturbance; 4, adopting a hybrid visual servo mode switching mechanism, and generating a mobile platform control instruction through an integral sliding mode surface in a position servo mode; and 5, designing a three-order composite controller to drive the mechanical arm so as to realize accurate grabbing. According to the method, the grabbing deviation caused by kinematics uncertainty of the mobile platform and dynamic environment disturbance is eliminated.

The invention discloses a marine ranch cruising task-oriented drum sailboat event triggering optimization control method. The method comprises the following steps of: constructing a three-degree-of-freedom nonlinear mathematical model of a drum sailboat; generating a virtual reference path based on the waypoints by using LVS so as to obtain a guidance law and a position error; improving the guidance law by using a limited boundary circle rule to obtain an improved ILVS guidance law, and further obtaining a heading error; designing a virtual control law based on the actor-commentator neural network; a dynamic surface control technology is introduced to filter the virtual control law and define a dynamic error, and a robust neural damping technology is used to carry out robustness processing on a nonlinear term; designing intermediate control input by combining the actor-commentator neural network, the dynamic error subjected to robust processing and the derivative of the filtered virtual control law; designing an integral dynamic event triggering mechanism, and designing a ship path tracking controller by combining an adaptive compensation technology and intermediate control input; according to the invention, the communication load of a guidance system can be effectively reduced, the practicability of the controller in a dynamic marine environment is effectively improved, and the cruise precision of a marine ranch is guaranteed.

The invention discloses a channel state information (CSI) positioning method based on improved piecewise averaging method (IPAA) dimensionality reduction and multi-scale feature extraction under multiple channels. In the data preprocessing stage, statistical features are extracted from obtained CSI measurement values by using an IPAA dimension reduction method, and data dimensions are reduced. Designing a parallel processing structure by using three deep neural networks, namely a 1D convolutional neural network (1DCNN), a long-short-term memory network (LSTM) and a Transform, extracting multi-scale features of the CSI measured value again, fusing the multi-scale features through a feature fusion strategy, and finally, performing offline classification learning on the fused features through a classification network, so as to obtain the multi-scale features of the CSI measured value. Therefore, a high-precision indoor positioning task is realized. The method has remarkable advantages in the aspects of positioning precision and calculation efficiency, and can effectively deal with a complex indoor environment.

A data analytics platform is provided for forecasting future states of commodities and other assets, based on processing of both textual and numerical data sources. The platform includes a multi-layer machine learning-based model that extracts sentiment from textual data in a natural language processing engine, evaluates numerical data in a time-series analysis, and generates an initial forecast for the commodity or asset being analyzed. The platform includes multiple applications of neural networks to develop augmented forecasts from further analysis of relevant information as it is collected. These include commodity-specific neural networks designed to continually develop taxonomies used to process commodity sentiment, and applications of reinforcement learning, symbolic networks, and unsupervised meta learning to improve overall performance and accuracy of the forecasts generated.

The invention provides a gas flow regulation system model correction method based on an NARX and ELM neural network, and the method comprises the steps: S1, constructing a gas flow regulation system mechanism model of a solid ramjet, and selecting to-be-compensated parameters of a steady-state model and a dynamic model; s2, based on an NARX neural network, designing a correction method of a solid ramjet gas flow steady-state regulation model, and based on experimental data, performing steady-state model parameter correction; s3, on the basis of the integrated ELM neural network, designing a correction method of the solid ramjet gas flow dynamic adjustment model, and performing dynamic model parameter correction on the basis of experimental data; and S4, in combination with the steps S1 to S3, digital simulation of the solid ramjet gas flow regulation control model after mixed compensation is carried out, and a gas generator response predicted value is obtained. According to the method, the experimental data can be effectively utilized to carry out iterative correction on the mechanism model, and the workload of accurate modeling is greatly reduced while the model precision is improved.

The invention discloses an optimal neural network differential classifier searching method based on a cryptographic structure. An existing method is improved from the three aspects of design of an ISRN neural network, data set format optimization based on a password structure, self-adaptive evolutionary optimizer construction and candidate difference set search. In the ISRN neural network design part, initial convolution layers with different convolution kernel sizes are combined with SE module adaptive screening, so that the utilization rate of the neural network to feature information is improved; in the data set format optimization part based on the password structure, more features in the differential propagation process can be provided for the data set input format of the password structure; in the self-adaptive evolutionary optimizer construction and candidate difference set search part, the information entropy evaluation is introduced, and the interlace operation and mutation probability are dynamically adjusted in different stages. The global search capability and convergence capability of the evolutionary optimizer are ensured, the search efficiency and convergence speed of the algorithm are improved, and falling into a locally optimal solution is avoided.

The invention relates to the technical field of robot control, in particular to a self-adaptive formation control and obstacle avoidance method for a distributed multi-robot cluster, which comprises the following steps: acquiring state information of neighbor robots, target points and environmental obstacles through wireless communication and sensors, constructing an error function for describing distance and speed deviation based on the information, and determining the position of the robot cluster according to the error function. A formation control and obstacle avoidance strategy with self-adaptive capability is designed, a self-adaptive controller is designed by using an RBF neural network, local obstacle avoidance and speed deviation items are output according to information changes of neighbor robots and obstacles, and the RBF neural network adopts three-dimensional space description, a three-layer structure, three neurons of an input layer and six neurons of a hidden layer. And one neuron of the output layer performs weighted summation on the output of the RBF neural network and the state of the neighbor robot to generate a complete control strategy, so that the robot cluster effectively avoids obstacles while keeping formation, and the adaptive capacity of the system to environment change is remarkably improved.

The invention discloses a piezoelectric micro-positioning platform preset performance control method based on an extended state observer, and the method mainly comprises the steps: firstly, building a piezoelectric micro-positioning platform system model considering input hysteresis and unknown disturbance; secondly, designing an extended state observer based on a radial basis function neural network to estimate an unmeasurable state of the system and lumped disturbance containing input hysteresis; then, a preset performance function is used, so that the overshoot performance of the system is effectively improved; thirdly, in combination with a preset performance function, a first-order sliding mode differentiator and an expansion state observer, providing a virtual control law and an adaptive control law; and finally, a preset performance controller is utilized, a Lyapunov stability theory is combined, proper parameters are selected to ensure that the closed-loop system is kept stable within preset time, and tracking errors are controlled within a set error range.

The invention provides an unmanned ship safety control method based on a bidirectional long-short-term memory neural network, and the method comprises the steps: designing a predictor and constructing a detection and compensation mechanism through employing the bidirectional long-short-term memory neural network when an unmanned ship is subjected to disturbance and mixed attack; based on a hybrid attack compensation mechanism, data driving control and sliding mode control are combined to design a safe and effective elastic control strategy. The effective data-driven sliding mode safety control method is provided for the unmanned ship which has ocean disturbance and suffers from DoS attacks and FDI attacks, the influence of mixed attacks on the unmanned ship is reduced, and the unmanned ship can safely and effectively track the expected trajectory. According to the invention, the data driving technology and the sliding mode technology are combined, compared with the traditional data driving technology, the control algorithm is higher in interpretability, and meanwhile, the controller designed based on the sliding mode technology is better in control effect.

The invention relates to a cloud computing resource load prediction method based on TCN-PatchTST (Trusted Cryptographic Network-PatchTST). Comprising the following steps: collecting CPU load time sequence data to form a training sample set; carrying out preprocessing and feature construction on the training sample, and selecting a memory utilization rate load value which has obvious influence on a CPU load value as an input feature to predict the CPU load value; according to the method, a TCN-PatchTST neural network is constructed, a two-channel feature extraction architecture is designed, local time sequence features are extracted through causal convolution and expansion convolution of TCN, a long-term dependency relationship is captured by using partitioning processing and a channel independence mechanism of PatchTST, and a feature fusion layer is designed to integrate multi-scale information; the TCN-PatchTST neural network is trained; and inputting a test sample into the trained TCN-PatchTST neural network to obtain a CPU load value prediction result of the test sample. According to the method, load prediction is carried out through the historical load time sequence data, the prediction precision of the CPU load value of the cloud computing resources is effectively improved, and technical support is provided for intelligent scheduling of the cloud resources.

The invention relates to a building outer wall crack detection method, which aims at a building outer wall sample image, designs a local feature encoder through a ResNet50 neural network to carry out local feature extraction, designs a global feature encoder through a Transform model to carry out global feature extraction, and carries out fusion sampling on local features and global features in cooperation with a fusion classification subsystem. A TransUNet to-be-trained model is constructed, training is carried out through the sample data, a building crack detection model is obtained, and then crack recognition on the actual building outer wall is executed; according to the design model, low-level features and high-level features are combined through jump connection, the building outer wall crack detection efficiency and accuracy are improved, early warning is carried out for falling-off potential safety hazards caused by cracks of the building outer wall, manpower, material resources and monitoring cost are reduced through the design method, and the falling-off potential safety hazards of the building outer wall are eliminated.

The invention discloses a five-phase permanent magnet synchronous motor harmonic suppression control method based on machine learning, and the method comprises the following steps: S1, constructing a mathematical model of a five-phase permanent magnet synchronous motor drive system in a rotating coordinate system, and designing a five-phase permanent magnet synchronous motor control system based on a deep neural network; and S2, deeply fusing the physical prior information of the motor into a training process to enhance the generalization ability of the model, and proposing a hybrid optimization strategy to improve the fitting precision of the predicted network voltage and the actual working voltage, suppress the fluctuation in the prediction process, and enable the predicted value to better meet the actual operation requirement. The invention overcomes the limitation of the existing dual-plane current control method, and makes full use of the technical advantages of the method.

The invention discloses a watermarking method and device suitable for any resolution image input, and the method comprises the steps: generating a fixed low-resolution watermark residual error through a watermark encoder according to an input carrier image and a watermark message, restoring the watermark residual error to the size of the carrier image, and superposing the watermark residual error with the carrier image to obtain a watermark-containing image; designing a watermark discriminator based on the residual neural network, wherein the watermark discriminator is used for distinguishing the difference between the watermark-containing image and the carrier image; a watermark noise layer is used for simulating various distortion operations suffered by an image under a digital channel, and disturbance is applied to the watermark-containing image processed by a watermark discriminator; and transmitting the watermark-containing image to which the disturbance is applied in a real channel, inputting the watermark-containing image into a robust watermark decoder, and decoding and extracting a watermark message. According to the method and the device, image input of any resolution can be processed, good invisibility and robustness can be kept, and copyright protection and leakage traceability of the image of any resolution can be realized.

The invention provides a push-broom radiometer antenna pattern optimization method based on reinforcement learning. The push-broom radiometer antenna pattern optimization method comprises the steps of S1, initializing an intelligent agent and an environment; s2, initializing network input and network output; s3, selecting an antenna pattern evaluation index, wherein the antenna pattern evaluation index comprises main beam efficiency, main beam width, side lobe height and a difference between the main beam efficiency, the main beam width and the side lobe height and an ideal antenna pattern; s4, constructing a deep neural network: designing the deep neural network into a structure combining a convolutional neural network and a deep neural network; s5, training a deep neural network; and step S6, obtaining an antenna pattern after network optimization. According to the optimization method, the sidelobe level can be lowered, and the main beam efficiency is improved, so that the system detection precision is improved.

The invention discloses a mechanical arm dynamic constraint control method and system oriented to model uncertainty compensation, and the method specifically comprises the steps: building a mathematical model of a mechanical arm system, and introducing a barrier function to convert a state constraint control problem of the system into a stability control problem; designing a neural network estimator based on the multilayer feedforward neural network, and estimating matching and non-matching endogenous disturbances suffered by the system; designing a disturbance observer based on a neural network estimator, and estimating other matching and non-matching disturbances of the system; designing a mechanical arm dynamic constraint intelligent controller and a neural network adaptive law; and selecting an initial value of a neural network weight parameter, the adaptive law matrix and a controller parameter, and carrying out dynamic constraint control on the mechanical arm. According to the method, the uncertainty of the model can be effectively compensated, meanwhile, high-precision trajectory tracking is achieved, it is strictly guaranteed that the system state does not exceed the dynamic constraint range, and the control precision, adaptability and reliability of the mechanical arm system under the complex working condition are improved.

The application discloses a design method and device of a metasurface optical neural network, and the design method comprises the following steps: obtaining an artificial neural network model and a metasurface optical neural network model, wherein the artificial neural network model is a trained model; selecting n intermediate layers in the artificial neural network model in the order from shallow to deep, one-to-one corresponding the n intermediate layers and n metasurface layers in sequence, and calculating an intermediate feature loss according to the similarity between the features output by each layer in the n intermediate layers and the features output by the corresponding metasurface layer; training the metasurface optical neural network model; and determining the process parameters of the metasurface according to the trained metasurface optical neural network model. The design method completely transplants the calculation capacity of the artificial neural network into the metasurface optical neural network, solves the problems of low design efficiency and low precision of the metasurface optical neural network, and is beneficial to the development of the metasurface optical neural network and the deployment in an environment with limited computing power and storage.

The invention discloses a lightweight neural network design and updating method and system based on dynamic combination of network units, belongs to the technical field of computer vision, machine learning and artificial intelligence, and aims to solve the technical problem of how to design dynamic updating and adjustment of a network to adapt to a constantly changing real environment. The adopted technical scheme is as follows: network unit design: a plurality of network units with various structures or functions are created in a modular design mode; constructing a complete neural network on the basis of the network units; constructing the neural network by combining different network units; dynamically adjusting a network structure: dynamically adding, deleting or reconfiguring a network unit, adjusting the depth and width of the network, exploring an optimal network structure by using reinforcement learning, updating network configuration in real time, and improving the performance and adaptability of the network; according to the freezing parameter updating strategy, frozen parameters and unfrozen fine adjustment parameters are intelligently selected according to network performance and task requirements.

This invention belongs to the field of image deblurring technology and discloses a neural network light field image deblurring method based on a multi-head cross-attention mechanism. Specifically, the network is trained on a LFDOF dataset generated by a light field; the network is fine-tuned using feature loss on a DPDD dataset collected by a secondary method to mitigate the difference between the defocus blurs present in the two domains; the features of the two defocus blur datasets, LFDOF and DPDD, are analyzed to develop a new deblurring training strategy based on a multi-head cross-attention mechanism; and an end-to-end network architecture is proposed, equipped with a novel dynamic residual block, to reconstruct a clear image in a coarse-to-fine manner. The end-to-end neural network designed in this invention can effectively eliminate spatially varying defocus blurs, solving the problem that simple secondary shooting cannot achieve pixel-level correspondence between defocused and in-focus image pairs.

The application relates to a personalized privacy risk dynamic measurement method based on a deep neural network, and specifically comprises the following steps: S1, collecting personalized privacy demands of a user, and establishing a personalized user privacy demand vector; S2, according to the personalized user privacy demand vector, extracting a feature matrix in a user picture by using a convolutional neural network, calculating a privacy feature vector from the feature matrix by using the convolutional neural network, multiplying the privacy feature vector with the personalized user privacy demand vector, obtaining a personalized privacy feature vector of the user in the picture, and identifying sensitive privacy information in the picture; and S3, predicting a leakage risk of the personalized privacy feature vector. The application is designed based on a deep neural network, solves the problem of privacy risk assessment from the perspective of user shared privacy image historical behavior characteristics, measures the risk size of user privacy information leakage, reduces the user privacy information leakage risk from the source, and enhances the user privacy consciousness.

The invention is suitable for the technical field of optical computing, and provides a knowledge distillation-based on-chip optical neural network design method, which comprises the following steps of: constructing an initial population of an optimization algorithm based on a preset process constraint condition; the plurality of individuals in the initial population are in one-to-one correspondence with the plurality of groups of optical neural network parameters; performing iterative optimization on the initial population by using an optimization algorithm to obtain an optimal optical neural network parameter meeting a preset process constraint condition; constructing an optical neural network based on the optimal optical neural network parameters, and taking the constructed optical neural network as an on-chip optical neural network; the calculation process of the fitness value of the individual in the optimization algorithm comprises the following steps: training the optical neural network corresponding to the individual by using the knowledge distillation framework, and determining the fitness value based on the training loss value and the physical parameter of the optical neural network. According to the invention, the area of the on-chip optical neural network is reduced, the complexity of the on-chip optical neural network is reduced, and the performance of the on-chip optical neural network in processing complex tasks is improved.