77 results about "Trust region" patented technology

A design method of a trust region satellite-borne dual-base interference SAR system is provided, relating to the field of radar measurement. Firstly, error sources of the interference SAR system are classified and analyzed, phase errors caused by image decoherence are calculated according to wave positions, and an analytical expression of baseline-related errors is reserved; a maximum phase error caused by imaging processing and an electronic device are estimated; and a baseline error is estimated. Then, the phase errors and the baseline errors are converted into elevation errors through an elevation fuzzy number, thereby obtaining an analytical expression of the elevation errors and the baseline parameters (length and dip angle). With a relative height measurement accuracy index of the system as an optimization object, a feasible solution interval between the baseline length and the baseline dip angle of the system is solved through a trust region algorithm. A minimum value of the baseline length is corrected according to a flying-around safety distance of a dual-satellite formation. According to the method of the present disclosure, the design of the satellite-borne dual-base interference SAR system can be guided, so that the satellite-borne dual-base interference SAR system is ensured to carry out terrain elevation surveying and mapping tasks with high accuracy and high reliability.

The invention discloses a cold chain transportation path optimization method based on intelligent scheduling, and the method comprises the steps: collecting the environmental data of a transportation path, and constructing a directed graph structure containing the temperature and length attributes of a path segment; defining a cooling capacity consumption function, calculating an estimated cooling capacity consumption value of the path section, and forming an environment diagram containing the refrigeration cost; constructing a state vector and an action space, and designing a path selection strategy network and a state value network; setting a multi-objective reward function including cold consumption, delivery completion, cold chain failure and supply behaviors, performing multi-round strategy training by adopting an improved PPO algorithm, and optimizing and updating network parameters through a trust region strategy; and finally, calling an optimal strategy output path selection action in actual transportation to realize dynamic avoidance and supply insertion regulation and control of the cold chain interruption risk. According to the invention, intelligent, self-adaptive and energy-efficient optimization of cold chain path selection can be realized.

The invention relates to an RC circuit health state assessment method, belongs to the technical field of circuit health management of equipment, and solves the technical problems of circuit health state assessment inaccuracy and high calculation cost caused by an algebraic ring in the prior art. The method comprises the following steps: constructing a multi-physics field coupling degradation model and a block diagram simulation model of the RC circuit device; the simulation module is used for acquiring a block diagram simulation model; converting the simulation module into a directed graph and obtaining an algebraic ring candidate set; determining a solving execution sequence of the simulation modules; establishing an algebraic ring solver according to a trust region broken line method; and obtaining a complete dynamic response result of the RC circuit device multi-physical field coupling degradation model in the whole simulation time by using an algebraic ring solver, and evaluating the health state of the RC circuit. According to the method, the algebraic ring in the degradation model can be accurately identified and efficiently processed, so that the physical authenticity of a degradation mechanism is ensured, and high-precision tracking of RC circuit performance and reliable evaluation of health state indexes are realized.

The invention discloses a ring main unit intelligent regulation and control method based on deep reinforcement learning, and the method comprises the following steps: building an electrical topological graph model through collecting the real-time operation data of a ring main unit and a power distribution network, extracting the characteristics of nodes and branches, and generating an initial state vector; a comparison strategy embedding layer is arranged in front of the strategy network, and state-action representation is optimized through comparison learning; constructing a reinforcement learning decision structure by adopting a trust region strategy optimization algorithm, and outputting a ring main unit control action sequence; before execution, an out-of-limit action is judged and corrected by a control barrier function convoy layer; and closing and opening control is completed according to the corrected action, rewards are recorded, and incremental learning and parameter fine adjustment are performed by using an experience playback buffer area. According to the invention, by fusing the graph neural network and deep reinforcement learning and introducing a comparison strategy embedding layer and a control barrier function escorting layer, safe, adaptive and efficient intelligent regulation and control of the ring main unit in a complex power distribution network are realized.

The invention relates to an overlapped chromatographic peak positioning and splitting method and system based on a numerical stable function. The method comprises the following steps: de-noising chromatographic data and correcting a baseline; calculating a second derivative by using SG filtering, grouping continuous points according to a negative threshold value, and determining an initial position of each peak by using a second-order minimum value; establishing a peak function of bidirectional exponential correction by taking a peak position as an initial value, setting parameters and boundaries such as area, center, width, trailing and the like, constructing a residual error minimization target, and performing robust fitting by adopting a trust region reflection algorithm; and indexes such as suitability, output peak area, center, broadening, separation degree, theoretical plate number and the like are verified by residual errors. According to the method, valley point or prior component information is not needed, the method is insensitive to noise and baseline drift, serious overlapping and trailing / leading edge peaks of multiple components can be effectively processed, convergence is high, the method is insensitive to initial values, parameters can be expanded to different peak shapes, and the method is suitable for qualitative and quantitative analysis of automatic and high-flux chromatographic data.

The invention relates to the technical field of monocular depth estimation, and discloses a monocular self-supervision depth estimation method based on credible region depth controllable generation. Firstly, a monocular image is collected and input into a depth estimation module for multi-scale feature extraction, an initial depth map of a whole scene is generated, and relative camera attitude transformation between adjacent frames is estimated through an attitude estimation module; identifying a credible area and an uncredible area according to the obtained initial depth map and relative camera attitude transformation; depth estimation of a dynamic region is guided through a pseudo depth and contrast learning method; then, constructing a multi-task joint loss function, and optimizing depth estimation of the untrusted region by utilizing depth information of the trusted region through a comparison loss mechanism; and iteratively optimizing network parameters of the depth estimation module and the attitude estimation module, and outputting a final optimized depth map. According to the method, the depth estimation precision and robustness of the model in a complex dynamic environment are improved through a credible region depth controllable generation strategy.

The invention relates to the technical field of thermal power generating unit operation optimization, and particularly discloses a thermal power generating unit operation optimization method, device and equipment and a storage medium, and the method comprises the steps: determining a target optimization function of a thermal power generating unit and a target constraint condition corresponding to the target optimization function; obtaining a nonlinear mapping function, and converting the target optimization function into representation of the nonlinear mapping function to obtain a target mapping function; determining a current initial operation parameter of the target mapping function in the target constraint condition based on a simulated annealing algorithm; iteratively optimizing the current initial operation parameter by adopting a trust region inner point method to obtain a current target optimization parameter of the target mapping function; updating the current operation parameters of the thermal power generating unit based on the current target optimization parameters to obtain current target operation parameters; and controlling the thermal power generating unit to operate based on the current target operation parameters. According to the scheme, the operation accuracy of the thermal power generating unit is improved by optimizing the operation data of the thermal power generating unit.

The invention discloses an intelligent agricultural instruction recognition method based on deep semantic analysis, and the method comprises the following steps: S1, receiving and preprocessing an agricultural task natural language instruction, and generating text data; s2, constructing a deep semantic analysis model, and generating a structured semantic feature vector; s3, collecting real-time data in a target agricultural environment, and fusing the real-time data with the structured semantic feature vector to construct a state information vector; s4, establishing a strategy function and a value function by adopting a trust region strategy optimization algorithm, and optimizing an agricultural operation strategy; s5, generating a control instruction of the agricultural equipment according to the optimized operation strategy, and driving the equipment to execute operation; and S6, collecting job feedback data, analyzing and updating a state information vector, and adjusting the deep semantic analysis model. The invention realizes an agricultural natural language instruction recognition and execution method fusing deep semantic analysis and strategy optimization, and is used for improving the intelligent processing capability and execution efficiency of agricultural operation.

The invention provides a generator set scheduling method and system in an electricity-carbon market, and the method comprises the steps: obtaining a market response model, a current decision, and a current adjustment range based on the data of the electricity-carbon market and a historical optimal reference point; based on the current decision and the current adjustment range, executing an accelerated solution step to obtain an optimal decision; based on the optimal decision, scheduling the target generator set to obtain a clearing result; the acceleration solving step comprises the following steps: based on a market response model, obtaining trust region sub-problems in a current decision and a current adjustment range, and then carrying out dimensionality reduction solving to obtain a decision adjustment step length; obtaining an update decision amount and an update adjustment range based on the decision adjustment step length; if the updating decision gradient vector is greater than the preset convergence threshold value, executing an accelerated solution step based on the updating decision and the updating adjustment range; otherwise, the accelerated solution is terminated, and the optimal decision is obtained. The scheduling efficiency of the generator set can be improved.

The invention provides a mining area continuous deformation monitoring method and system based on an InSAR technology, and belongs to the technical field of mining area surface deformation monitoring, and the method comprises the steps: connecting a discontinuous InSAR deformation monitoring result based on a mining area single-point settlement rule, and achieving the mining area deformation monitoring in an SAR data coverage period; a point-by-point mining area deformation connection model is constructed based on a function conforming to a mining area single-point deformation evolution law (following an S-type growth mode), namely a Weibull function. Meanwhile, adopting GA-PSO calculation to construct model initial parameters, and adjusting the model initial parameters according to the point position residual error and the model parameter distribution rule; and setting upper and lower limits according to initial parameters of the model, and solving final model parameters by adopting a least square method based on a trust region reflection algorithm, so as to obtain the deformation of the mining area in the monitoring period. The space-time filling of discontinuous InSAR deformation monitoring is realized, and the detection accuracy is improved.

The invention relates to the technical field of meteorological data assimilation, discloses an assimilation method suitable for a nonlinear problem, and aims to solve the problem that an incremental variation assimilation method has no trust region constraint and may not converge in practical application. A scheme for solving convergence and calculation efficiency of non-linear cloud radiance assimilation in theoretical and practical application is discussed. According to the method, the comparison of the total variation assimilation method and the increment variation assimilation method is realized under the same framework, the comparison of the multi-grid method and the multi-external circulation method is realized under the same framework, the problem that direct comparison is difficult to realize in different platforms due to complex parameters in different methods is solved, and the system operation complexity is greatly simplified through unified model configuration. Experiments show that by adopting the linear search total variation assimilation method under the multi-grid framework, the accuracy of an analysis field is stably improved, and the minimization process of a nonlinear total variation objective function is accelerated.

The application is suitable for the technical field of artificial intelligence security, and provides a jailbreaking security evaluation method and system based on dictionary guided sparse logic editing, which comprises the following steps: first, generating a static intensity vector; then, intercepting an original logit vector corresponding to each autoregressive decoding time step of a target large language model, and generating a dynamic gating signal based on the original logit vector; meanwhile, constructing a preliminary sparse intervention vector from the static intensity vector; and then, generating a modified adversarial logit vector based on the preliminary sparse intervention vector, so as to obtain response content for evaluating security. Through the application of extremely sparse local logic editing based on temperature scaling and trust region constraints in the vocabulary space, the high computing power overhead of traditional gradient attacks and the auxiliary model memory overhead of traditional logic intervention methods can be completely eliminated, and efficient, smooth, interpretable and robust large model jailbreaking security evaluation can be realized.

The present application relates to the field of artificial intelligence, and relates to a model trust region-based copyright protection method for smart grid deep learning models. The method comprises: acquiring a smart grid deep model set; for each model in the smart grid deep model set, searching an exclusive dataset for feature data samples having a model prediction value approximate to a model discrimination boundary, so as to obtain a trust region feature point set; performing dimensionality reduction on gradient vectors of the trust region feature point set on the model discrimination boundary according to a linear discriminant analysis method, so as to obtain perturbation vectors of the trust region feature point set; on the basis of predicted label changes of each model before and after the trust region feature point set is combined with the perturbation vectors, generating a model feature identifier set corresponding to the smart grid deep model set; and, on the basis of the model feature identifier set, training a copyright detection model to be trained, so as to obtain a pre-trained copyright detection model.

The present application belongs to the technical field of intelligent transportation system, and relates to a signal-vehicle cooperative optimization method based on MPC and trust region Bayesian optimization. Firstly, traffic flow data collection and road network simulation modeling are performed, and initial path allocation of vehicles is performed, then real-time speed control is performed on intelligent connected vehicles entering the road network, and the best driving speed is calculated by using a Bayesian optimization algorithm based on a trust region according to signal light information, distance from the intersection and driving conditions of surrounding vehicles. Then, the road impedance is updated, and the best driving path of the intelligent connected vehicle is recalculated. Then, the future traffic state is predicted by using model predictive control, and the best green light time of each intersection in different phases is calculated by using trust region Bayesian optimization. The best green light time is imported into the simulation platform, and vehicle path optimization and speed control are performed again, and the cycle feedback is performed. Finally, it is judged whether the vehicle leaves the road network or reaches the simulation end time, and the vehicle control and signal optimization are ended.

The invention relates to the technical field of automatic driving, in particular to a rapid solution optimization method and system for occupying grid trajectory planning, which combines occupying grid downsampling and a spatial index structure of a low-resolution occupying grid, initial trajectory planning based on a fitted curve and local fine path planning based on a trust domain. Through triple optimization and a synergistic effect, the solution speed in actual measurement is increased by 8-10 times, so that real-time planning based on occupied grids becomes possible. The basis function based on dynamics ensures that the generated trajectory itself conforms to the vehicle kinematics, and the quality and the performability of the trajectory are improved. The sparsity perception and the hierarchical optimization strategy make full use of the structural features of the problem, unnecessary calculation is avoided, and accurate putting of calculation resources is achieved.

This invention provides a hybrid signature method and system based on quantum key distribution and PUF, belonging to the field of information security technology. The method includes: generating a quantum key through a quantum key distribution protocol and generating a derived key seed based on a physically non-cloning function, while simultaneously configuring a trusted region for location and timing; acquiring location and timing information and generating location and timing data; generating a dynamic root key and deriving classical signature key pairs and post-quantum signature key pairs; preprocessing the message to be signed and the location and timing data, matching hierarchical security policies, and selecting the corresponding signature method to generate a signature data packet; during the signature verification process, verifying the location and timing data for signature validity and legality, and verifying the signature result using the corresponding verification method selected according to the hierarchical security policy. This invention achieves dynamic generation of signature keys and hierarchical security policy scheduling, balancing digital signatures in terms of quantum attack resistance, device binding capability, and computational efficiency.

A system performs operations of a neural network agent and a circuit simulator for simulating circuit sizing. The system receives input indicative of a specification and design parameters of a simulated circuit. The system iteratively searches a design space until a circuit size is found that satisfies the specification and the design parameters. In each iteration, the neural network agent computes measurement estimates for random samples generated in a trust region, which is a portion of the design space. Based on the measurement estimates, the system identifies a candidate size corresponding to an optimized value indicator. The circuit simulator receives the candidate size and generates a simulation measurement value. The system computes updates to weights of the neural network agent and the trust region based at least in part on the simulation measurement value for a next iteration.

This invention relates to the field of aviation emergency rescue auxiliary decision analysis, and particularly to a classification and grading method for aviation emergency rescue scenarios. The method first acquires historical case information and designs prompt templates, then uses a large language model to parse unstructured text to obtain structured scenario element data. Next, it constructs a hierarchical structure model for classifying and grading forest fire scenarios, and uses the analytic hierarchy process (AHP) to calculate element weights. Subsequently, it constructs a scoring mapping model, sets monotonicity constraints, and aims to maximize the Pearson correlation between scenario scores and the scale of rescue force deployment, using a trust region algorithm to iteratively solve the model parameters. Finally, it calculates the total scenario score, performs clustering operations based on the deployment scale, and determines the grading threshold based on the cluster centers and dispersion. This invention achieves in-depth utilization of unstructured rescue information, enabling scientific quantification and accurate grading of complex scenarios such as forest fires, effectively improving the scientific nature and response efficiency of aviation emergency rescue decision-making.

The invention discloses a manifold time warping method for a robot motion pose sequence, and mainly relates to the field of teaching learning of robot motion. Comprising the following steps: S1, collecting and importing data; s2, establishing a primary function sequence and a structured path, and constructing constraints; s3, respectively calculating the variances of all position items and the variances of the attitude items, and obtaining the weight ratio of the variances and the attitude items through a variance normalization thought; s4, calculating a center mean value of all current to-be-structured tracks, and calculating a loss function; and S5, solving the regular primary function weight by using a trust region interior point method, and obtaining a final regular path. The method has the beneficial effects that the contributions of the position and attitude characteristics to the loss function can be considered when the position and attitude characteristics are in different change ranges, and high-quality regularity is realized; the time collapse phenomenon in the traditional time warping is avoided, so that the warping result is more uniformly and reasonably distributed on the time axis; the analysis gradient scheme is higher in precision, faster in speed and more reliable.

The application discloses a trajectory optimization method for an electric unmanned aerial vehicle based on fuzzy neural network sequence convex optimization, and belongs to the unmanned aerial vehicle field.The application realizes the method as follows: a residual energy equation is introduced on a particle dynamics equation of the unmanned aerial vehicle, a state equation of a hybrid energy system is constructed, the state equation and an obstacle avoidance constraint are convexed in a trust region range, and a convex optimization model of a hybrid electric unmanned aerial vehicle flight trajectory problem is constructed; in view of a trust region size adjustment problem, a fuzzy neural network is designed according to a constraint violation degree and an objective function increment, the trust region is adaptively adjusted through the fuzzy neural network, the optimality of a sequence convex optimization method is improved, and the convergence speed of the sequence convex optimization method is accelerated; the hybrid electric unmanned aerial vehicle is iteratively solved through the convex optimization method, and a low energy consumption flight trajectory of the hybrid electric unmanned aerial vehicle is obtained. The unmanned aerial vehicle flies according to the obtained trajectory, solar cells can be maximally utilized, energy consumption of the hybrid energy system is reduced, and flight endurance of the unmanned aerial vehicle is increased.

A non-transitory computer-readable storage medium storing a quantization program that causes at least one computer to execute a process, the process includes calculating, for all layers of a neural network, differences each between a trust region radius threshold and a quantization error of a first bit width that is narrower by one step than a second bit width, calculating, based on the differences, each scaling coefficient for all the layers, updating a trust region radius by using a smallest value among the scaling coefficients, and quantizing a parameter of the neural network by a third bit width set based on the trust region.

The application discloses a water pump device characteristic curve correction method based on clustering analysis and a trust region algorithm, and comprises the following steps: clustering steady-state working condition data of a water pump device by using a K-means clustering method, a large number of similar working conditions are combined through iteration, and typical working condition points of the water pump device are screened out; based on the typical working condition points of the water pump device, an original characteristic curve of the water pump device is obtained by using a moving least square method or cubic spline or polynomial fitting; a three-dimensional affine transformation is realized by introducing parameter modification to the original characteristic curve of the water pump device, and the characteristic curve correction of the water pump device is realized based on a Dogbox algorithm in the trust region algorithm. The method has the advantages that: the method does not need to carry out model tests or field tests, but is corrected according to basic operation monitoring data of general large axial flow / mixed flow pump stations, is relatively strong in universality, can still maintain a certain accuracy in a working condition interval with insufficient data, and can provide scientific basis and technical support for actual scheduling of the pump station.

The invention provides a signal-to-noise ratio estimation method and system based on a horizontal visual graph quadratic form, and belongs to the technical field of signal processing. Firstly, a power spectrum of an observation signal is calculated; then carrying out grouping summation on the power spectrums to obtain a grouping summation sequence of the power spectrums; calculating an autocorrelation module value square sequence of the grouping summation sequence, and normalizing and quantifying the autocorrelation module value square sequence; performing horizontal visual graph conversion on the quantized sequence to obtain a graph unsigned Laplace matrix; taking the quantization sequence as a graph signal, and calculating a graph quadratic form mean value under each signal-to-noise ratio based on a graph unsigned Laplace matrix; performing nonlinear fitting on the relationship between the signal-to-noise ratio and the quadratic value of the average graph by using a trust region fitting algorithm to obtain a signal-to-noise ratio estimation expression; and estimating the signal-to-noise ratio of the observation signal diagram according to the signal-to-noise ratio estimation expression and the obtained quadratic form value of the observation signal diagram. Under the condition that training samples are not increased, the estimation performance of the algorithm under the conditions of low signal-to-noise ratio and fading channels is effectively improved.

To reduce a time required for reading encrypted data to a trusted area and decrypting the data in secret information processing to increase the speed of the secret information processing .SOLUTION: A computing system comprises a processor and a memory. The processor has an execution environment for setting a trusted area in the memory, and the execution environment has a plurality of isolated execution processes which are logically isolated. The computing system executes preparatory processing which includes a plurality of pieces of block processing having an execution order determined and receives an execution request of processing on encrypted data and reads out the encrypted data for use in block processing to the trusted area to decrypt the encrypted data, and then executes the block processing using the trusted area and executes, in parallel, the block processing and preparatory processing for block processing later than that block processing in the execution order.SELECTED DRAWING: Figure 7

The invention discloses an advanced material structure field effect transistor model parameter extraction system and a method thereof, relates to the technical field of integrated circuit device modeling, and solves the problems that in the prior art, the precision level cannot fully meet the actual requirement, and the method implementation and circuit simulation cannot be effectively integrated. Comprising a data reading and processing module, a circuit simulation module and a parameter extraction and optimization module, the data reading and processing module is configured to read the preprocessed experimental data, and process and store the drain voltage, the gate voltage and the drain current based on the preprocessed experimental data; the circuit simulation module is configured to perform simulation according to a preset circuit netlist file, and calculate the output current of the full-surrounding gate field effect transistor model based on the input drain voltage and gate voltage; and the parameter extraction and optimization module is configured to extract model parameters according to the output current and the drain current based on a trust region reflection algorithm so as to obtain the model parameters of the field effect transistor meeting the requirements.

This application discloses a sequential convex optimization cooperative trajectory design method, device, medium, and product based on Transformer prediction, relating to the field of aerospace mid-course computational guidance technology. The method includes: inputting state information into a Transformer target trajectory predictor to obtain control action values; integrating the control action values ​​based on dynamic equations to obtain the state information at the next moment; repeating the process to obtain a time-series predicted state matrix; inputting the state information and the time-series predicted state matrix into a deep neural network model to obtain the predicted intercept point and predicted remaining flight time; constructing a trajectory optimization model based on the predicted intercept point and predicted remaining flight time; and solving the trajectory optimization model using an adaptive dynamic adjustment method based on trust region and proximity terms to obtain the trajectory optimization result. This application can achieve accurate prediction and rapid trajectory optimization under limited resources.

The invention belongs to the field of traffic signal control, and relates to a traffic signal optimization method based on model predictive control and trust region Bayesian optimization. The method comprises the following steps: firstly, based on intersection video data, using a target detection and target tracking algorithm to realize automatic collection of traffic flows in different turning directions; then designing a road section flow and intersection steering ratio estimation method based on a traffic flow theory based on three kinds of heterogeneous data including radar data, video data and map data; and finally, establishing a microscopic traffic simulation model based on a traffic state reconstruction result, designing a traffic signal optimization combination framework based on centralized model predictive control and trust region Bayesian optimization, and solving a real-time optimal regional traffic signal timing scheme according to the dynamic change of traffic flow. The vehicle average delay in the network can be effectively reduced, and high efficiency and expandability are achieved when the problem of large-scale signal optimization is solved.

The application discloses a hydraulic pipe network simulation system, and relates to the technical field of pipe network simulation, which comprises the following steps: establishing a connection relation graph of nodes and pipe sections; determining a peak array, and deriving an augmented incidence matrix, a basic incidence matrix, a basic loop matrix, a tree branch matrix, a residual branch matrix, a tree branch vector and a residual branch pipe section vector; inputting corresponding parameters; adopting a basic loop method optimized by a magnetotactic bacteria method to perform first layer iteration, and calculating a residual branch pipe section flow increment iteration step; adopting an LM algorithm, and combining with the magnetotactic bacteria method to perform second layer iteration and calculate an end point flow iteration step; adopting an LM algorithm modified by a trust region, and combining with the magnetotactic bacteria method to perform third layer iteration and calculate a controlled variable iteration step; and calculating and outputting hydraulic properties of each node and pipe section. The application solves the problem that the existing iteration algorithm causes singular values due to too many control points or numerical solutions cannot converge due to improper initial value setting.

The application relates to the technical field of electronic signal processing, and discloses an anti-interference design method based on multi-element mixed electromagnetic interference, which establishes a mixed model of a useful signal static sparse dictionary and a parameterized atomic generating function of an interference signal; initial estimation of interference parameters is obtained by utilizing discrete anchor point dictionary matching; a joint optimization objective function containing a data fidelity term and a sparse constraint term is constructed; initial estimation values are used for initialization, and alternating iteration solving is executed based on a trust region constraint; useful signal sparse coefficients, interference parameter sets and amplitudes are alternately updated by minimizing the objective function, wherein the interference parameters are updated in a multi-element continuous parameter space containing time delay and frequency shift based on a trust region method to adaptively match waveform deformation; and finally, the useful signal is reconstructed according to the converged sparse coefficients and the static dictionary and is output. The application solves the problem of mixed interference separation under multi-system coexistence, and improves the parameter estimation precision and the reliability of algorithm convergence.

The present invention discloses a path optimization method for water cleaning equipment based on deep reinforcement learning, comprising the following steps: S1, using sensors to collect water environment data and perform preprocessing; S2, establishing a state space and action space for path optimization, and building a reward mechanism; S3, using a Transformer network to extract features from the preprocessed data; S4, using an improved trust region policy optimization algorithm to train the policy network and the value network, optimizing the update step size of the policy network based on the KL divergence constraint, and optimizing the path planning strategy based on the policy gradient method; S5, using the trained policy network and value network to perform path planning; S6, calculating the energy consumption of the water cleaning equipment and adjusting the operating mode or operation path of the water cleaning equipment. The present invention combines Transformer and the improved trust region policy optimization algorithm to optimize the path planning of water cleaning equipment, and has the advantages of strong environmental adaptability, low energy consumption, high cleaning efficiency and strong operation stability.