Unmanned intelligent inspection equipment collaborative scheduling method and system in photovoltaic power generation scenarios

Abstract

The present invention provides a method and system for collaborative scheduling of unmanned intelligent inspection equipment in a photovoltaic power generation scenario, which relates to the technical field of intelligent inspection of photovoltaic power stations, including: performing task priority classification and time series correlation analysis based on a dynamic adaptive neural network and a sequence predictor, constructing a task dependency graph, and generating candidate task allocation schemes based on a deep reinforcement learning framework, with the goal of minimizing completion time, energy consumption and maximizing coverage, optimizing the candidate schemes by combining a collaborative decision-making system constructed with a multi-objective genetic algorithm and a graph attention mechanism, updating task allocation constraints by using a confidence propagation network, judging abnormal situations by a Bayesian reasoning algorithm, and finally outputting the optimal scheduling scheme, which can effectively improve the inspection efficiency of photovoltaic power stations, reduce energy consumption, and increase inspection coverage.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent inspection of photovoltaic power stations, and in particular to a method and system for collaborative scheduling of unmanned intelligent inspection equipment in a photovoltaic power generation scenario. Background Art

[0002] As the scale of photovoltaic power generation continues to expand, it becomes increasingly important to ensure the stable operation and efficient power generation of photovoltaic power stations. As an efficient and reliable means of operation and maintenance, unmanned intelligent inspection technology is widely used in the daily inspection of photovoltaic power stations.

[0003] Traditional inspection methods mainly rely on manual work, which is inefficient and easily affected by human factors. However, automated inspection based on drones can effectively improve inspection efficiency, reduce operation and maintenance costs, and improve the accuracy of inspection data. Existing drone inspection technology usually adopts pre-planned fixed routes or simple rule-based scheduling methods, which are difficult to adapt to the complex and changeable environment of photovoltaic power stations.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for coordinated scheduling of unmanned intelligent inspection equipment in a photovoltaic power generation scenario, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for coordinated scheduling of unmanned intelligent inspection equipment in a photovoltaic power generation scenario, comprising:

[0007] Obtain inspection tasks in photovoltaic power generation scenarios and convert them into inspection task feature vectors, build a task evaluation system based on a dynamic adaptive neural network containing multiple resonance layers, set different warning parameters and competitive learning mechanisms for each resonance layer, add the inspection task feature vectors to the task evaluation system, obtain task priority classification results, add the task priority classification results to a pre-set sequence predictor, model the inspection task feature vectors at different time scales to output task time series association sequences, build a recursive attention network based on the task time series association sequence to perform multi-task association analysis, establish a task dependency graph in combination with position encoding, and obtain a task dependency matrix based on the task dependency graph;

[0008] Based on the task priority classification results, task timing association sequence and task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved in combination with an improved variable neighborhood search algorithm to obtain a candidate task allocation scheme;

[0009] Based on the candidate task allocation scheme, a state space is constructed for each inspection device. Each inspection device is regarded as an independent intelligent agent and equipped with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool. The candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on the graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, decision outputs are generated and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing, and the candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

[0010] In an optional embodiment,

[0011] The inspection tasks in the photovoltaic power generation scenario are obtained and converted into inspection task feature vectors. A task evaluation system is constructed based on a dynamic adaptive neural network containing multiple resonance layers. Different warning parameters and competitive learning mechanisms are set for each resonance layer. The inspection task feature vectors are added to the task evaluation system to obtain task priority classification results. The task priority classification results are added to a pre-set sequence predictor. By modeling the inspection task feature vectors at different time scales, the output task time series association sequences include:

[0012] Obtain an inspection task, based on the inspection task, obtain the task type feature through one-hot encoding, map the task location coordinates to spatial features through a geographic information system to obtain the task location feature, quantify the task urgency into a continuous value to obtain the task urgency feature, determine the equipment feature required for the task through a multi-hot encoding method, determine the task completion time feature through a normalization algorithm, and splice the task type feature, the task location feature, the task urgency feature, the equipment feature required for the task, and the task completion time feature to form the inspection task feature vector;

[0013] Constructing a dynamic adaptive neural network, wherein the dynamic adaptive neural network comprises a plurality of resonance layers, each of the resonance layers comprises a plurality of neurons connected by synapses, information is transmitted between the resonance layers by feedforward connections and feedback connections, the neurons in the resonance layers have a competitive learning mechanism, when one neuron is activated, the activities of other neurons are inhibited, the neurons in the resonance layer are activated according to input features and warning parameters, the warning parameters control the activation threshold of the neurons, and the neurons in the resonance layer control the update speed of weights by the learning rate;

[0014] The inspection task feature vector is added to the dynamic adaptive neural network, and the feature transformation is performed step by step through the neurons of the resonance layer, and the task priority classification result is obtained by combining the forward propagation algorithm;

[0015] The task priority classification result is added to a pre-set sequence predictor constructed based on a long short-term memory network. The sequence predictor updates the network state according to the current input and the state of the previous time step, models the inspection task feature vector at different time scales, and outputs a task timing association sequence containing the association information between task features and historical task features.

[0016] In an optional embodiment,

[0017] Based on the task temporal association sequence, a recursive attention network is constructed to perform multi-task association analysis, and a task dependency graph is established in combination with position encoding. The task dependency matrix obtained based on the task dependency graph includes:

[0018] A recursive attention network is constructed to process the task temporal association sequence, wherein the recursive attention network includes an encoder and a decoder, the task temporal association sequence is added to the recursive attention network, the encoder encodes the task temporal information through a position encoding mechanism, and maps the task temporal association sequence to a semantic feature space through a multi-head attention mechanism, and the decoder adopts a multi-head attention mechanism to perform weighted aggregation on the semantic feature space to generate a task dependency graph, and the task dependency graph represents the dependency relationship between tasks in the form of a directed acyclic graph;

[0019] Construct a task dependency matrix according to the total number of inspection tasks, the number of rows and columns of the task dependency matrix are both the total number of inspection tasks, establish a correspondence between the rows and columns of the task dependency matrix and the identifiers of the inspection tasks, set the initial values ​​of all elements in the task dependency matrix to the second preset value, traverse all edges in the task dependency graph, obtain the source task node, the target task node and the corresponding dependency weight value of each edge, determine the corresponding row index and column index in the task dependency matrix based on the identifiers of the source task node and the target task node, and set the element value corresponding to the row index and the column index position in the task dependency matrix to the first preset value;

[0020] Performing validity verification on the task dependency matrix, judging whether there is a cyclic dependency relationship based on whether the element values ​​on the diagonal of the task dependency matrix all satisfy the second preset value and whether the task dependency matrix satisfies the directed acyclic graph characteristic, if all are satisfied, it is considered that the validity verification has passed, otherwise it is considered that the validity verification has failed, and the first preset value and the second preset value are modified;

[0021] The task dependency matrix is ​​optimized for storage based on compressed storage, wherein the location information of the storage element value being a first preset value and its corresponding row and column index information are stored, a mapping relationship table between the task dependency matrix and the inspection task identifier is generated, and the task dependency matrix is ​​output.

[0022] In an optional embodiment,

[0023] Based on the task priority classification results, task timing association sequence and task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved by combining the improved variable neighborhood search algorithm, and the candidate task allocation schemes obtained include:

[0024] Based on the task priority classification results, the task timing association sequence and the task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network;

[0025] Input the inspection task feature vector into the first strategy branch, the second strategy branch and the third strategy branch of the task allocation main network, wherein the first strategy branch uses a first-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a global feature map, the second strategy branch uses a second-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a local feature map, and the third strategy branch uses a third-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a detail feature map;

[0026] The global feature map, the local feature map and the detail feature map are superimposed to obtain a comprehensive feature representation, and the comprehensive feature representation is input into a fully connected layer to generate an initial task allocation plan including a task sequence and a time window constraint;

[0027] Calculate the spatial distance between tasks based on the position coordinates of adjacent tasks in the task sequence, calculate the time interval between tasks based on the time window constraint, and combine the spatial distance and the time interval to construct a task distance matrix;

[0028] An initial temperature parameter is set for the initial task allocation scheme based on the state assessment auxiliary network. According to the initial temperature parameter, a selection is made among an exchange operation, an insertion operation and a reversal operation in combination with a preset operation probability, wherein the exchange operation is used to exchange the positions of two tasks in the task sequence, the insertion operation is used to move a task in the task sequence to another position, and the reversal operation is used to reverse the order of subsequences in the task sequence. After executing the operation, the initial temperature parameter is decayed, and an operation is selected again based on the attenuated temperature. The selection and execution of the operation and the temperature decay are repeated until a preset maximum number of iterations is reached, so as to obtain multiple optimal inspection routes, and generate a candidate task allocation scheme based on the optimal inspection route.

[0029] In an optional embodiment,

[0030] After the operation is performed, the initial temperature parameter is attenuated as shown in the following formula:

[0031] ;

[0032] Among them, T k+1 represents the temperature value at the k+1th iteration, T k represents the temperature value at the kth iteration, k represents the number of iterations, β represents the temperature attenuation adjustment factor, ω represents the periodic fluctuation factor, α represents the temperature attenuation rate control factor, and λ represents the KL divergence adjustment factor. represents the inferred distribution q(z|x k ) and the prior distribution p(z).

[0033] In an optional embodiment,

[0034] Based on the candidate task allocation scheme, a state space is constructed for each inspection device. Each inspection device is used as an independent intelligent agent and equipped with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool. The candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage rate. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on a graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, a decision output is generated, and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme, including:

[0035] Collect inspection equipment status information and construct a three-dimensional state space matrix, map the inspection equipment number to the first dimension, map the location information, power status, task completion progress and other state variable types to the second dimension, and map the time step to the third dimension to form a dynamic representation of the equipment state. Based on the three-dimensional state space matrix, deploy a local decision network and a value network on each inspection device, input the state space dimension to the input layer of the three-layer fully connected neural network, output the action space dimension after hidden layer processing, and use the value network to evaluate the state value;

[0036] According to the decision data generated during the execution of the equipment, an asynchronous experience replay pool is created and the state transition data is stored. The current state, execution action, reward, and next state information are recorded. The sampling probability is assigned to the state transition data according to the task priority. For each inspection device, the mapping relationship between the device number, task number, and time window is encoded into a chromosome sequence. The three optimization objective functions of task completion time, energy consumption, and inspection coverage are calculated according to the encoding results. After the population is initialized, a fast non-dominated sort is performed. The non-dominated individuals are divided into the first layer by counting the dominance relationship between individuals, and the dominated individuals are randomly divided into other layers except the first layer. The crowding degree of the individuals that have completed the layer division is calculated. By sorting individuals at different layers and measuring the target space distance, the distance value of the objective function is accumulated.

[0037] According to the individual level and crowding information obtained, the crossover probability and mutation probability parameters are dynamically adjusted. Based on the adjusted parameters, the parent individuals are selected to perform single-point crossover to generate offspring individuals. The sites of the offspring individuals are randomly selected to perform gene mutations and the task constraint satisfaction is verified in real time. The parent and offspring populations are merged and the sorting and crowding calculations are re-executed. High-quality individuals are selected to form a new population. The iteration is continued until the termination condition is met. The individual with the highest fitness value is selected from the optimal solution set as the improved solution.

[0038] Based on the improved solution, a collaborative decision-making system is constructed according to the graph attention mechanism and the interaction between agents is modeled. Corresponding agent attention weights are assigned to different agents, and decision outputs are generated based on the agent attention weights. A confidence propagation network is constructed based on the decision outputs to model the uncertainty of the decision outputs. Each agent calculates the decision confidence based on local observations and influences the confidence estimates of other agents through network propagation. The residual information is exchanged and updated in combination with the residual message passing method. The residual convergence is monitored in real time and converges to a globally consistent task allocation constraint. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an anomaly. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

[0039] In an optional embodiment,

[0040] Monitor the residual convergence in real time and converge to the globally consistent task allocation constraints, output candidate scheduling solutions and combine the Bayesian reasoning algorithm to determine whether there are any abnormalities. If so, optimize again. Otherwise, output the current candidate scheduling solution as the optimal scheduling solution, including:

[0041] Obtain the real-time status data of each agent, record the residual value between each agent and its neighboring nodes, store the residual data in the residual matrix, calculate the norm of the residual matrix and compare it with the preset convergence threshold. The norm of the residual matrix reflects the consistency of the allocation scheme between agents;

[0042] Monitor the residual norm variation trend in multiple consecutive time slices. When the residual norm is continuously less than a preset threshold, determine that the system has reached a convergence state, and summarize the local constraint information of all agents, which includes device execution capability, time window restriction, and path reachability constraint.

[0043] Based on the local constraints in the converged state, a global constraint set including task time window, path accessibility, and energy consumption is constructed. Constraint conflict detection is used to ensure that there is no conflict between constraints, and candidate scheduling solutions that meet the global constraints are output.

[0044] Obtain historical normal scheduling data, extract scheduling characteristic parameters, the scheduling characteristic parameters include task time distribution, energy consumption distribution, equipment load distribution, task completion rate distribution, establish a multi-dimensional Gaussian mixture model of normal scheduling data, estimate model parameters through expectation maximization algorithm, the model parameters include the mean vector of each Gaussian component, covariance matrix and mixing weight, perform feature extraction on candidate scheduling schemes, and obtain candidate feature vectors of the same dimension as historical data, the candidate feature vectors include task time features, energy consumption features, equipment load features, and task completion rate features;

[0045] The log-likelihood value of the candidate feature vector is calculated based on the established Gaussian mixture model, and the log-likelihood value is the logarithmic value of the weighted sum of the conditional probability of the feature vector under each Gaussian component and the mixture weight. An anomaly detection threshold is set, and the anomaly detection threshold is determined based on the likelihood value distribution of historical data. When the log-likelihood value of the candidate scheduling scheme is lower than the threshold, the current candidate scheduling scheme is marked as abnormal. For the candidate scheduling scheme marked as abnormal, the scheduling scheme is regenerated based on non-dominated sorting and adaptive genetic operators. For the candidate scheduling scheme not marked as abnormal, the objective function value is calculated, and the candidate scheduling scheme not marked as abnormal with the highest objective function value is selected as the optimal scheduling scheme for output.

[0046] A second aspect of an embodiment of the present invention provides an unmanned intelligent inspection equipment collaborative dispatching system in a photovoltaic power generation scenario, comprising:

[0047] The first unit is used to obtain the inspection tasks in the photovoltaic power generation scenario and convert them into inspection task feature vectors, build a task evaluation system based on a dynamic adaptive neural network containing multiple resonance layers, set different warning parameters and competitive learning mechanisms for each resonance layer, add the inspection task feature vectors to the task evaluation system, obtain task priority classification results, add the task priority classification results to a pre-set sequence predictor, model the inspection task feature vectors at different time scales to output task time series association sequences, build a recursive attention network based on the task time series association sequence to perform multi-task association analysis, establish a task dependency graph in combination with position encoding, and obtain a task dependency matrix based on the task dependency graph;

[0048] The second unit is used to construct a deep reinforcement learning framework based on the task priority classification results, the task timing association sequence and the task dependency matrix, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved in combination with an improved variable neighborhood search algorithm to obtain a candidate task allocation scheme;

[0049] The third unit is used to construct a state space for each inspection device based on the candidate task allocation scheme, treat each inspection device as an independent intelligent agent and equip it with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool, and the candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage rate. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on a graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, decision outputs are generated, and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing, and the candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

[0050] According to a third aspect of the embodiments of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] A fourth aspect of the embodiments of the present invention is:

[0056] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0057] In the present invention, priority classification and timing correlation analysis are performed through the task evaluation system, and task allocation and path planning are performed in combination with a deep reinforcement learning framework, so that inspection tasks can be effectively arranged, inspection time can be reduced, and inspection efficiency can be improved. The task allocation scheme is optimized by a multi-objective genetic algorithm with the goal of minimizing energy consumption, which can effectively reduce the energy consumption of inspection equipment and extend the service life of equipment. The multi-objective genetic algorithm is optimized with the goal of maximizing coverage, and the task allocation constraints are updated in combination with a collaborative decision-making system and a confidence propagation network, which can improve the inspection coverage, reduce missed inspections, and ensure the safe and stable operation of photovoltaic power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of a method for coordinated dispatching of unmanned intelligent inspection equipment in a photovoltaic power generation scenario according to an embodiment of the present invention;

[0059] Figure 2 It is a structural schematic diagram of an unmanned intelligent inspection equipment collaborative scheduling system in a photovoltaic power generation scenario according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a method for coordinating unmanned intelligent inspection equipment in a photovoltaic power generation scenario according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] S1. Obtain the inspection task in the photovoltaic power generation scenario and convert it into an inspection task feature vector, build a task evaluation system based on a dynamic adaptive neural network containing multiple resonance layers, set different warning parameters and competitive learning mechanisms for each resonance layer, add the inspection task feature vector to the task evaluation system, obtain the task priority classification result, add the task priority classification result to a pre-set sequence predictor, model the inspection task feature vectors at different time scales to output the task time series association sequence, build a recursive attention network based on the task time series association sequence to perform multi-task association analysis, establish a task dependency graph in combination with position encoding, and obtain a task dependency matrix based on the task dependency graph;

[0064] The resonance layer is a layer used in a neural network to capture the interaction of signals or features. It enhances the response to specific patterns by simulating resonance phenomena of different frequencies. It is usually used to process complex input data, such as speech, images, etc., to extract meaningful features. The dynamic adaptive neural network is a neural network model whose structure and parameters can be adjusted in real time according to input data and environmental changes. The alert parameter refers to a parameter used to monitor and control the model training process in a neural network or other learning system. The competitive learning mechanism is a machine learning mechanism in which multiple learning units or models compete on a given task, and only the best performing unit can be updated or optimized. The sequence predictor is a model or algorithm used to predict the future value of sequence data (such as time series, text sequence, etc.). The task dependency graph is a graph model used to represent the dependency relationship between tasks. In the task dependency graph, each task is a node, and the dependency relationship between tasks is represented by directed edges. The task dependency matrix is ​​a matrix representation method used to describe the dependency relationship between tasks.

[0065] In an optional embodiment,

[0066] The inspection tasks in the photovoltaic power generation scenario are obtained and converted into inspection task feature vectors. A task evaluation system is constructed based on a dynamic adaptive neural network containing multiple resonance layers. Different warning parameters and competitive learning mechanisms are set for each resonance layer. The inspection task feature vectors are added to the task evaluation system to obtain task priority classification results. The task priority classification results are added to a pre-set sequence predictor. By modeling the inspection task feature vectors at different time scales, the output task time series association sequences include:

[0067] Obtain an inspection task, based on the inspection task, obtain the task type feature through one-hot encoding, map the task location coordinates to spatial features through a geographic information system to obtain the task location feature, quantify the task urgency into a continuous value to obtain the task urgency feature, determine the equipment feature required for the task through a multi-hot encoding method, determine the task completion time feature through a normalization algorithm, and splice the task type feature, the task location feature, the task urgency feature, the equipment feature required for the task, and the task completion time feature to form the inspection task feature vector;

[0068] Constructing a dynamic adaptive neural network, wherein the dynamic adaptive neural network comprises a plurality of resonance layers, each of the resonance layers comprises a plurality of neurons connected by synapses, information is transmitted between the resonance layers by feedforward connections and feedback connections, the neurons in the resonance layers have a competitive learning mechanism, when one neuron is activated, the activities of other neurons are inhibited, the neurons in the resonance layer are activated according to input features and warning parameters, the warning parameters control the activation threshold of the neurons, and the neurons in the resonance layer control the update speed of weights by the learning rate;

[0069] The inspection task feature vector is added to the dynamic adaptive neural network, and the feature transformation is performed step by step through the neurons of the resonance layer, and the task priority classification result is obtained by combining the forward propagation algorithm;

[0070] The task priority classification result is added to a pre-set sequence predictor constructed based on a long short-term memory network. The sequence predictor updates the network state according to the current input and the state of the previous time step, models the inspection task feature vector at different time scales, and outputs a task timing association sequence containing the association information between task features and historical task features.

[0071] The one-hot encoding is an encoding method commonly used to convert categorical variables into numerical data. Each category is mapped to an independent binary vector, in which only one position is 1 and the rest are 0, which is used to represent different categorical values. The multi-hot encoding is an extended form of one-hot encoding, which is suitable for situations where each instance may belong to multiple categories. In multi-hot encoding, the position of each category can be 1 or 0, allowing a data instance to belong to multiple categories at the same time. It is commonly used in text classification, recommendation systems and other fields. The activation threshold is a value in a neural network used to determine whether a neuron is activated. When the input signal of a neuron reaches or exceeds the activation threshold, the neuron will be activated and transmit a signal, otherwise it will remain in an inactive state.

[0072] Get the inspection tasks of the photovoltaic power station. These tasks may come from various sources such as routine inspection plans, sensor alarms, and manual reports. For example, if the temperature of a photovoltaic component rises abnormally, the system will automatically generate an inspection task; for another example, if the monthly routine inspection plan includes the inspection of all inverters, the system will also generate corresponding inspection tasks.

[0073] Convert the acquired inspection tasks into feature vectors. This process includes extracting multiple key features of the task and converting these features into numerical vector representations. Specifically, features can be extracted from the following aspects:

[0074] Task type feature: Use one-hot encoding to represent the task type. For example, suppose there are three task types: component cleaning, equipment maintenance, and troubleshooting. If the current task is equipment maintenance, the corresponding task type feature vector is [0, 1, 0].

[0075] Task location feature: Use the geographic information system to map the latitude and longitude coordinates of the task location to the pre-divided spatial grid and use one-hot encoding to represent it. For example, if the photovoltaic power station is divided into 100 grids and the task is located in the 57th grid, the corresponding task location feature vector is a vector of length 100, in which only the 57th element is 1 and the rest are 0.

[0076] Task urgency feature: quantify the urgency of the task as a continuous value between 0 and 1. For example, the most urgent fault task is assigned a value of 1, the routine inspection task is assigned a value of 0.1, and other tasks are assigned corresponding values ​​according to their urgency. Assuming that the urgency of a task is medium, it is assigned a value of 0.6.

[0077] Task-required equipment features: Multi-hot encoding is used to represent the equipment required for a task. For example, suppose there are three types of equipment: a wrench, a multimeter, and an infrared thermometer. If a task requires a wrench and a multimeter, the corresponding task-required equipment feature vector is [1, 1, 0].

[0078] Task completion time feature: Use a normalization algorithm to convert the estimated completion time of a task into a value between 0 and 1. For example, assuming the shortest completion time is 1 hour, the longest completion time is 8 hours, and the estimated completion time of a task is 4 hours, then the corresponding task completion time feature is (4-1) / (8-1)=0.43.

[0079] The above five feature vectors are concatenated to form the final inspection task feature vector. For example, an equipment maintenance task is located in the 57th grid, has an urgency of 0.6, requires a wrench and a multimeter, and is expected to be completed in 4 hours. Its corresponding inspection task feature vector can be expressed as [0, 1, 0, …, 1, …, 0.6, 1, 1, 0, 0.43], where the ellipsis represents other elements of the task location feature vector.

[0080] A dynamic adaptive neural network is constructed for task evaluation, which includes multiple resonance layers. Each resonance layer contains multiple neurons, and the neurons are connected through synapses. Information is transmitted between different resonance layers through feedforward connections and feedback connections. Each resonance layer is set with different alert parameters to control the activation threshold of neurons. There is a competitive learning mechanism between neurons in the resonance layer. When a neuron is activated, it inhibits the activity of other neurons. The weights of neurons are updated according to the input features and learning rate. The inspection task feature vector is input into the dynamic adaptive neural network, and the task priority classification result is finally obtained through the step-by-step feature transformation of the resonance layer and the forward propagation algorithm. For example, the tasks are divided into three priorities: high, medium, and low.

[0081] The task priority classification results are input into the sequence predictor built based on the long short-term memory network (LSTM). The sequence predictor updates the network state according to the current input and the state of the previous time step, models the inspection task feature vector at different time scales, and outputs the task time series association sequence containing the association information between task features and historical task features. For example, it predicts the priority change trend of different tasks in the future period of time, as well as the association relationship between different tasks.

[0082] In this embodiment, by prioritizing tasks, urgent and important tasks can be given priority, thereby improving inspection efficiency and reducing resource waste. By analyzing the task timing association sequence, potential failure risks can be discovered and preventive measures can be taken in advance to avoid failures. Based on the task priority and timing association information, the scheduling of inspection personnel and equipment can be optimized to improve resource utilization.

[0083] In an optional embodiment,

[0084] Based on the task temporal association sequence, a recursive attention network is constructed to perform multi-task association analysis, and a task dependency graph is established in combination with position encoding. The task dependency matrix obtained based on the task dependency graph includes:

[0085] A recursive attention network is constructed to process the task temporal association sequence, wherein the recursive attention network includes an encoder and a decoder, the task temporal association sequence is added to the recursive attention network, the encoder encodes the task temporal information through a position encoding mechanism, and maps the task temporal association sequence to a semantic feature space through a multi-head attention mechanism, and the decoder adopts a multi-head attention mechanism to perform weighted aggregation on the semantic feature space to generate a task dependency graph, and the task dependency graph represents the dependency relationship between tasks in the form of a directed acyclic graph;

[0086] Construct a task dependency matrix according to the total number of inspection tasks, the number of rows and columns of the task dependency matrix are both the total number of inspection tasks, establish a correspondence between the rows and columns of the task dependency matrix and the identifiers of the inspection tasks, set the initial values ​​of all elements in the task dependency matrix to the second preset value, traverse all edges in the task dependency graph, obtain the source task node, the target task node and the corresponding dependency weight value of each edge, determine the corresponding row index and column index in the task dependency matrix based on the identifiers of the source task node and the target task node, and set the element value corresponding to the row index and the column index position in the task dependency matrix to the first preset value;

[0087] Performing validity verification on the task dependency matrix, judging whether there is a cyclic dependency relationship based on whether the element values ​​on the diagonal of the task dependency matrix all satisfy the second preset value and whether the task dependency matrix satisfies the directed acyclic graph characteristic, if all are satisfied, it is considered that the validity verification has passed, otherwise it is considered that the validity verification has failed, and the first preset value and the second preset value are modified;

[0088] The task dependency matrix is ​​optimized for storage based on compressed storage, wherein the location information of the storage element value being a first preset value and its corresponding row and column index information are stored, a mapping relationship table between the task dependency matrix and the inspection task identifier is generated, and the task dependency matrix is ​​output.

[0089] The source task node refers to the task node that initially needs to be processed in the task graph or task dependency graph. The validity verification is a process used to check and verify whether the task, model or system meets the expected requirements or specifications. The circular dependency refers to the fact that certain tasks or modules in the task or system are interdependent, forming a closed loop. The inspection task identifier refers to the symbol or code used to uniquely identify the inspection task in the system, which helps manage and track different inspection tasks and ensure the accurate execution and monitoring of tasks. It is often used in equipment maintenance, quality control and other fields.

[0090] Prepare the inspection task timing association sequence, which records the execution time sequence of each inspection task. For example, if task A is executed before task B, then task A is before task B in the sequence. Assume that there are three inspection tasks: check device A, check device B, and check device C. Their execution order is: check device A->check device B->check device C. Then the task timing association sequence is: [check device A, check device B, check device C].

[0091] Construct a recursive attention network, which includes an encoder and a decoder. The encoder uses a positional encoding mechanism to incorporate the position information of the task in the time series into the representation of the task. For example, if the inspection device A is at the first position in the sequence, its position information will be encoded and added to its representation. The encoder uses a multi-head attention mechanism to capture the semantic association between tasks and map the task sequence to a semantic feature space. The decoder also uses a multi-head attention mechanism to perform weighted aggregation on the semantic features output by the encoder to generate a task dependency graph. The graph shows the dependency relationship between tasks in the form of a directed acyclic graph. For example, there is a directed edge from "inspection device A" to "inspection device B" in the graph, indicating that "inspection device B" depends on "inspection device A".

[0092] Construct a task dependency matrix based on the total number of inspection tasks. Assuming there are three inspection tasks in total, construct a 3x3 matrix. The rows and columns of the matrix correspond to the identifiers of each inspection task, for example, the first row and the first column correspond to "Inspection device A", the second row and the second column correspond to "Inspection device B", and the third row and the third column correspond to "Inspection device C". In the initial state, all elements in the matrix are set to the second preset value, such as "0".

[0093] Traverse all the edges in the task dependency graph, and obtain the source task node, target task node, and their corresponding dependency weight values ​​for each edge. For example, there is an edge pointing from "Inspection Device A" to "Inspection Device B" with a weight value of "1". According to the identifiers of the source task node and the target task node, determine their corresponding row index and column index in the task dependency matrix. For example, "Inspection Device A" corresponds to the first row, and "Inspection Device B" corresponds to the second column. Set the element value at the corresponding position in the matrix to the first preset value, such as "1". In this example, set the element value of the first row and second column of the matrix to "1".

[0094] Verify the validity of the task dependency matrix and check whether the element values ​​on the diagonal of the matrix are all the second preset values, such as "0". This ensures that the task itself will not depend on itself. Secondly, check whether the matrix satisfies the directed acyclic graph property, that is, determine whether there is a circular dependency. For example, if "inspection device A" depends on "inspection device B", "inspection device B" depends on "inspection device C", and "inspection device C" depends on "inspection device A", then there is a circular dependency. If the verification fails, it is necessary to modify the first preset value and the second preset value and rebuild the task dependency matrix.

[0095] The task dependency matrix is ​​optimized for storage, and compressed storage is used to store only the position information of the element value being the first preset value and its corresponding row and column index information. For example, only the row and column indexes of the elements with a value of "1" in the matrix are recorded, and a mapping relationship table between the task dependency matrix and the inspection task identifier is generated, for example, recording information such as "the first row corresponds to inspection device A" and "the second column corresponds to inspection device B". Finally, the task dependency matrix and the mapping relationship table are output.

[0096] In this embodiment, through the recursive attention network and position encoding mechanism, the temporal dependencies and semantic associations between tasks can be effectively captured, so as to more accurately construct the task dependency graph and matrix, and express the task dependency in the form of a matrix, which is more concise and intuitive, and convenient for the subsequent task scheduling and optimization algorithm. It adopts a compressed storage method to store only the necessary element information, effectively reducing the storage space occupied by the task dependency matrix.

[0097] S2. Based on the task priority classification results, task timing association sequence and task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved in combination with an improved variable neighborhood search algorithm to obtain a candidate task allocation scheme;

[0098] The hierarchical dual network structure is a deep learning model structure, usually composed of two independent neural networks, one of which is used for the main task (main network) and the other is used for the auxiliary task (auxiliary network). The task allocation main network refers to the neural network part responsible for the main task decision in the hierarchical dual network structure. The state evaluation auxiliary network is the auxiliary part in the hierarchical dual network, which is responsible for evaluating the environmental state. The policy branch refers to a part of the network structure in the deep learning network that is specifically used to generate policy or decision outputs. The task sequence refers to a group of tasks arranged in sequence in multi-task learning or task scheduling. The task sequence defines the order of task execution and may be affected by factors such as time windows and dependencies. The time window constraint means that in task scheduling or resource allocation problems, tasks must be started or completed within a specific time range. The variable neighborhood search algorithm is an optimization algorithm that optimizes the objective function by searching for improvements to the solution in the neighborhood space of the current solution.

[0099] In an optional embodiment,

[0100] Based on the task priority classification results, task timing association sequence and task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved by combining the improved variable neighborhood search algorithm, and the candidate task allocation schemes obtained include:

[0101] Based on the task priority classification results, the task timing association sequence and the task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network;

[0102] Input the inspection task feature vector into the first strategy branch, the second strategy branch and the third strategy branch of the task allocation main network, wherein the first strategy branch uses a first-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a global feature map, the second strategy branch uses a second-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a local feature map, and the third strategy branch uses a third-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a detail feature map;

[0103] The global feature map, the local feature map and the detail feature map are superimposed to obtain a comprehensive feature representation, and the comprehensive feature representation is input into a fully connected layer to generate an initial task allocation plan including a task sequence and a time window constraint;

[0104] Calculate the spatial distance between tasks based on the position coordinates of adjacent tasks in the task sequence, calculate the time interval between tasks based on the time window constraint, and combine the spatial distance and the time interval to construct a task distance matrix;

[0105] An initial temperature parameter is set for the initial task allocation scheme based on the state assessment auxiliary network. According to the initial temperature parameter, a selection is made among an exchange operation, an insertion operation and a reversal operation in combination with a preset operation probability, wherein the exchange operation is used to exchange the positions of two tasks in the task sequence, the insertion operation is used to move a task in the task sequence to another position, and the reversal operation is used to reverse the order of subsequences in the task sequence. After executing the operation, the initial temperature parameter is decayed, and an operation is selected again based on the attenuated temperature. The selection and execution of the operation and the temperature decay are repeated until a preset maximum number of iterations is reached, so as to obtain multiple optimal inspection routes, and generate a candidate task allocation scheme based on the optimal inspection route.

[0106] Extract features from the inspection task to obtain a feature vector of the inspection task. For example, a feature vector of an inspection task may include information such as the latitude and longitude of the task location, the urgency of the task, and the estimated completion time of the task. Assume that the feature vector of an inspection task is [30.5, 114.3, 5, 30], which represents latitude, longitude, urgency, and estimated completion time, respectively.

[0107] The feature vector of the inspection task is input into the main task allocation network. The main task allocation network contains multiple parallel strategy branches, each of which uses convolution kernels of different sizes to process the feature vector. For example, the first strategy branch uses a convolution kernel of size 3x1 to extract global features, the second strategy branch uses a convolution kernel of size 2x1 to extract local features, and the third strategy branch uses a convolution kernel of size 1x1 to extract detail features. After convolution operation on the above example feature vector [30.5, 114.3, 5, 30], the global feature map, local feature map and detail feature map are obtained respectively.

[0108] The global feature map, local feature map and detail feature map are superimposed to obtain a comprehensive feature representation. For example, the three feature maps are simply spliced ​​together. The comprehensive feature representation is input to the fully connected layer to generate an initial task allocation scheme, which contains task sequences and time window constraints. For example, the generated initial task allocation scheme is: [Task A, Task B, Task C], and the time window corresponding to each task, such as the time window of Task A is [8:00, 9:00], the time window of Task B is [9:00, 10:00], and the time window of Task C is [10:00, 11:00].

[0109] The spatial distance between tasks is calculated based on the position coordinates of adjacent tasks in the task sequence. For example, if the position of task A is (30.5, 114.3) and the position of task B is (30.6, 114.4), the spatial distance between task A and task B can be obtained by calculating the distance between the two points. At the same time, the time interval between tasks is calculated based on the time window constraint. For example, if the time window of task A is [8:00, 9:00] and the time window of task B is [9:00, 10:00], the time interval between task A and task B is 0. The spatial distance and time interval are combined to construct the task distance matrix.

[0110] An initial temperature parameter is set for the initial task allocation scheme, for example, to 100. According to the initial temperature parameter, combined with the pre-set operation probability, for example, the probability of the exchange operation is 0.4, the probability of the insertion operation is 0.3, and the probability of the reversal operation is 0.3, a selection is made among the exchange operation, the insertion operation, and the reversal operation. Assume that the exchange operation is selected, and the positions of two tasks in the task sequence are exchanged, for example, the positions of task B and task C are exchanged. After the operation is executed, the initial temperature parameter is attenuated, for example, multiplied by the attenuation coefficient of 0.9, to obtain a new temperature parameter of 90. Based on the attenuated temperature, the operation is selected again, and the selection and execution of the operation and the temperature attenuation are repeated until the preset maximum number of iterations, for example, 1000 times, is reached. Finally, multiple optimal inspection routes are obtained, and candidate task allocation schemes are generated based on these routes. For example, the optimal inspection route obtained is: [Task A, Task C, Task B].

[0111] In this embodiment, through the deep reinforcement learning framework and variable neighborhood search algorithm, a better inspection route can be obtained, the inspection time and cost can be reduced, thereby improving the inspection efficiency. Tasks can be allocated according to task priority, timing association and dependency, resource utilization can be optimized, and resource waste can be avoided. The deep reinforcement learning framework can dynamically adjust the task allocation plan according to actual conditions, enhance the adaptability and robustness of the model, and cope with complex and changeable inspection environments.

[0112] In an optional embodiment,

[0113] After the operation is performed, the initial temperature parameter is attenuated as shown in the following formula:

[0114] ;

[0115] Among them, T k+1 represents the temperature value at the k+1th iteration, T k represents the temperature value at the kth iteration, k represents the number of iterations, β represents the temperature attenuation adjustment factor, ω represents the periodic fluctuation factor, α represents the temperature attenuation rate control factor, and λ represents the KL divergence adjustment factor. represents the inferred distribution q(z|x k ) and the prior distribution p(z).

[0116] In this embodiment, by dynamically adjusting the temperature parameters, the exploration and utilization in the generation process can be better controlled, thereby generating clearer and more realistic samples. In the early stages of training, higher temperature values ​​encourage the model to explore different possibilities, thereby increasing the diversity of generated samples. As training progresses, the temperature value gradually decreases, and the model gradually converges to a better solution, thereby generating more representative samples. By introducing the KL divergence adjustment factor, the temperature decay rate can be dynamically adjusted according to the difference between the inferred distribution and the prior distribution, thereby avoiding drastic fluctuations during training and improving the stability of training.

[0117] S3. Based on the candidate task allocation scheme, a state space is constructed for each inspection device. Each inspection device is treated as an independent intelligent agent and equipped with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through the asynchronous experience replay pool. The candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on the graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, and decision outputs are generated and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

[0118] The asynchronous experience replay pool is a mechanism used in reinforcement learning to store and replay experience data generated during the interaction between the intelligent agent and the environment. The adaptive genetic operator is a method or method for dynamically adjusting the parameters of genetic operations (such as selection, crossover, mutation, etc.) in the genetic algorithm according to the characteristics of the problem or the performance of the individual. The residual message passing is a graph-based learning method, which is commonly used in graph neural networks (GNNs). The feature information of the nodes is transmitted to the adjacent nodes through the edges in the network, and the transmitted information is corrected or enhanced through the residual learning mechanism, thereby improving the effect of information transmission. The Bayesian inference algorithm is a type of inference method based on Bayes' theorem, which is used to derive the probability distribution of unknown variables from observed data.

[0119] In an optional embodiment,

[0120] Based on the candidate task allocation scheme, a state space is constructed for each inspection device. Each inspection device is used as an independent intelligent agent and equipped with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool. The candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage rate. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on a graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, a decision output is generated, and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme, including:

[0121] Collect inspection equipment status information and construct a three-dimensional state space matrix, map the inspection equipment number to the first dimension, map the location information, power status, task completion progress and other state variable types to the second dimension, and map the time step to the third dimension to form a dynamic representation of the equipment state. Based on the three-dimensional state space matrix, deploy a local decision network and a value network on each inspection device, input the state space dimension to the input layer of the three-layer fully connected neural network, output the action space dimension after hidden layer processing, and use the value network to evaluate the state value;

[0122] According to the decision data generated during the execution of the equipment, an asynchronous experience replay pool is created and the state transition data is stored. The current state, execution action, reward, and next state information are recorded. The sampling probability is assigned to the state transition data according to the task priority. For each inspection device, the mapping relationship between the device number, task number, and time window is encoded into a chromosome sequence. The three optimization objective functions of task completion time, energy consumption, and inspection coverage are calculated according to the encoding results. After the population is initialized, a fast non-dominated sort is performed. The non-dominated individuals are divided into the first layer by counting the dominance relationship between individuals, and the dominated individuals are randomly divided into other layers except the first layer. The crowding degree of the individuals that have completed the layer division is calculated. By sorting individuals at different layers and measuring the target space distance, the distance value of the objective function is accumulated.

[0123] According to the individual level and crowding information obtained, the crossover probability and mutation probability parameters are dynamically adjusted. Based on the adjusted parameters, the parent individuals are selected to perform single-point crossover to generate offspring individuals. The sites of the offspring individuals are randomly selected to perform gene mutations and the task constraint satisfaction is verified in real time. The parent and offspring populations are merged and the sorting and crowding calculations are re-executed. High-quality individuals are selected to form a new population. The iteration is continued until the termination condition is met. The individual with the highest fitness value is selected from the optimal solution set as the improved solution.

[0124] Based on the improved solution, a collaborative decision-making system is constructed according to the graph attention mechanism and the interaction between agents is modeled. Corresponding agent attention weights are assigned to different agents, and decision outputs are generated based on the agent attention weights. A confidence propagation network is constructed based on the decision outputs to model the uncertainty of the decision outputs. Each agent calculates the decision confidence based on local observations and influences the confidence estimates of other agents through network propagation. The residual information is exchanged and updated in combination with the residual message passing method. The residual convergence is monitored in real time and converges to a globally consistent task allocation constraint. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an anomaly. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

[0125] The chromosome sequence is a coding structure used to represent individual solutions in the genetic algorithm. Each chromosome sequence corresponds to a possible solution, which is usually represented by binary, real numbers or other formats. The high-quality individuals are individuals that perform well in the evolution of the genetic algorithm. High-quality individuals are usually evaluated based on the fitness function and can produce better solutions or show higher performance on the objective function of the problem. The residual convergence refers to the change of the residual term during the training process of the model when using residual learning (such as residual network).

[0126] Collect the status information of the inspection equipment, including the location of each device, the remaining power, and the completion progress of the current task. Organize this information into a three-dimensional matrix. The first dimension of this matrix represents the number of the inspection equipment, such as device 1, device 2, and so on. The second dimension represents the type of state variable of the device, such as location information, power status, task completion progress, and so on. The third dimension represents the time step, recording the status of the device at each time point. In this way, a dynamic representation of the device status is formed. For example, the position coordinates of device 1 at the 10th time step are (x=10, y=20), the power is 90%, and the task completion progress is 50%.

[0127] Deploy a local decision network and a value network on each inspection device. Input the dimension information of the three-dimensional state space matrix constructed earlier into the input layer of a three-layer fully connected neural network. The hidden layer of this neural network processes the input state information and finally outputs the action space dimension. The action space represents the actions that the device can take, such as moving to a certain location, performing a certain task, etc. The value network is used to evaluate the value of the current state, such as the expected benefits of completing the task, the energy consumed, etc. For example, the state information of the input device 1 at the 10th time step, the output of the neural network may be the action of moving to the location (x=20, y=30), and the value evaluated by the value network is 10.

[0128] Create an asynchronous experience replay pool to store state transition data. Each time a device performs an action, it records the current state, the action performed, the reward obtained, and the next state information. For example, device 1 is in state A at the 10th time step, performs the action of moving to position B, obtains a reward of 5, and reaches state B at the 11th time step. All this information will be stored in the experience replay pool. Different sampling probabilities are assigned to different state transition data according to the priority of the task. The state transition data corresponding to tasks with higher priority have a higher probability of being sampled. For example, the sampling probability of emergency tasks is set to 0.8, while the sampling probability of regular tasks is set to 0.2.

[0129] The candidate task allocation scheme is optimized using a multi-objective genetic algorithm. The mapping relationship between device number, task number and time window is encoded into a chromosome sequence. For example, the chromosome sequence "1-2-3" means that device 1 performs task 2 in time window 3. The values ​​of the three objective functions of task completion time, energy consumption and inspection coverage are calculated based on the encoding results. For example, the task completion time corresponding to the chromosome sequence "1-2-3" is 10 minutes, the energy consumption is 20 units, and the inspection coverage is 80%. After the population is initialized, a fast non-dominated sort is performed. By counting the dominance relationship between individuals, the non-dominated individuals are divided into the first layer, and the dominated individuals are randomly divided into other layers. The crowding degree is calculated for the individuals at each layer. By sorting individuals at different layers and measuring the target space distance, the objective function distance value is accumulated. For example, if the task completion time, energy consumption and inspection coverage of individual A are better than those of individual B, then individual A dominates individual B.

[0130] According to the individual level and crowding information obtained, the crossover probability and mutation probability parameters are dynamically adjusted. Based on the adjusted parameters, the parent individuals are selected to perform single-point crossover to generate offspring individuals. Gene mutation is performed on random sites of offspring individuals, and the satisfaction of task constraints is verified in real time. For example, a single-point crossover of parent chromosomes "1-2-3" and "2-1-3" may generate offspring chromosomes "1-1-3" and "2-2-3". After merging the parent and offspring populations, the sorting and crowding calculations are re-executed to select high-quality individuals to form a new population, and the iteration is continued until the termination condition is met. The individual with the highest fitness value is selected from the optimal solution set as the improved solution, where the high-quality individual refers to the individual with the highest crowding.

[0131] Construct a collaborative decision-making system based on the graph attention mechanism. Assign corresponding attention weights to different agents. Generate decision outputs based on attention weights. For example, device 1 and device 2 need to collaborate to complete task 3. According to the state and task requirements, they are assigned attention weights of 0.7 and 0.3 respectively. Then, a confidence propagation network is constructed based on the decision outputs. Each agent calculates the decision confidence based on local observations and propagates through the network to influence the confidence estimates of other agents. For example, device 1 determines that the confidence of completing task 3 is 0.9 based on its own state, and passes this information to device 2. Device 2 combines its own state and the received information to update its confidence estimate of completing task 3.

[0132] Combined with the residual message passing method, the residual information is exchanged and updated, and the residual convergence is monitored in real time until it converges to the globally consistent task allocation constraint. Output the candidate scheduling scheme, and combine the Bayesian reasoning algorithm to determine whether there is an anomaly. If there is an anomaly, optimize again. Otherwise, output the current candidate scheduling scheme as the optimal scheduling scheme. For example, if a device fails and the task cannot be completed, it is necessary to re-optimize and adjust the task allocation scheme.

[0133] In this embodiment, by optimizing the task allocation plan, the task completion time can be minimized and the inspection efficiency can be improved. By minimizing energy consumption, energy can be effectively saved and operating costs can be reduced. By optimizing the scheduling plan, the inspection coverage can be maximized to ensure that all key areas are effectively inspected.

[0134] In an optional embodiment,

[0135] Monitor the residual convergence in real time and converge to the globally consistent task allocation constraints, output candidate scheduling solutions and combine the Bayesian reasoning algorithm to determine whether there are any abnormalities. If so, optimize again. Otherwise, output the current candidate scheduling solution as the optimal scheduling solution, including:

[0136] Obtain the real-time status data of each agent, record the residual value between each agent and its neighboring nodes, store the residual data in the residual matrix, calculate the norm of the residual matrix and compare it with the preset convergence threshold. The norm of the residual matrix reflects the consistency of the allocation schemes between agents.

[0137] Monitor the residual norm change trend in multiple consecutive time slices. When the residual norm is continuously less than the preset threshold, determine that the system has reached a convergence state, and summarize the local constraint information of all agents, which includes device execution capability, time window limit, and path reachability constraint;

[0138] Based on the local constraints in the converged state, a global constraint set including task time window, path reachability, and energy consumption is constructed. Constraint conflict detection is used to ensure that there is no conflict between constraints, and candidate scheduling solutions that meet the global constraints are output.

[0139] Obtain historical normal scheduling data, extract scheduling characteristic parameters, the scheduling characteristic parameters include task time distribution, energy consumption distribution, equipment load distribution, task completion rate distribution, establish a multi-dimensional Gaussian mixture model of normal scheduling data, estimate model parameters through expectation maximization algorithm, the model parameters include the mean vector of each Gaussian component, covariance matrix and mixing weight, perform feature extraction on candidate scheduling schemes, and obtain candidate feature vectors of the same dimension as historical data, the candidate feature vectors include task time features, energy consumption features, equipment load features, and task completion rate features;

[0140] The log-likelihood value of the candidate feature vector is calculated based on the established Gaussian mixture model, and the log-likelihood value is the logarithmic value of the weighted sum of the conditional probability of the feature vector under each Gaussian component and the mixture weight. An anomaly detection threshold is set, and the anomaly detection threshold is determined based on the likelihood value distribution of historical data. When the log-likelihood value of the candidate scheduling scheme is lower than the threshold, the current candidate scheduling scheme is marked as abnormal. For the candidate scheduling scheme marked as abnormal, the scheduling scheme is regenerated based on non-dominated sorting and adaptive genetic operators. For the candidate scheduling scheme not marked as abnormal, the objective function value is calculated, and the candidate scheduling scheme not marked as abnormal with the highest objective function value is selected as the optimal scheduling scheme for output.

[0141] The residual matrix norm is an indicator for measuring the residual matrix (i.e., the difference between the actual result and the model prediction result). Common norms such as the L2 norm (Euclidean norm) can measure the size of the residual. The smaller the residual matrix norm, the closer the model prediction result is to the actual result. Otherwise, it indicates that there is a large error in the model. The path accessibility constraint is a condition that restricts path selection in path planning, graph search and other problems. The likelihood value is a numerical value used in statistics to measure the possibility of observed data under a given model. In Bayesian reasoning and maximum likelihood estimation, the likelihood value is used to evaluate the probability of a model generating observed data under a certain parameter set.

[0142] Obtain the real-time status data of each agent, such as location, remaining energy, task execution progress, etc. At the same time, record the residual value between each agent and its neighboring nodes, which can be expressed as the difference in the allocation schemes between agents. Store these residual data in a residual matrix. Then, calculate the norm of the residual matrix, which reflects the consistency of the allocation schemes between agents. Compare the calculated norm of the residual matrix with the pre-set convergence threshold.

[0143] Continuously monitor the residual norm change trend in multiple consecutive time slices. If the residual norm is continuously less than the preset threshold, the system is judged to have reached a convergence state. At this time, the local constraint information of all agents is summarized, including the execution capability of the equipment, time window restrictions, and path accessibility constraints.

[0144] After the system reaches convergence, a global constraint set including task time window, path accessibility, energy consumption, etc. is constructed based on the summarized local constraint information. Through the constraint conflict detection mechanism, such as checking whether there is time conflict, resource conflict, etc., it is ensured that there is no conflict between global constraints. Then, based on the conditions that meet the global constraints, the candidate scheduling scheme is output.

[0145] In order to determine whether the candidate scheduling scheme is abnormal, it is necessary to use the historical normal scheduling data for modeling, extract the scheduling characteristic parameters from the historical normal scheduling data, such as task time distribution, energy consumption distribution, equipment load distribution, task completion rate distribution, etc., and use the scheduling characteristic parameters to establish a multidimensional Gaussian mixture model. The parameters of the model are estimated by the expectation maximization algorithm, including the mean vector, covariance matrix and mixing weight of each Gaussian component.

[0146] Perform feature extraction on the generated candidate scheduling schemes to obtain candidate feature vectors with the same dimensions as the historical data, such as task time features, energy consumption features, equipment load features, task completion rate features, etc. Based on the established Gaussian mixture model, calculate the log-likelihood value of the candidate feature vector. The log-likelihood value is the logarithmic value of the weighted sum of the conditional probability of the feature vector under each Gaussian component and the mixture weight.

[0147] An anomaly detection threshold is pre-set, which can be determined based on the likelihood value distribution of historical data. If the log-likelihood value of a candidate scheduling solution is lower than the threshold, the candidate scheduling solution is marked as abnormal. For the candidate scheduling solution marked as abnormal, optimization algorithms such as non-dominated sorting and adaptive genetic operators are used to regenerate the scheduling solution. For the candidate scheduling solutions that are not marked as abnormal, their objective function values ​​are calculated, such as task completion efficiency, total energy consumption, etc. The candidate scheduling solution that is not marked as abnormal and has the highest objective function value is selected as the optimal scheduling solution output.

[0148] For example, suppose there are three agents and three tasks, the initial residual matrix norm is 0.8, and the preset threshold is 0.1. After several time slices, the residual matrix norm is reduced to 0.05, which is less than the threshold, and the system is judged to be converged. Suppose the local constraint of agent 1 is that it can only perform tasks in time periods 1-3, the local constraint of agent 2 is that it can only perform tasks 1 and 2, and the local constraint of agent 3 is that it can only perform tasks in time periods 2-4. After constructing the global constraint, a candidate scheduling scheme is generated: agent 1 performs task 1, agent 2 performs task 2, and agent 3 performs task 3. Suppose historical data shows that the task completion rate is concentrated above 90%, and the task completion rate of the candidate scheme is 70%, and its log-likelihood value is lower than the threshold, it is judged to be abnormal. After regenerating the scheduling scheme, a new candidate scheme is obtained, whose task completion rate is 95%, and the log-likelihood value is higher than the threshold, then the scheme is selected as the optimal scheme output.

[0149] In this embodiment, by real-time monitoring of residual convergence, it is possible to quickly determine whether the system has reached a consistent state, thereby reducing unnecessary iterative calculations and improving the efficiency of task allocation. By combining global constraints to generate a scheduling plan, it is possible to ensure that the final output scheduling plan meets all constraints and avoids local optimal solutions, thereby ensuring the global consistency of the scheduling plan. By performing anomaly detection through the Bayesian reasoning algorithm, abnormal scheduling plans can be effectively identified and optimized, thereby improving the robustness and reliability of the scheduling plan.

[0150] Figure 2 FIG. 1 is a schematic diagram of the structure of the unmanned intelligent inspection equipment collaborative dispatching system in the photovoltaic power generation scenario according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0151] The first unit is used to obtain the inspection tasks in the photovoltaic power generation scenario and convert them into inspection task feature vectors, build a task evaluation system based on a dynamic adaptive neural network containing multiple resonance layers, set different warning parameters and competitive learning mechanisms for each resonance layer, add the inspection task feature vectors to the task evaluation system, obtain task priority classification results, add the task priority classification results to a pre-set sequence predictor, model the inspection task feature vectors at different time scales to output task time series association sequences, build a recursive attention network based on the task time series association sequence to perform multi-task association analysis, establish a task dependency graph in combination with position encoding, and obtain a task dependency matrix based on the task dependency graph;

[0152] The second unit is used to construct a deep reinforcement learning framework based on the task priority classification results, the task timing association sequence and the task dependency matrix, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved in combination with an improved variable neighborhood search algorithm to obtain a candidate task allocation scheme;

[0153] The third unit is used to construct a state space for each inspection device based on the candidate task allocation scheme, treat each inspection device as an independent intelligent agent and equip it with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool, and the candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage rate. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on a graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, decision outputs are generated, and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing, and the candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

[0154] According to a third aspect of the embodiments of the present invention,

[0155] An electronic device is provided, comprising:

[0156] processor;

[0157] a memory for storing processor-executable instructions;

[0158] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0159] A fourth aspect of the embodiments of the present invention is:

[0160] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0161] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The unmanned intelligent inspection equipment collaborative scheduling method in the photovoltaic power generation scenario is characterized by: include: Obtain inspection tasks in photovoltaic power generation scenarios and convert them into inspection task feature vectors, build a task evaluation system based on a dynamic adaptive neural network containing multiple resonance layers, set different warning parameters and competitive learning mechanisms for each resonance layer, add the inspection task feature vectors to the task evaluation system, obtain task priority classification results, add the task priority classification results to a pre-set sequence predictor, model the inspection task feature vectors at different time scales to output task time series association sequences, build a recursive attention network based on the task time series association sequence to perform multi-task association analysis, establish a task dependency graph in combination with position encoding, and obtain a task dependency matrix based on the task dependency graph; Based on the task priority classification results, task timing association sequence and task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved in combination with an improved variable neighborhood search algorithm to obtain a candidate task allocation scheme; Based on the candidate task allocation scheme, a state space is constructed for each inspection device. Each inspection device is treated as an independent intelligent agent and equipped with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool. The candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage rate. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on a graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, a decision output is generated, and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme. Based on the task priority classification results, task timing association sequence and task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved by combining the improved variable neighborhood search algorithm, and the candidate task allocation schemes obtained include: Based on the task priority classification results, the task timing association sequence and the task dependency matrix, a deep reinforcement learning framework is constructed, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network; Input the inspection task feature vector into the first strategy branch, the second strategy branch and the third strategy branch of the task allocation main network, wherein the first strategy branch uses a first-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a global feature map, the second strategy branch uses a second-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a local feature map, and the third strategy branch uses a third-size convolution kernel to perform a convolution operation on the inspection task feature vector to obtain a detail feature map; The global feature map, the local feature map and the detail feature map are superimposed to obtain a comprehensive feature representation, and the comprehensive feature representation is input into a fully connected layer to generate an initial task allocation plan including a task sequence and a time window constraint; Calculate the spatial distance between tasks based on the position coordinates of adjacent tasks in the task sequence, calculate the time interval between tasks based on the time window constraint, and combine the spatial distance and the time interval to construct a task distance matrix; An initial temperature parameter is set for the initial task allocation scheme based on the state assessment auxiliary network. According to the initial temperature parameter, a selection is made among an exchange operation, an insertion operation and a reversal operation in combination with a preset operation probability, wherein the exchange operation is used to exchange the positions of two tasks in the task sequence, the insertion operation is used to move a task in the task sequence to another position, and the reversal operation is used to reverse the order of subsequences in the task sequence. After executing the operation, the initial temperature parameter is decayed, and an operation is selected again based on the attenuated temperature. The selection and execution of the operation and the temperature decay are repeated until a preset maximum number of iterations is reached, so as to obtain multiple optimal inspection routes, and generate a candidate task allocation scheme based on the optimal inspection route.

2. The method according to claim 1, characterized in that The inspection tasks in the photovoltaic power generation scenario are obtained and converted into inspection task feature vectors. A task evaluation system is constructed based on a dynamic adaptive neural network containing multiple resonance layers. Different warning parameters and competitive learning mechanisms are set for each resonance layer. The inspection task feature vectors are added to the task evaluation system to obtain task priority classification results. The task priority classification results are added to a pre-set sequence predictor. By modeling the inspection task feature vectors at different time scales, the output task time series association sequences include: Obtain an inspection task, based on the inspection task, obtain the task type feature through one-hot encoding, map the task location coordinates to spatial features through a geographic information system to obtain the task location feature, quantify the task urgency into a continuous value to obtain the task urgency feature, determine the equipment feature required for the task through a multi-hot encoding method, determine the task completion time feature through a normalization algorithm, and splice the task type feature, the task location feature, the task urgency feature, the equipment feature required for the task, and the task completion time feature to form the inspection task feature vector; Constructing a dynamic adaptive neural network, wherein the dynamic adaptive neural network comprises a plurality of resonance layers, each of the resonance layers comprises a plurality of neurons connected by synapses, information is transmitted between the resonance layers by feedforward connections and feedback connections, the neurons in the resonance layers have a competitive learning mechanism, when one neuron is activated, the activities of other neurons are inhibited, the neurons in the resonance layer are activated according to input features and warning parameters, the warning parameters control the activation threshold of the neurons, and the neurons in the resonance layer control the update speed of weights by the learning rate; The inspection task feature vector is added to the dynamic adaptive neural network, and the feature transformation is performed step by step through the neurons of the resonance layer, and the task priority classification result is obtained by combining the forward propagation algorithm; The task priority classification result is added to a pre-set sequence predictor constructed based on a long short-term memory network. The sequence predictor updates the network state according to the current input and the state of the previous time step, models the inspection task feature vector at different time scales, and outputs a task timing association sequence containing the association information between task features and historical task features.

3. The method according to claim 1, characterized in that Based on the task temporal association sequence, a recursive attention network is constructed to perform multi-task association analysis, and a task dependency graph is established in combination with position encoding. The task dependency matrix obtained based on the task dependency graph includes: A recursive attention network is constructed to process the task temporal association sequence, wherein the recursive attention network includes an encoder and a decoder, the task temporal association sequence is added to the recursive attention network, the encoder encodes the task temporal information through a position encoding mechanism, and maps the task temporal association sequence to a semantic feature space through a multi-head attention mechanism, and the decoder adopts a multi-head attention mechanism to perform weighted aggregation on the semantic feature space to generate a task dependency graph, and the task dependency graph represents the dependency relationship between tasks in the form of a directed acyclic graph; Construct a task dependency matrix according to the total number of inspection tasks, the number of rows and columns of the task dependency matrix are both the total number of inspection tasks, establish a correspondence between the rows and columns of the task dependency matrix and the identifiers of the inspection tasks, set the initial values ​​of all elements in the task dependency matrix to the second preset value, traverse all edges in the task dependency graph, obtain the source task node, the target task node and the corresponding dependency weight value of each edge, determine the corresponding row index and column index in the task dependency matrix based on the identifiers of the source task node and the target task node, and set the element value corresponding to the row index and the column index position in the task dependency matrix to the first preset value; Performing validity verification on the task dependency matrix, judging whether there is a cyclic dependency relationship based on whether the element values ​​on the diagonal of the task dependency matrix all satisfy the second preset value and whether the task dependency matrix satisfies the directed acyclic graph characteristic, if all are satisfied, it is considered that the validity verification has passed, otherwise it is considered that the validity verification has failed, and the first preset value and the second preset value are modified; The task dependency matrix is ​​optimized for storage based on compressed storage, wherein the location information of the storage element value being a first preset value and its corresponding row and column index information are stored, a mapping relationship table between the task dependency matrix and the inspection task identifier is generated, and the task dependency matrix is ​​output.

4. The method according to claim 1, characterized in that: After the operation is performed, the initial temperature parameter is attenuated as shown in the following formula: Among them, T k+1 represents the temperature value at the k+1th iteration, T k represents the temperature value at the kth iteration, k represents the number of iterations, β represents the temperature attenuation adjustment factor, ω represents the periodic fluctuation factor, α represents the temperature attenuation rate control factor, λ represents the KL divergence adjustment factor, and D KL (q(z|x k )||p(z)) represents the inferred distribution q(z|x) in the variational autoencoder k ) and the prior distribution p(z).

5. The method according to claim 1, characterized in that Based on the candidate task allocation scheme, a state space is constructed for each inspection device. Each inspection device is used as an independent intelligent agent and equipped with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool. The candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage rate. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on a graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, a decision output is generated, and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme, including: Collect inspection equipment status information and construct a three-dimensional state space matrix, map the inspection equipment number to the first dimension, map the location information, power status, task completion progress and other state variable types to the second dimension, and map the time step to the third dimension to form a dynamic representation of the equipment state. Based on the three-dimensional state space matrix, deploy a local decision network and a value network on each inspection device, input the state space dimension to the input layer of the three-layer fully connected neural network, output the action space dimension after hidden layer processing, and use the value network to evaluate the state value; According to the decision data generated during the execution of the equipment, an asynchronous experience replay pool is created and the state transition data is stored. The current state, execution action, reward, and next state information are recorded. The sampling probability is assigned to the state transition data according to the task priority. For each inspection device, the mapping relationship between the device number, task number, and time window is encoded into a chromosome sequence. The three optimization objective functions of task completion time, energy consumption, and inspection coverage are calculated according to the encoding results. After the population is initialized, a fast non-dominated sort is performed. The non-dominated individuals are divided into the first layer by counting the dominance relationship between individuals, and the dominated individuals are randomly divided into other layers except the first layer. The crowding degree of the individuals that have completed the layer division is calculated. By sorting individuals at different layers and measuring the target space distance, the distance value of the objective function is accumulated. According to the individual level and crowding information obtained, the crossover probability and mutation probability parameters are dynamically adjusted. Based on the adjusted parameters, the parent individuals are selected to perform single-point crossover to generate offspring individuals. The sites of the offspring individuals are randomly selected to perform gene mutations and the task constraint satisfaction is verified in real time. The parent and offspring populations are merged and the sorting and crowding calculations are re-executed. High-quality individuals are selected to form a new population. The iteration is continued until the termination condition is met. The individual with the highest fitness value is selected from the optimal solution set as the improved solution. Based on the improved solution, a collaborative decision-making system is constructed according to the graph attention mechanism and the interaction between agents is modeled. Corresponding agent attention weights are assigned to different agents, and decision outputs are generated based on the agent attention weights. A confidence propagation network is constructed based on the decision outputs to model the uncertainty of the decision outputs. Each agent calculates the decision confidence based on local observations and influences the confidence estimates of other agents through network propagation. The residual information is exchanged and updated in combination with the residual message passing method. The residual convergence is monitored in real time and converges to a globally consistent task allocation constraint. The candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an anomaly. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

6. The method according to claim 5, characterized in that Monitor the residual convergence in real time and converge to the globally consistent task allocation constraints, output candidate scheduling solutions and combine the Bayesian reasoning algorithm to determine whether there are any abnormalities. If so, optimize again. Otherwise, output the current candidate scheduling solution as the optimal scheduling solution, including: Obtain the real-time status data of each agent, record the residual value between each agent and its neighboring nodes, store the residual data in the residual matrix, calculate the norm of the residual matrix and compare it with the preset convergence threshold. The norm of the residual matrix reflects the consistency of the allocation scheme between agents; Monitor the residual norm variation trend in multiple consecutive time slices. When the residual norm is continuously less than a preset threshold, determine that the system has reached a convergence state, and summarize the local constraint information of all agents, which includes device execution capability, time window restriction, and path reachability constraint. Based on the local constraints in the converged state, a global constraint set including task time window, path accessibility, and energy consumption is constructed. Constraint conflict detection is used to ensure that there is no conflict between constraints, and candidate scheduling solutions that meet the global constraints are output. Obtain historical normal scheduling data, extract scheduling characteristic parameters, the scheduling characteristic parameters include task time distribution, energy consumption distribution, equipment load distribution, task completion rate distribution, establish a multi-dimensional Gaussian mixture model of normal scheduling data, estimate model parameters through expectation maximization algorithm, the model parameters include the mean vector of each Gaussian component, covariance matrix and mixing weight, perform feature extraction on candidate scheduling schemes, and obtain candidate feature vectors of the same dimension as historical data, the candidate feature vectors include task time features, energy consumption features, equipment load features, and task completion rate features; The log-likelihood value of the candidate feature vector is calculated based on the established Gaussian mixture model, and the log-likelihood value is the logarithmic value of the weighted sum of the conditional probability of the feature vector under each Gaussian component and the mixture weight. An anomaly detection threshold is set, and the anomaly detection threshold is determined based on the likelihood value distribution of historical data. When the log-likelihood value of the candidate scheduling scheme is lower than the threshold, the current candidate scheduling scheme is marked as abnormal. For the candidate scheduling scheme marked as abnormal, the scheduling scheme is regenerated based on non-dominated sorting and adaptive genetic operators. For the candidate scheduling scheme not marked as abnormal, the objective function value is calculated, and the candidate scheduling scheme not marked as abnormal with the highest objective function value is selected as the optimal scheduling scheme for output.

7. An unmanned intelligent inspection equipment collaborative dispatching system in a photovoltaic power generation scenario, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain the inspection tasks in the photovoltaic power generation scenario and convert them into inspection task feature vectors, build a task evaluation system based on a dynamic adaptive neural network containing multiple resonance layers, set different warning parameters and competitive learning mechanisms for each resonance layer, add the inspection task feature vectors to the task evaluation system, obtain task priority classification results, add the task priority classification results to a pre-set sequence predictor, model the inspection task feature vectors at different time scales to output task time series association sequences, build a recursive attention network based on the task time series association sequence to perform multi-task association analysis, establish a task dependency graph in combination with position encoding, and obtain a task dependency matrix based on the task dependency graph; The second unit is used to construct a deep reinforcement learning framework based on the task priority classification results, the task timing association sequence and the task dependency matrix, wherein the deep reinforcement learning framework adopts a hierarchical dual network structure, including a task allocation main network and a state evaluation auxiliary network, multiple parallel strategy branches are set in the task allocation main network and multi-scale processing is performed on the inspection task feature vector, a branch feature graph is generated and fused to obtain a comprehensive feature representation, an initial task allocation scheme is generated based on the comprehensive feature representation, the task sequence and time window constraints in the initial task allocation scheme are extracted to construct a task distance matrix, and the optimal inspection route is solved in combination with an improved variable neighborhood search algorithm to obtain a candidate task allocation scheme; The third unit is used to construct a state space for each inspection device based on the candidate task allocation scheme, treat each inspection device as an independent intelligent agent and equip it with a local decision network and a value network. For each intelligent agent, the state transition data is shared according to the task priority through an asynchronous experience replay pool, and the candidate task allocation scheme is optimized in combination with a multi-objective genetic algorithm, with the goal of minimizing the completion time, minimizing the energy consumption and maximizing the coverage rate. The solution is solved through non-dominated sorting and adaptive genetic operators, and an improved solution is generated. A collaborative decision-making system is constructed based on a graph attention mechanism. Based on the collaborative decision-making system, the interaction of intelligent agents is modeled, decision outputs are generated, and a confidence propagation network is constructed. The task allocation constraints are updated through residual message passing, and the candidate scheduling scheme is output and combined with the Bayesian reasoning algorithm to determine whether there is an abnormality. If so, it is optimized again. Otherwise, the current candidate scheduling scheme is output as the optimal scheduling scheme.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images