Artificial intelligence-based live working platform intelligent allocation method and system

Abstract

The application relates to the technical field of industrial control systems, and specifically discloses an intelligent distribution method and system for a live working platform based on artificial intelligence. The scheme collects multi-dimensional data through a working demand perception module, a platform state monitoring module and an environment situation assessment module, a multi-target collaborative decision engine fuses features, quantifies targets and searches for an optimal strategy based on deep reinforcement learning, dynamic balance of working efficiency, resource utilization and safety is achieved, and finally, a scheduling instruction is output by a dynamic distribution executor. The application improves the intelligent distribution level and self-adaptive decision-making capability of the live working platform.

Description

Technical Field

[0001] This invention belongs to the field of industrial control system technology, specifically relating to an intelligent allocation method and system for live-line working platforms based on artificial intelligence. Background Technology

[0002] In the field of power system operation and maintenance, the efficient allocation and scheduling of live-line working platforms is a key link in ensuring the stable operation of the power grid and improving operational efficiency. Traditional allocation methods mainly rely on human experience for decision-making, which is difficult to cope with complex and ever-changing field environments and real-time operational needs, resulting in problems such as low platform utilization, response delays, and resource waste. With the development of artificial intelligence technology, its potential in optimizing decision-making and intelligent scheduling provides new solutions for live-line working management.

[0003] Among them, the AI-based intelligent allocation method for live-line working platforms aims to achieve automatic task matching and optimal resource allocation through algorithmic models. This method needs to comprehensively consider multiple dimensions such as task type, platform status, and personnel skills to establish an intelligent allocation decision-making mechanism, thereby improving overall operational efficiency and safety.

[0004] In existing technologies, traditional allocation methods often employ fixed rules or simple priority strategies, which cannot adapt to dynamically changing work scenarios. These methods lack the ability to deeply mine and learn from historical work data, making it difficult to accurately predict work demands and platform availability. Furthermore, when dealing with multi-objective optimization problems, existing systems often fail to balance the conflicts between work efficiency, resource utilization, and safety requirements, leading to allocation results that deviate from the optimal solution.

[0005] Furthermore, traditional methods have limited responsiveness to real-time environmental factors and lack interpretability of allocation decisions, affecting the system's practicality and reliability. Therefore, there is an urgent need for an intelligent, adaptive, and interpretable allocation scheme for live-line working platforms. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent allocation method and system for live-line working platforms based on artificial intelligence, in order to solve the problems of low platform utilization, response delay and resource waste caused by existing technologies that rely on human experience, have fixed rules, are difficult to adapt to dynamic scenarios, lack the ability to balance multi-objective conflicts, and lack interpretability of decision-making.

[0007] To achieve the above objectives, the present invention provides an intelligent allocation system for live-line working platforms based on artificial intelligence. The system includes a work demand perception module, a platform status monitoring module, an environmental situation assessment module, a multi-objective collaborative decision engine, and a dynamic allocation executor.

[0008] The job requirement awareness module is used to collect and analyze multi-dimensional feature data of the job to be executed in real time. These feature data include at least the job type code, estimated job duration, required skill level, job geographical coordinates, and task urgency indicator.

[0009] The platform status monitoring module is used to continuously acquire real-time status information of all available live-line working platforms. This status information specifically includes the platform's unique identifier, the platform's current location latitude and longitude, the platform's current working status, the platform's remaining power percentage, the platform's load capacity, and the list of tools configured on the platform.

[0010] The environmental situation assessment module is used to access and integrate external environmental data sources to obtain real-time environmental parameters that affect operational safety and feasibility. These parameters include, but are not limited to, detailed weather forecast data for the next 6 hours, real-time power grid load data, topographic data of the work area, and historical accident statistical characteristics.

[0011] The multi-objective collaborative decision engine is the core of the system. It receives all output data from the task requirement perception module, platform status monitoring module, and environmental situation assessment module, and performs multi-objective optimization calculations through the built-in deep reinforcement learning network to generate the optimal matching scheme between the platform and the task.

[0012] The dynamic allocation executor generates specific platform scheduling instructions based on the matching scheme output by the multi-objective collaborative decision engine and issues them to the corresponding physical platform execution units. At the same time, it records the key basis and logical path of the allocation decision to the system log.

[0013] Furthermore, the internal workflow of the multi-objective collaborative decision engine specifically includes four stages: feature fusion, objective quantification, strategy search, and solution generation.

[0014] During the feature fusion stage, the engine standardizes and vectorizes the heterogeneous data from the three upstream modules to construct a unified high-dimensional feature tensor.

[0015] During the target quantification phase, the engine quantifies the three core optimization targets—operational efficiency, resource utilization, and operational safety—into computable utility functions.

[0016] The job efficiency utility function is mainly constructed based on the reciprocal of the total task completion time; the resource utilization utility function takes into account the weighted combination of platform vacancy rate and platform mobile energy consumption; and the job safety utility function is derived through a composite calculation of environmental risk factors and platform safety history.

[0017] During the policy search phase, the engine utilizes a pre-trained deep reinforcement learning network, taking a high-dimensional feature tensor as input, to explore allocation strategies that can simultaneously maximize the sum of the three utility functions mentioned above.

[0018] This deep reinforcement learning network uses a proximal policy optimization algorithm for iterative updates. Its value network is used to evaluate the value of a state, and its policy network is used to output the probability of an action.

[0019] During the scheme generation phase, the engine decodes the optimal action probability distribution output by the policy network into specific platform task matching pairs, along with a confidence score for each matching pair.

[0020] Furthermore, as one embodiment of the present invention, the training process of the deep reinforcement learning network is based on a training framework that combines offline and online methods.

[0021] During the offline training phase, supervised pre-training is performed using historical job datasets to initially establish the mapping relationship between state features and assignment decisions.

[0022] During the online training phase, new state-action-reward samples are collected through real-time interaction with the environment, and the network parameters are updated periodically using an experience replay mechanism.

[0023] The design of the reward function is closely related to the three quantitative objectives mentioned above. Specifically, it is a weighted sum of operational efficiency utility, resource utilization utility, and operational safety utility, while introducing a maximum negative reward term for behaviors that violate hard safety constraints.

[0024] Furthermore, the task demand perception module also integrates a short-term demand forecasting submodule. This submodule uses a time series analysis model to predict the types and quantities of new tasks that may be added within the next 24 hours based on historical task data patterns. The forecast results are then used as prior information to input into a multi-objective collaborative decision-making engine to support forward-looking resource reservation decisions.

[0025] Furthermore, the platform status monitoring module also includes a health assessment unit. This unit calculates a comprehensive health index for each platform by analyzing the platform's operation logs, maintenance records, and real-time sensor data. This health index is input as an important feature into the multi-objective collaborative decision-making engine to mitigate potential failure risks in allocation decisions and prioritize assigning critical tasks to platforms with better health status.

[0026] Furthermore, the environmental situation assessment module incorporates a dynamic risk assessment model. This model utilizes machine learning algorithms to comprehensively analyze current and predicted environmental parameters, calculating in real time the safety risk levels of different work areas at different times. The calculated risk levels not only serve as the output of the environmental situation assessment module but also directly participate in the construction of the work safety utility function in the multi-objective collaborative decision-making engine, enabling allocation decisions to dynamically respond to environmental changes.

[0027] Furthermore, before executing scheduling instructions, the dynamic allocation executor includes an instruction verification and conflict detection step. This step simulates the execution of the allocation scheme and checks for platform path conflicts, resource overruns, or violations of security procedures. If a conflict is detected, the information is fed back to the multi-objective collaborative decision engine for scheme replanning; if the verification passes, the instruction is immediately issued and the relevant resources are locked.

[0028] Furthermore, the entire system is deployed on a distributed computing architecture, which includes edge computing nodes and a cloud-based central server. The edge computing nodes are responsible for handling real-time data acquisition and preliminary analysis, while the cloud-based central server undertakes complex computing tasks such as the multi-objective collaborative decision engine. The two systems synchronize data efficiently and securely via an encrypted data link.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1. This invention constructs a complete system integrating job demand perception, platform status monitoring, environmental situation assessment, and multi-objective collaborative decision-making, achieving end-to-end automation and intelligence from data acquisition to decision execution. The multi-objective collaborative decision-making engine employs deep reinforcement learning technology, enabling simultaneous optimization of multiple conflicting objectives such as job efficiency, resource utilization, and job safety. Through quantified utility functions and strategy search, it finds the globally optimal or near-optimal allocation scheme, fundamentally overcoming the limitations of traditional fixed rules or simple priority strategies, and improving platform utilization and overall job efficiency.

[0031] 2. This invention introduces an adaptive decision-making mechanism based on deep reinforcement learning, enabling the system to continuously learn and evolve from historical data and real-time interactions, thereby dynamically adapting to complex and ever-changing operational scenarios and environmental conditions. The collaborative work of sub-modules such as short-term demand forecasting, platform health assessment, and dynamic risk assessment provides the decision engine with more comprehensive and forward-looking input information, enhancing the system's ability to cope with uncertainty and reducing response delays and decision-making errors caused by sudden environmental changes or information lag.

[0032] 3. The system design of this invention emphasizes the interpretability and reliability of the decision-making process. The multi-objective collaborative decision-making engine not only outputs the allocation results but also includes a confidence score, and the key decision-making logic and basis are recorded. The instruction verification and conflict detection links in the dynamic allocation executor further ensure the safety and feasibility of the allocation scheme before actual execution. This transparent and robust design greatly enhances the system's usability and operator trust, while also ensuring the safe conduct of live-line work by mitigating potential risks. Attached Figure Description

[0033] Figure 1This is a schematic diagram of the overall technical solution architecture of the intelligent distribution system for live-line working platforms based on artificial intelligence proposed in this invention;

[0034] Figure 2 This is a schematic diagram of the core principle framework of the multi-objective collaborative decision-making engine in this invention;

[0035] Figure 3 This is a schematic diagram of the multi-level interaction relationship and data flow between the operation requirement perception module, the platform status monitoring module, and the environmental situation assessment module in this invention;

[0036] Figure 4 This is a flowchart illustrating the instruction verification and conflict detection logic of the dynamically allocated executor in this invention. Detailed Implementation

[0037] The specific implementation method of the intelligent allocation system for live-line working platforms based on artificial intelligence is as follows.

[0038] Please refer to the attached document. Figure 1 The system is physically deployed on a distributed computing architecture, which is clearly divided into two levels: edge computing nodes and cloud central servers.

[0039] Edge computing nodes are deployed directly on each live-line working platform and in the regional operation management center. Their core hardware components include multi-core embedded processors, high-speed data acquisition cards, various industrial communication protocol interface modules, and high-speed network communication units with encryption functions.

[0040] Edge computing nodes undertake the most real-time data acquisition and preliminary analysis tasks in the system. Their specific workflow includes, but is not limited to, directly connecting to the GPS receiver on the platform to obtain the platform's latitude and longitude, connecting to the platform's battery management system to read the remaining power percentage, connecting to the platform's controller LAN bus to obtain the platform's operating status word and tool list status bit, and obtaining raw environmental parameters through the interface of a dedicated meteorological sensor network deployed in the operating area and the power grid data acquisition and monitoring control system.

[0041] The cloud-based central server adopts a high-availability cluster architecture, typically consisting of multiple high-performance servers connected by a load balancer. It is equipped with large-capacity memory and high-speed solid-state drive arrays to store massive amounts of historical job data, model parameters, and real-time operation logs.

[0042] The cloud-based central server hosts the most computationally intensive multi-objective collaborative decision-making engine and its related modules in this system.

[0043] Edge computing nodes and cloud central servers continuously synchronize data via a virtual private network link employing advanced encryption standard algorithms, ensuring the security and integrity of data transmission.

[0044] All data packets transmitted over the network follow a predefined application layer protocol format, which includes a fixed-length header for storing sequence numbers, timestamps, data source identifiers and checksums, and a variable-length payload for carrying specific structured business data.

[0045] The job requirement perception module is the starting point for system data input. Its main function is to collect and analyze the multi-dimensional feature data of all job tasks to be executed in real time.

[0046] This module is implemented at the software level as a continuously running background service process, deployed on an edge computing node that is directly connected to the job task management system.

[0047] The job requirement awareness module polls for new job tasks once per second by calling the application programming interface provided by the task management system.

[0048] Once a new task is detected, the module immediately starts a data parsing pipeline.

[0049] The first stage of the pipeline is data extraction, which reads all the raw fields associated with the task from the database table of the task management system. These fields include at least an 8-digit job type code, which uniquely identifies the category of the job, such as line inspection, equipment replacement, or fault handling.

[0050] An estimated job duration, expressed in minutes, represented by a 32-bit floating-point number.

[0051] The required skill level is an integer, typically ranging from 1 to 5, where 5 represents the highest skill requirement.

[0052] A job location coordinate containing two 64-bit floating-point numbers: longitude and latitude. And a task urgency indicator, typically an enumeration type, such as 0 for normal, 1 for important, and 2 for urgent.

[0053] The second stage of the pipeline is data cleaning and standardization. The module checks whether all read field values ​​are within the preset valid range. For example, the estimated job duration must be greater than 0 and less than 480 minutes, the skill level must be between 1 and 5, and the geographical coordinates must be within the valid range of latitude and longitude.

[0054] For any data that is out of range or incorrectly formatted, the module will record a detailed error log and attempt to obtain alternative values ​​from the task history or use preset default values. If both fail, the task will be marked as a data exception and will not be submitted for further processing.

[0055] The third stage of the pipeline is feature vector construction. The module concatenates and normalizes the cleaned field values ​​in a predefined order to form a fixed-dimensional numerical feature vector for subsequent processing.

[0056] An important extended function of the job demand perception module is its integrated short-term demand forecasting submodule.

[0057] This submodule is called by the job requirement awareness module as an independent dynamic link library in the software.

[0058] The core of the short-term demand forecasting submodule is a time series analysis model, specifically a seasonal autoregressive integral moving average model.

[0059] During the initialization phase, the model loads historical job data from the past 365 days, which includes the number of different job types generated each day and each hour.

[0060] The model training process first performs a stationarity test on historical data and then eliminates seasonal effects through differencing.

[0061] The model parameters include the autoregressive order, the differencing order, and the moving average order. These parameters are optimized and selected through grid search combined with the Akaike information content criterion.

[0062] The trained model, taking historical data from the 720 hours prior to the current time as input, can predict the number of new tasks of each job type that may be added per hour within the next 24 hours. The prediction results are output as a two-dimensional array, where one dimension represents the 24-hour time slice and the other dimension represents the different job type codes. The job demand awareness module packages these prediction results together with the actual task data collected in real time as part of its output.

[0063] The output data packet is encoded using a specific binary serialization format and includes a version number field, a data generation timestamp field, a number of real-time tasks field, a complete feature vector array for each real-time task, a number of prediction tasks field, and a generalized array of prediction task features.

[0064] Please refer to the attached document. Figure 3 The platform status monitoring module is another key data source for the system, responsible for continuously acquiring and maintaining the real-time status information of all available live-line working platforms.

[0065] The module is divided into two parts in terms of system architecture: one part is a lightweight agent program that runs locally on each live-line working platform, and the other part is a status aggregation service that runs on the cloud central server.

[0066] The platform's local agent program interacts directly with the platform's hardware system, reading a series of status parameters every 100ms via the platform's controller LAN bus. These parameters include the platform's unique identifier, a globally unique 16-byte identifier.

[0067] The platform's current location latitude and longitude are provided by the platform's built-in GPS module, with an accuracy typically reaching sub-meter level.

[0068] The platform's current operating status is a status word, where each bit represents a specific state. For example, the lowest bit indicates the platform's power switch status, the second lowest bit indicates the platform's movement status, the third bit indicates the platform's lifting mechanism status, and the remaining bits are reserved for future expansion. The platform's remaining battery percentage is an integer value between 0 and 100, read directly from the platform's battery management system.

[0069] The platform's load capacity is a floating-point number in kilograms, representing the platform's maximum design load. Also included is a list of configured tools, a list structure showing the model codes and status of all tools currently installed on the platform.

[0070] After collecting this raw data, the local agent performs preliminary verification, such as checking for jumps in latitude and longitude values ​​and whether the battery percentage changes within a reasonable range. Data that passes verification is encapsulated into data frames and sent to the cloud-based status aggregation service via a transmission control protocol connection.

[0071] The platform status monitoring module also includes a health assessment unit, which is implemented as a subroutine of the status aggregation service.

[0072] The health assessment unit maintains a platform health information database, which records the historical operating data of each platform, including total operating hours, historical maintenance records, fault alarm history, and wear sensor readings of key components.

[0073] The health assessment unit uses a model based on a weighted scoring algorithm to calculate the overall health index for each platform.

[0074] The model first assigns weights to different health-influencing factors, such as operating hours (0.2 weight), the time since the last maintenance (0.3 weight), fault alarm frequency (0.4 weight), and wear sensor readings (0.1 weight).

[0075] Then, a normalized score is applied to each factor. For example, the score for running hours is equal to 1 minus the current running hours of the platform divided by the platform's designed lifespan in hours, and the score for the most recent maintenance time is equal to 1 minus the current time minus the last maintenance time divided by the recommended maintenance cycle.

[0076] Finally, the scores of all factors are multiplied by their respective weights and summed to obtain a comprehensive health index between 0 and 1, where 1 represents the best health status.

[0077] This health index, together with the platform's real-time status information, constitutes the complete output of the platform's status monitoring module.

[0078] The output data packets of the platform status monitoring module adopt a binary format similar to but with a different structure than the operation requirement perception module. They include a total number of platforms field, a unique identifier for each platform, latitude and longitude coordinates, working status word, remaining power percentage, load capacity, tool list array, and health index value.

[0079] The environmental situation assessment module is responsible for accessing and integrating real-time information from external environmental data sources to assess environmental conditions that affect operational safety and feasibility.

[0080] This module is deployed on a cloud-based central server and connects to external systems through multiple data interfaces.

[0081] These interfaces include an application programming interface that connects to the National Meteorological Administration's data center to obtain detailed weather forecast data for the next 6 hours, including temperature, humidity, wind speed, wind direction, precipitation probability, and lightning activity index.

[0082] The interface connected to the power grid dispatch center's data bus is used to obtain real-time power grid load data, including line current, voltage, power factor, and regional load factor.

[0083] An application programming interface that connects to geographic information system services is used to obtain topographic data of the work area, including altitude, slope, surface type, and obstacle distribution.

[0084] It also includes a data interface that connects to the safety supervision and management system database to obtain historical accident statistics, including the frequency and severity of accidents in different types of operations under different environmental conditions.

[0085] The environmental situation assessment module pulls the latest data from these data sources once per minute and performs data fusion processing.

[0086] The environmental situation assessment module has a built-in dynamic risk assessment model, which is a trained machine learning model that specifically uses the gradient boosting decision tree algorithm.

[0087] The model uses current and predicted environmental parameters as input feature vectors to output the safety risk level of different work areas at different times. The risk level is represented by a value between 0 and 100, with higher values ​​indicating greater risk.

[0088] The training data for the dynamic risk assessment model comes from historical operation records, corresponding environmental data, and actual accident records.

[0089] During model training, feature importance analysis showed that wind speed, lightning activity index, grid load factor, and terrain complexity were the factors that contributed most to risk prediction.

[0090] The calculated risk level not only serves as an independent output of the environmental situation assessment module, but also directly participates in the construction of the operational safety utility function in the subsequent multi-objective collaborative decision engine.

[0091] The final output of the environmental situation assessment module is a structured data object, which includes data collection timestamps, an array of weather forecasts for the next 6 hours, a current power grid load data object, a terrain feature vector of the work area, and a risk level matrix for each area and time period.

[0092] Please refer to the attached document. Figure 2 The multi-objective collaborative decision-making engine is the intelligent core of the entire system. It receives all the output data from the job requirement perception module, the platform status monitoring module, and the environmental situation assessment module, and generates the optimal platform task matching scheme through a complex calculation process.

[0093] The engine is implemented in software as a high-performance computing service, running in a graphics processor-accelerated environment on a cloud-based central server.

[0094] The engine's workflow strictly follows four stages: feature fusion, target quantification, strategy search, and solution generation.

[0095] In the feature fusion stage, the engine first standardizes and vectorizes the heterogeneous data input from the three upstream modules.

[0096] Each task feature vector from the task requirement awareness module is converted into a 128-dimensional floating-point array.

[0097] Each platform status information from the platform status monitoring module is converted into a 256-dimensional floating-point array.

[0098] Environmental data from the environmental situation assessment module is converted into a 64-dimensional floating-point array. These vectors of different dimensions are then concatenated into a unified 448-dimensional high-dimensional feature tensor through a feature concatenation layer.

[0099] This tensor also contains a time dimension to represent data from multiple consecutive time steps. Typically, the system retains features from the most recent 10 time steps to capture dynamic trends.

[0100] During the objective quantization phase, the engine quantifies the system's three core optimization objectives into computable utility functions.

[0101] Operational efficiency utility function Defined as the reciprocal of the total task completion time. Total task completion time The calculation takes into account the time it takes for the platform to move to the work site. With the estimated duration of the task The specific calculation formula is as follows:

[0102] ;

[0103] in The great circle distance calculated based on the latitude and longitude between the platform's current location and the work site is divided by the platform's average moving speed. Therefore, the work efficiency utility function is expressed as:

[0104] ;

[0105] Resource utilization utility function It is a composite function that takes into account both the platform vacancy rate and the platform's mobile energy consumption.

[0106] Platform vacancy rate Defined as the percentage of time the platform does not perform tasks out of the total time. Platform mobile energy consumption. The resource utilization utility function is calculated based on the product of the platform's travel distance and its energy consumption coefficient per unit distance. The specific form of the utility function is:

[0107] ;

[0108] in and These are preset weighting coefficients, typically set to 0.7 and 0.3. (Job safety utility function) This is derived through a composite calculation of environmental risk factors and the platform's security history.

[0109] Environmental risk factors The risk level obtained directly from the environmental situation assessment module is normalized to the range of 0 to 1.

[0110] Platform security history Based on the number and severity of security incidents that occurred on the platform in the past 30 days, the fewer the security incidents and the less severe the incidents, the better. The higher the value, the better. The specific form of the job safety utility function is:

[0111] ;

[0112] During the policy search phase, the engine utilizes a pre-trained deep reinforcement learning network, taking the high-dimensional feature tensor generated during the feature fusion phase as input, to explore an allocation strategy that can simultaneously maximize the sum of the three utility functions mentioned above.

[0113] This deep reinforcement learning network employs a proximal policy optimization algorithm. Its network structure includes a common feature extraction layer and two dedicated output heads: a value network and a policy network.

[0114] The feature extraction layer consists of a 3-layer convolutional neural network followed by a 2-layer long short-term memory network, used to extract spatiotemporal features from the input tensor.

[0115] The value network is a fully connected neural network. Its input is the extracted features, and its output is a scalar value representing the long-term expected return of the current state.

[0116] The policy network is also a fully connected neural network. Its input is the extracted features, and its output is a probability distribution representing the probability of each action among all possible platform task matching actions.

[0117] The objective function of the near-end policy optimization algorithm is designed to maximize the expected reward while ensuring stable policy updates. Its mathematical expression is:

[0118] ;

[0119] in This represents the probability ratio between the old and new strategies. Represents the dominance function. For the clipping function, The trimming parameter is typically set to 0.2. Dominance function. The generalized advantage estimation algorithm is used to calculate the value network output and the actual reward obtained.

[0120] During the inference phase, the action probability distribution output by the policy network is normalized by the softmax function, and the action with the highest probability is selected as the optimal allocation decision for the current state.

[0121] During the solution generation phase, the engine decodes the optimal action probability distribution output by the policy network into specific platform task matching pairs. The decoding process first discretizes the continuous action space into all possible combinations of platform tasks, then selects the top 5 combinations as candidate solutions based on action probabilities. Each candidate solution is assigned a confidence score, which is based on the probability value output by the policy network and the value network's assessment of the state's value.

[0122] Ultimately, the engine selects the scheme with the highest confidence score as the final output. This output is a structured allocation scheme object containing an array of matching platform-unique identifiers, an array of matching task identifiers, the estimated start and finish times for each matching pair, and the confidence score for each matching pair.

[0123] Please refer to the attached document. Figure 4 The dynamic allocation executor is the final link in the system's decision-making and execution process. It receives the matching scheme output by the multi-objective collaborative decision engine and is responsible for converting it into specific executable instructions.

[0124] The dynamic allocation executor is implemented in software as a transactional service, ensuring the atomicity, consistency, isolation, and durability of allocation operations.

[0125] The executor's workflow begins with the instruction generation phase, where the executor transforms the abstract matching scheme into specific platform scheduling instructions.

[0126] These instructions include platform movement instructions, which include the target's latitude and longitude coordinates and recommended route. Platform task loading instructions include task details and a list of required tools.

[0127] The platform status change command changes the platform status from idle to working.

[0128] All instructions are encoded according to a predefined instruction set architecture to ensure that the platform controller can parse them correctly.

[0129] Before executing scheduling instructions, the dynamic allocation executor includes a crucial instruction verification and conflict detection step.

[0130] This step uses a high-fidelity simulation environment to virtually execute the allocation scheme.

[0131] The simulation environment calculates the movement path for each assigned platform and checks for path intersections or time overlaps between platforms, i.e., path conflict detection.

[0132] The simulation environment also checks whether the total workload of the platform exceeds its load capacity or whether the power supply is sufficient to support the entire task cycle when the platform is executing a new task, i.e., resource over-limit detection.

[0133] At the same time, the simulation environment will check against the safety regulations database to see if the allocation plan violates any hard safety constraints, such as prohibiting certain types of operations under specific weather conditions.

[0134] The simulation environment uses the same control logic and physical model as the real platform to ensure the accuracy of the simulation results.

[0135] If any type of conflict or violation is detected, the simulation environment will generate a detailed conflict report, including the conflict type, the platform and task identifiers involved, and the conflict severity score.

[0136] The report will be immediately fed back to the multi-objective collaborative decision-making engine, triggering a targeted replanning of the solution.

[0137] The replanning process considers conflict information, adjusts the weights of the utility function, or adds constraints to generate new candidate solutions. If the verification passes and no conflicts are detected, the dynamic allocation executor immediately sends instructions to the corresponding physical platform execution unit via a message queue.

[0138] After the instruction is issued, the executor will lock the relevant platform and task resources to prevent them from being reused by other allocation operations. It will also record the key basis for this allocation decision, including the feature vector used, utility function value, decision timestamp, and verification results, in the system's distributed log database for subsequent auditing and analysis.

[0139] The data flow and control flow of the entire system follow strict timing logic. Edge computing nodes collect raw data and perform preliminary processing at a speed of 100ms.

[0140] The processed data is transmitted to the cloud central server via an encrypted link, with the transmission delay typically controlled within 1 second.

[0141] Each module on the cloud-based central server consumes and processes data on demand. The computation cycle of the multi-objective collaborative decision engine is typically 5 to 10 seconds to adapt to dynamic changes in the operational scenario without causing system oscillations due to excessive frequency. The instruction verification and execution cycle of the dynamically allocated executors is synchronized with the decision engine.

[0142] The system ensures time consistency of all components through a distributed clock synchronization protocol, providing a reliable foundation for time-based decision-making and logging.

[0143] The system also has a robust exception handling mechanism. When any module detects an abnormal state, it will report it to the central monitoring terminal through a predefined exception code and level system, and automatically trigger the corresponding degradation strategy according to the exception level. For example, it can use the last known valid data when communication is interrupted, or switch to a rule-based backup decision mode when the decision engine fails, to ensure that the system can maintain its basic functions under various abnormal conditions.

[0144] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0145] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent distribution system for live-line working platforms based on artificial intelligence, characterized in that, include: The task requirement perception module is used to collect and parse multi-dimensional feature data of the task to be executed in real time. The multi-dimensional feature data includes the task type code, the estimated duration of the task, the required skill level, the geographical coordinates of the task, and the task urgency indicator. The platform status monitoring module is used to continuously acquire the real-time status information of all available live-line working platforms. The real-time status information includes the platform's unique identifier, the platform's current location latitude and longitude, the platform's current working status, the platform's remaining power percentage, the platform's load capacity, and the list of tools configured on the platform. The environmental situation assessment module is used to access and integrate external environmental data sources to obtain real-time environmental parameters that affect operational safety and feasibility. These real-time environmental parameters include real-time power grid load data, operational area topography data, and historical accident statistical characteristics. The multi-objective collaborative decision engine receives all output data from the task requirement perception module, platform status monitoring module, and environmental situation assessment module. It performs multi-objective optimization calculations through a built-in deep reinforcement learning network to generate the optimal matching scheme between the platform and the task. The dynamic allocation executor is used to generate specific platform scheduling instructions based on the matching scheme output by the multi-objective collaborative decision engine and issue them to the corresponding physical platform execution units. At the same time, it records the key basis and logical path of the allocation decision to the system log. The internal workflow of the multi-objective collaborative decision-making engine includes four stages: feature fusion, objective quantification, strategy search, and solution generation. During the feature fusion stage, the engine standardizes and vectorizes the heterogeneous data from the three upstream modules to construct a unified high-dimensional feature tensor. During the target quantification phase, the engine quantifies the three core optimization targets—operational efficiency, resource utilization, and operational safety—into computable utility functions. During the policy search phase, the engine utilizes a pre-trained deep reinforcement learning network, taking a high-dimensional feature tensor as input, to explore allocation strategies that can simultaneously maximize the sum of three utility functions. During the solution generation phase, the engine decodes the optimal action probability distribution output by the policy network into specific platform task matching pairs, and attaches a confidence score for each matching pair. The task efficiency utility function is constructed based on the reciprocal of the total task completion time; The resource utilization utility function comprehensively considers the weighted combination of platform vacancy rate and platform mobile energy consumption; The operational safety utility function is derived through a composite calculation of environmental risk factors and platform safety history records; The deep reinforcement learning network is iteratively updated using a proximal policy optimization algorithm. The deep reinforcement learning network includes a value network and a policy network; Value networks are used to assess the value of a state; The policy network is used to output action probabilities.

2. The intelligent distribution system for a live-line working platform based on artificial intelligence according to claim 1, characterized in that, The task demand perception module also integrates a short-term demand forecasting sub-module. The short-term demand forecasting submodule uses a time series analysis model to predict the types and quantities of new tasks that may be added in the future based on the patterns in historical operation data. The prediction results are used as prior information to input into a multi-objective collaborative decision engine to support forward-looking resource reservation decisions.

3. The intelligent distribution system for a live-line working platform based on artificial intelligence according to claim 1, characterized in that, The platform status monitoring module also includes a health assessment unit; The health assessment unit calculates the comprehensive health index of each platform by analyzing the platform's operation logs, maintenance records, and real-time sensor data. The health index is used as an important feature input into the multi-objective collaborative decision engine to avoid potential failure risks in allocation decisions.

4. The intelligent distribution system for a live-line working platform based on artificial intelligence according to claim 1, characterized in that, The environmental situation assessment module has a built-in dynamic risk assessment model. The dynamic risk assessment model uses machine learning algorithms to calculate the safety risk level of different work areas in different time periods in real time, by combining current and forecast environmental parameters. The risk level is used to construct the job safety utility function in the multi-objective collaborative decision engine.

5. The intelligent distribution system for a live-line working platform based on artificial intelligence according to claim 1, characterized in that, The dynamic allocation executor includes instruction verification and conflict detection steps before executing scheduling instructions; The instruction verification and conflict detection process simulates the execution of the allocation scheme to check for platform path conflicts, resource overruns, or violations of security procedures. If a conflict is detected, the information is fed back to the multi-objective collaborative decision engine for replanning. If the verification passes, an instruction will be issued immediately and the relevant resources will be locked.

6. The intelligent distribution system for a live-line working platform based on artificial intelligence according to claim 1, characterized in that, The system is deployed on a distributed computing architecture; The distributed computing architecture includes edge computing nodes and a cloud central server; Edge computing nodes are responsible for handling data acquisition and preliminary analysis with high real-time requirements; The cloud-based central server undertakes complex computing tasks such as multi-objective collaborative decision-making engines; The two systems achieve efficient and secure data synchronization via an encrypted data link.

7. An intelligent allocation method for live-line working platforms based on artificial intelligence, wherein the intelligent allocation of live-line working platforms is achieved using the intelligent allocation system for live-line working platforms based on artificial intelligence as described in any one of claims 1-6.

Images