Power plant safety monitoring method based on task priority coefficient and computing power allocation

Abstract

This invention belongs to the field of Internet of Things (IoT) technology and discloses a power plant safety monitoring method based on task priority coefficients and computing power allocation. The method constructs a model library with a shared basic model; maps collected and aligned operating environment features and task-specific features to graph nodes in the topology of power plant equipment; dynamically sorts all tasks and allocates computing resources using task priority coefficients obtained by extracting risk propagation features between nodes using a graph neural network; when resources for a task with the first priority coefficient are insufficient, resource preemption for a task with the second priority coefficient is evaluated and executed; after obtaining resources, the corresponding adapter weights and basic model are loaded from the model library, and batch processing inference is performed on tasks with the same basic model; finally, monitoring instructions are generated based on the inference results, improving resource utilization, realizing intelligent scheduling decisions based on real-time risk quantification assessment and ultra-low latency switching response, and ensuring the reliability and real-time performance of power plant safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, and in particular to a power plant safety monitoring method based on task priority coefficients and computing power allocation. Background Technology

[0002] With the development of intelligent power plants and industrial IoT technologies, using artificial intelligence, especially large-scale models, for power plant safety monitoring has become an important trend. However, traditional power plant safety systems mainly rely on fixed rule engines, shallow machine learning models, or full-scale deep learning models with extremely high computational resource requirements. These methods have significant limitations: first, they cannot handle multimodal data and complex natural language commands; second, the deployment and fine-tuning of full-scale models are costly and difficult to adapt to the rapid iteration of private domain data in power plants; and third, traditional scheduling mechanisms cannot perceive the real-time operating status and risk level of equipment, resulting in rigid resource allocation and difficulty in coping with the instantaneous computing power demands of high-real-time, high-priority emergencies.

[0003] Existing attempts to apply large models to industrial scenarios have improved the level of intelligence, but significant bottlenecks remain in computing power allocation. Traditional cloud computing scheduling algorithms focus on throughput or fairness, but fail to address the core contradiction between the high real-time constraints and limited computing resources unique to the industrial safety field. In particular, in scenarios requiring the loading of large models, frequent switching between different tasks can lead to significant switching latency and memory fragmentation due to the time-consuming model loading process, making it impossible to ensure the reliability and low latency of power plant safety monitoring while improving the utilization of scarce computing resources.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a power plant safety monitoring method based on task priority coefficients and computing power allocation, aiming to solve the technical problem of ensuring the reliability and low latency of power plant safety monitoring while improving the utilization rate of scarce computing power resources.

[0006] To achieve the above objectives, this invention provides a power plant safety monitoring method based on task priority coefficients and computing power allocation. The power plant safety monitoring method based on task priority coefficients and computing power allocation includes the following steps: For the basic large model, task adapter weights are trained and generated based on the power plant private domain security dataset to build a model library with a shared basic model; Real-time acquisition of power plant operating environment characteristics and task-specific characteristics, followed by time series alignment; A topology graph of power plant equipment is constructed, and the aligned operating environment features and task features are mapped to the nodes of the topology graph. A graph neural network is used to extract the risk propagation features between nodes, and the risk propagation quantification value of each task is output as the task priority coefficient. All pending tasks are dynamically sorted according to the task priority coefficient, and computing resources are allocated to each task according to the sorting result. When a first priority task requires computational resources and the currently available resources are insufficient, the feasibility of preempting the allocated resources for a second priority task is evaluated, wherein the priority of the second priority task is lower than that of the first priority task. If the evaluation is passed, the execution of the second priority task is paused, the computing resources of the second priority task are released, and the released computing resources are redistributed to the first priority task to achieve real-time dynamic allocation of computing power. After the first priority task obtains computing resources, the task adapter weights corresponding to the first priority task and the shared basic large model are loaded from the model library of the shared basic model into the allocated computing resources. Perform batch inference on multiple tasks that have been loaded with the same underlying large model; Based on the output of batch processing inference, generate power plant safety monitoring instructions or alarm information, and execute the corresponding safety monitoring response.

[0007] In one embodiment, the step of training and generating task adapter weights based on a power plant private domain security dataset to construct a model library with a shared base model includes: Freeze all original pre-trained parameters of the base model; A low-rank adaptation technique is used to inject new trainable low-rank matrices into the preset layers of the basic large model; Based on the power plant private domain security dataset, only the new trainable low-rank matrix is ​​trained, and the trained and optimized low-rank matrix is ​​stored as task adapter weights. The task adapter weights are independent of the original pre-trained parameters of the base model and together with the base model constitute a shared model library.

[0008] In one embodiment, the time series alignment includes: Based on the real-time requirements of the task itself, a unified reference sampling frequency is set; For feature sequences with sampling frequencies lower than the reference sampling frequency, an interpolation method is used to align the operating environment features and the task's own features to the reference sampling frequency; For feature sequences with a sampling frequency higher than the benchmark sampling frequency, a time window aggregation method is used to align the operating environment features and the task's own features to the benchmark sampling frequency; Using the reference sampling frequency as the time granularity, the characteristics of the operating environment and the characteristics of the task itself are synchronized and aligned according to the timestamp.

[0009] In one embodiment, the construction of the power plant equipment topology graph involves mapping aligned features to nodes in the topology graph, extracting risk propagation features between nodes using a graph neural network, and outputting a risk propagation quantification value for each task as a task priority coefficient, including: Based on the physical connection or logical control relationship of power plant equipment, a topology graph is constructed, wherein the nodes of the topology graph represent equipment or sensors, and the edges of the topology graph represent the association and dependency between equipment. The aligned operating environment features and task features are used as node attributes and input into the graph neural network. The graph neural network aggregates the feature information of adjacent nodes through a multi-layer message passing mechanism to simulate the propagation path and attenuation law of risk between devices. The node embedding representations output by the graph neural network are mapped through a fully connected layer to obtain the risk propagation quantification value corresponding to each task, wherein the risk propagation quantification value serves as the priority coefficient of the task.

[0010] In one embodiment, allocating computing resources to each task according to the sorting result includes: The required video memory and computing power for the current task are determined based on the sorting results. Traverse all available computing resources and filter out the set of computing resources that meet the memory capacity and computing power requirements of the task. From the set of computing resources, the computing resource with the smallest remaining video memory capacity is selected for allocation; The task is assigned to the selected computing resources, and the corresponding task adapter weights and basic large model are loaded.

[0011] In one embodiment, the feasibility of assessing the task of preempting the allocated resources with a second priority coefficient includes: Calculate the delay risk loss caused by the preemption of the second priority task; Calculate the security risk reduction benefit obtained by immediately executing the task with the first priority coefficient; Determine whether the benefits of reducing the security risk outweigh the losses from the delay risk. If the benefits of reducing the security risk outweigh the losses from the delay risk, then the assessment is deemed successful.

[0012] In one embodiment, pausing the execution of the second priority task, releasing the computing resources of the second priority task, and reallocating the released computing resources to the first priority task includes: Before preemption, save the current execution state and computation context of the task with the second priority coefficient; Suspend the execution of the second priority task and release the computing resources occupied by the second priority task; The freed computing resources are allocated to the task with the first priority coefficient, and a fast switch is performed; After the first priority task is completed, the second priority task is put back into the scheduling queue, and the current execution state of the second priority task is restored to the execution state according to the calculation context saved by the second priority task.

[0013] Furthermore, to achieve the above objectives, this invention also proposes a power plant safety monitoring device based on task priority coefficients and computing power allocation. This device is applied to the power plant safety monitoring method based on task priority coefficients and computing power allocation described above. The device includes: The task weight allocation module is used to train and generate task adapter weights based on the power plant private domain security dataset for the basic large model, so as to build a model library with a shared basic model. The time series alignment module is used to collect the operating environment characteristics and task characteristics of the power plant in real time and perform time series alignment. The priority coefficient calculation module is used to construct a topology graph of power plant equipment, map the aligned operating environment features and task features to the nodes of the topology graph, and use a graph neural network to extract the risk propagation features between nodes, and output the risk propagation quantification value of each task as the task priority coefficient. The computing resource allocation module is used to dynamically sort all pending tasks according to the task priority coefficient, and allocate computing resources to each task according to the sorting result. The resource preemption assessment module is used to assess the feasibility of preempting the allocated resources of a second priority task when a first priority task needs to compute resources and the currently available resources are insufficient. The priority of the second priority task is lower than that of the first priority task. The resource preemption execution module is used to pause the execution of the second priority task if the evaluation is passed, release the computing resources of the second priority task, and redistribute the released computing resources to the first priority task to achieve real-time dynamic allocation of computing power. The power plant model loading module is used to load the task adapter weights and the shared basic large model corresponding to the first priority task into the allocated computing resources after the first priority task obtains computing resources. The task batch processing module is used to perform batch inference on multiple tasks that have been loaded and share the same basic large model; The power plant safety monitoring module is used to generate power plant safety monitoring instructions or alarm information based on the output results of batch processing inference, and to execute corresponding safety monitoring responses.

[0014] Furthermore, to achieve the above objectives, the present invention also proposes a power plant safety monitoring device based on task priority coefficient and computing power allocation. The power plant safety monitoring device based on task priority coefficient and computing power allocation includes: a memory, a processor, and a power plant safety monitoring program based on task priority coefficient and computing power allocation stored in the memory and executable on the processor. The power plant safety monitoring program based on task priority coefficient and computing power allocation is configured to implement the steps of the power plant safety monitoring method based on task priority coefficient and computing power allocation as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a power plant safety monitoring program based on task priority coefficients and computing power allocation. When the power plant safety monitoring program based on task priority coefficients and computing power allocation is executed by a processor, it implements the steps of the power plant safety monitoring method based on task priority coefficients and computing power allocation described above.

[0016] This invention constructs a model library of shared basic models; it maps collected and aligned operating environment features and task-specific features to graph nodes in the topology diagram of power plant equipment; it uses task priority coefficients obtained by extracting risk propagation features between nodes using graph neural networks to dynamically sort all tasks and allocate computing resources; when resources for a task with the first priority coefficient are insufficient, it evaluates and executes resource preemption for a task with the second priority coefficient; after obtaining resources, it loads the corresponding adapter weights and basic models from the model library of basic models, performs batch processing inference on tasks with the same basic model, and finally generates monitoring instructions based on the inference results, thereby improving resource utilization, realizing intelligent scheduling decisions based on real-time risk quantification assessment and ultra-low latency switching response, and ensuring the reliability and real-time performance of power plant safety monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the first embodiment of the power plant safety monitoring method based on task priority coefficient and computing power allocation of the present invention. Figure 2This is a structural block diagram of the first embodiment of the power plant safety monitoring device based on task priority coefficient and computing power allocation of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] This invention provides a power plant safety monitoring method based on task priority coefficients and computing power allocation, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of a power plant safety monitoring method based on task priority coefficients and computing power allocation according to the present invention.

[0021] In this embodiment, the power plant safety monitoring method based on task priority coefficient and computing power allocation includes the following steps: Step S10: For the basic large model, train and generate task adapter weights based on the power plant private domain security dataset to build a model library with a shared basic model.

[0022] In this embodiment, the execution entity is a power plant safety monitoring device based on task priority coefficient and computing power allocation. This power plant safety monitoring device based on task priority coefficient and computing power allocation has functions such as data processing, data communication, and program execution. The power plant safety monitoring device based on task priority coefficient and computing power allocation can be a computer terminal device or other network device, or other devices with similar functions. This embodiment does not limit the scope of such devices.

[0023] It should be noted that traditional power plant safety systems primarily rely on fixed rule engines, shallow machine learning models, or full-scale deep learning models with extremely high computational resource requirements. These methods have significant limitations: first, they cannot handle multimodal data and complex natural language commands; second, the deployment and fine-tuning of full-scale models are costly and difficult to adapt to the rapid iteration of power plant private domain data; and third, traditional scheduling mechanisms cannot perceive the real-time operating status and risk level of equipment, leading to rigid resource allocation and difficulty in coping with the instantaneous computational demands of high-real-time, high-priority emergencies. Existing attempts to apply large models to industrial scenarios, while improving the level of intelligence, still face significant bottlenecks in computational resource allocation. Traditional cloud computing scheduling algorithms focus on throughput or fairness but fail to address the core contradiction between the high real-time constraints and limited computing resources unique to the industrial safety field. In particular, in scenarios requiring the loading of large models, frequent switching between different tasks results in significant switching latency and memory fragmentation due to model loading time, failing to ensure the reliability and low latency of power plant safety monitoring while simultaneously improving the utilization of scarce computing resources.

[0024] To address the aforementioned technical issues, this embodiment constructs a model library of shared basic models. The collected and aligned operational environment features and task-specific features are mapped to graph nodes in the power plant equipment topology diagram. Task priority coefficients, obtained by extracting risk propagation features between nodes using a graph neural network, are used to dynamically sort all tasks and allocate computing resources. When resources for a task with the first priority coefficient are insufficient, resource preemption for a task with the second priority coefficient is evaluated and executed. After acquiring resources, the corresponding adapter weights and basic models are loaded from the basic model library. Batch processing inference is performed on tasks with the same basic model, and finally, monitoring instructions are generated based on the inference results. This improves resource utilization, achieves intelligent scheduling decisions based on real-time risk quantification assessment, and enables ultra-low latency switching responses, ensuring the reliability and real-time performance of power plant safety monitoring. Specifically, this can be implemented as follows.

[0025] In the specific implementation, for the basic large model, task adapter weights are trained and generated based on the power plant private domain safety dataset to construct a model library sharing the basic model. This specifically includes: freezing all original pre-trained parameters of the basic large model; employing low-rank adaptation techniques to inject new trainable low-rank matrices into preset layers of the basic large model; training only the new trainable low-rank matrices based on the power plant private domain safety dataset, and storing the trained and optimized low-rank matrices as task adapter weights. The task adapter weights are independent of the original pre-trained parameters of the basic large model and together with the basic large model constitute the shared model library.

[0026] It's important to note that a large, pre-trained model with strong understanding and reasoning capabilities is selected. This model typically employs a Transformer architecture, and its original parameter set is considered a core asset shared by the system. During system deployment, all original pre-trained parameters of the base model are frozen by configuring the model settings. These original parameters remain unchanged throughout subsequent domain-specific fine-tuning, ensuring that the general knowledge inherent in the base model is not compromised.

[0027] For key weight matrices within the base model, such as the query, key-value matrices in the self-attention mechanism or the weights of fully connected layers, the system injects a pair of new, trainable low-rank matrices in parallel. The rank of the low-rank matrix is ​​much smaller than the dimension of the original weight matrix, ensuring that the total number of newly introduced parameters is extremely small. The system utilizes a labeled and cleaned power plant private domain security dataset. For each specific security task, a corresponding fine-tuning objective function is set. During fine-tuning training, the system strictly focuses computational resources on updating the new trainable low-rank matrix. All original pre-trained parameters remain frozen and do not participate in gradient updates. After training and optimization, only the converged low-rank matrix is ​​stored as the task adapter weights for that specific task. This adapter weight file is extremely small and stored independently from the base model file.

[0028] Each task's adapter weights are stored as independent, lightweight files, logically and physically independent of the base model's original pre-trained parameters. The shared base model library consists of two parts: the base model and its frozen original parameters; and multiple sets of task adapter weights trained for different security tasks.

[0029] Through the above steps, we have achieved efficient reuse of the basic model and lightweighting of the task adapter weights, so that during actual operation and inference, computing resources only need to load the original parameters of the basic large model once.

[0030] Step S20: Collect the operating environment characteristics and task characteristics of the power plant in real time, and perform time series alignment.

[0031] In the specific implementation, the alignment process involves setting a unified reference sampling frequency based on the real-time requirements of the task's own characteristics; for feature sequences with a sampling frequency lower than the reference sampling frequency, an interpolation method is used to align the operating environment features and the task's own features to the reference sampling frequency; for feature sequences with a sampling frequency higher than the reference sampling frequency, a time window aggregation method is used to align the operating environment features and the task's own features to the reference sampling frequency; and using the reference sampling frequency as the time granularity, the operating environment features and the task's own features are synchronized and aligned according to the timestamp.

[0032] It should be noted that, firstly, a unified baseline sampling frequency is established based on the real-time requirements of the target security task. This frequency is typically determined based on the response requirements of the most latency-sensitive critical security tasks, such as real-time fire early warning or P0-level violation detection. All subsequent data processing and feature vector construction will use this baseline sampling frequency as the time granularity.

[0033] For feature sequences with frequencies below the reference frequency, the system identifies and processes these sequences. Examples include power plant shift information, long-term equipment operating status, or daily updated equipment criticality levels. For these low-frequency feature sequences, the system uses interpolation to align their time series to the reference sampling frequency. This embodiment employs a zero-order hold-the-difference scheme, directly using the value from the previous known time point until the next time point arrives, ensuring that every time point in the low-frequency feature sequence has an estimated value synchronized with the reference frequency.

[0034] For feature sequences with frequencies higher than the reference frequency, the system identifies and processes these sequences, such as high-frequency vibration sensor data and real-time current or voltage data streams. The system defines a time window that matches the reference sampling period. For high-frequency data within this window, the system uses a time window aggregation method to downsample it to the reference sampling frequency. Common aggregation methods include extreme value extraction, statistical calculation, and outlier counting. This embodiment uses a statistical calculation scheme, calculating the mean, standard deviation, or trend slope within the window to reflect the stability and trend of the data. After alignment, the system ensures that the high-frequency feature sequence is condensed into one or a few representative values ​​and synchronized with the reference frequency.

[0035] The system uses a baseline sampling frequency as the time granularity and strictly synchronizes all upsampled and downsampled operational environment features and task-specific features according to timestamps. Finally, at each baseline sampling moment, the system outputs a complete and time-synchronized feature vector, which is used as input to the dynamic priority regression model for subsequent priority coefficient calculation.

[0036] Through the above steps, the system effectively solves the problem of time inconsistency in multi-source data, ensuring the accuracy and real-time performance of the input features of the regression model.

[0037] Step S30: Construct a topology diagram of power plant equipment, map the aligned operating environment features and task features to the nodes of the topology diagram, and use a graph neural network to extract risk propagation features between nodes, outputting the risk propagation quantification value of each task as the task priority coefficient.

[0038] In the specific implementation, a topology graph of power plant equipment is constructed. Aligned features are mapped onto the nodes of the topology graph, and a graph neural network is used to extract risk propagation features between nodes. The risk propagation quantification value of each task is output as a task priority coefficient. This includes constructing a topology graph based on the physical connections or logical control relationships of power plant equipment, where the nodes of the topology graph represent equipment or sensors, and the edges of the topology graph represent the dependencies between equipment. Aligned operating environment features and task-specific features are used as node attributes and input into the graph neural network. The graph neural network aggregates the feature information of adjacent nodes through a multi-layer message passing mechanism to simulate the propagation path and attenuation law of risk between equipment. The node embedding representation output by the graph neural network is mapped through a fully connected layer to obtain the risk propagation quantification value corresponding to each task, where the risk propagation quantification value serves as the task priority coefficient.

[0039] It should be noted that, based on the actual physical layout, piping connections, circuit topology, and logical control dependencies of the power plant, a directed or undirected graph G=(V,E) is constructed. The set of nodes V represents critical equipment, sensors, or monitoring areas, such as steam turbines, boilers, pumps, valves, and temperature sensors. The set of edges E represents the direct connections between equipment, such as steam pipe connections, electrical lines, and control signal links. Each edge can be assigned a weight based on its actual physical distance or dependency strength to simulate the propagation and attenuation of risk.

[0040] The time-series aligned features are mapped to node attributes. At each baseline sampling time, the system merges the aligned operating environment features and task-specific features according to the device or sensor they belong to. If a node corresponds to multiple sensors, its features are concatenated or aggregated to form a fixed-dimensional feature vector. These features serve as the initial node input features for the graph neural network. For tasks that are not directly related to the current task, their features can be appended to the corresponding nodes based on the target device or region being monitored.

[0041] In this embodiment, a Graph Attention Network (GAT) is used as the basic architecture. The model contains multiple graph convolutional layers, each of which updates the representation of the central node by aggregating the features of neighboring nodes, thereby simulating the propagation and coupling of risk between devices. Specifically, the first... The formula for updating the node features of a layer is: ; in, Represents a node The set of neighboring nodes, Attention coefficient For a trainable weight matrix, This is a non-linear activation function. After... After layer message passing, each node obtains an embedded representation that incorporates global topology information. To transform node embeddings into task priority coefficients, the model appends a fully connected layer after the last layer. The priority coefficient of a task is mapped to a scalar value. If a task involves multiple nodes, such as a cross-regional inspection task, the weighted average or maximum value of the outputs of each node is used as the priority coefficient of the task. The model is trained based on historical task data. The system collects the operational environment characteristics, task characteristics, and actual security risk losses caused by task delays during historical task execution. After aligning the historical features, a topology graph is constructed and input into GAT to obtain the predicted priority coefficient. and the actual risk loss value Construct a loss function, such as mean squared error: ; The trainable parameters in the GNN are optimized using the backpropagation algorithm. After training, the model can be used for online real-time inference. During system operation, the scheduler only needs to input the aligned feature sequence obtained at each benchmark sampling time into the trained graph neural network, and the priority coefficient corresponding to each task can be obtained after forward propagation.

[0042] Through the above steps, this method deeply integrates graph neural networks with the physical topology of power plants, realizes priority calculation for risk propagation perception, overcomes the limitation of traditional regression models that ignore the coupling relationship between equipment, and provides high-quality input for subsequent heuristic scheduling.

[0043] Step S40: Dynamically sort all tasks to be processed according to the task priority coefficient, and allocate computing resources to each task according to the sorting result.

[0044] In the specific implementation, the memory capacity and computing power required for the current task are determined according to the sorting results; all available computing resources are traversed to select a set of computing resources that meet the memory capacity and computing power requirements of the task; in the set of computing resources, the computing resource with the smallest remaining memory capacity is selected for allocation; the task is assigned to the selected computing resource, and the corresponding task adapter weights and basic large model are loaded.

[0045] It should be noted that resource allocation involves searching for the optimal resource matching and preemption decision scheme. This includes treating all currently running and waiting tasks as scheduling objects, constructing an optimization objective function to minimize the weighted sum of risk losses for all tasks, where the weights are the priority coefficients corresponding to each task; setting constraints, including the upper limit of GPU memory capacity, the number of computing cores, task dependencies, and the context switching overhead caused by preemption operations; employing a pre-defined heuristic search strategy to search for the optimal task-resource mapping scheme in the solution space that satisfies the constraints; when the optimal scheme found contains a preemption operation, this scheme is output as the preemption decision; otherwise, the current resource allocation is maintained or a waiting strategy is executed.

[0046] Specifically, suppose there is a set of running tasks in the system at the current moment. and waiting task set The total set of tasks is Each task It possesses resource attributes such as priority level, required GPU memory capacity, number of computing cores, estimated execution time, and task dependencies. Let the set of available computing resources be... Each resource Has a maximum video memory capacity and the upper limit of the number of computing cores Define decision variables. Indicates task Is it assigned to a resource? ,as well as Indicates task Whether it has been seized.

[0047] Step S50: When the first priority task requires computing resources and the currently available resources are insufficient, assess the feasibility of preempting the allocated resources for the second priority task.

[0048] In this embodiment, the feasibility assessment specifically includes calculating the delay risk loss caused by the preemption of the second priority task; calculating the security risk reduction benefit obtained by immediately executing the first priority task; and determining whether the security risk reduction benefit is greater than the delay risk loss. If the security risk reduction benefit is greater than the delay risk loss, the assessment is determined to be successful.

[0049] It should be noted that the objective function is to minimize the weighted sum of risk losses for all tasks, where the weights are priority coefficients, and the additional risk loss caused by preemption is also considered. ; in For the task The estimated delay from the current moment to actual completion, assuming the task is already running and not preempted, then ); The time overhead of context switching caused by preemptive operations; This is an adjustable coefficient used to balance the impact of direct delay and preemption overhead on risk.

[0050] The constraints include: 1. Resource capacity constraint: For each resource, the total video memory allocated to all tasks cannot exceed [a certain limit]. The total number of cores calculated does not exceed 2. Task dependency constraints: If the task... Task-dependent Then the task The start time must be later than the task. 3. Preemption Consistency Constraint: A preempted task must be paused and its resources released, and it must be rescheduled afterwards.

[0051] To solve this optimization problem, this invention employs a heuristic algorithm to search for an approximate optimal solution within a finite time. This embodiment uses a genetic algorithm as the heuristic strategy, and the specific steps are as follows: Encoding: Each possible scheduling scheme is encoded as a chromosome, which includes a sequence of task-to-resource mappings and a sequence of task execution orders. For example, an integer array represents the resource number allocated to each task, and a floating-point array represents the priority weight of the tasks to determine the execution order.

[0052] Population initialization: Randomly generate several initial scheduling schemes to ensure that basic resource capacity constraints are met.

[0053] Fitness function: The fitness is the reciprocal of the objective function value; a higher fitness indicates a better solution. For solutions that do not meet dependency constraints or exceed resource limits, a penalty term is applied to lower their fitness.

[0054] Selection operation: Use roulette or tournament selection mechanism to select the best individuals from the current population as the parent generation.

[0055] Crossover operation: Perform single-point crossover on the task mapping part of the two parent chromosomes and partial mapping crossover on the execution order part to generate offspring.

[0056] Mutation operation: randomly changes the resource allocation or execution order of a task with a certain probability, introducing a new solution space.

[0057] Iteration and Termination: Repeat selection, crossover, and mutation until the preset maximum number of iterations is reached or the fitness converges, and output the scheduling scheme corresponding to the current best individual.

[0058] In addition, other heuristic algorithms can be used, such as simulated annealing, particle swarm optimization, or heuristic rules specifically designed for scheduling problems, depending on the actual needs of the scenario.

[0059] Finally, the decision scheme is output. The optimal solution obtained by the heuristic algorithm is parsed into specific scheduling instructions: if the scheme includes a preemption operation, the scheduler outputs the scheme to the resource preemption execution module; otherwise, the current resource allocation is maintained, and the waiting task will remain in the queue, or a non-preemptive resource allocation will be executed according to the scheme.

[0060] Through the above steps, the scheduler dynamically runs heuristic optimization in each decision cycle, minimizing the overall system security risk while satisfying complex constraints, and realizing intelligent scheduling decision based on risk quantification.

[0061] Step S60: If the evaluation is successful, the execution of the second priority task is paused, the computing resources of the second priority task are released, and the released computing resources are redistributed to the first priority task to achieve real-time dynamic allocation of computing power.

[0062] In the specific implementation, before preemption, the current execution state and computation context of the second priority task are saved; the execution of the second priority task is paused and the computation resources occupied by the second priority task are released; the released computation resources are allocated to the first priority task and a fast switch is performed; after the first priority task is completed, the second priority task is put back into the scheduling queue, and the current execution state of the second priority task is restored to the execution state according to the computation context saved by the second priority task.

[0063] It's important to note that after receiving a preemption command, the scheduler first locks the computing resources occupied by the preempted low-priority task. Before officially pausing the low-priority task, the system saves its current execution state and computational context, including the batch progress of model inference, the cache state of intermediate activation layers, the program counter, and memory pointers. This saving operation ensures that the low-priority task can seamlessly resume from the point of interruption in the future. Subsequently, the system immediately suspends the execution of the low-priority task and marks the computing cores and some GPU memory resources it occupies as available.

[0064] The scheduler allocates released computing resources to high-priority tasks waiting to be executed. During the rapid switching operation, the scheduler immediately unloads the task adapter weights corresponding to the preempted low-priority tasks from the GPU memory of the computing resources. Because these weights are extremely small, the unloading process is very short. Subsequently, the scheduler immediately loads the task adapter weights corresponding to the high-priority tasks into the GPU memory of the computing resources. Throughout this process, the original pre-trained parameters of the underlying basic model remain loaded in GPU memory without needing to be reloaded, thus minimizing task switching latency to the millisecond level. After acquiring resources and adapters, the high-priority tasks immediately begin inference computation.

[0065] Once a high-priority task completes its execution, the system outputs its result and releases the occupied adapter weight. The scheduler then re-enters the low-priority task into the scheduling queue, but temporarily increases its priority. When the low-priority task is selected by the scheduler again and acquires resources, the system loads its corresponding adapter weight and, based on the previously saved current execution state and computation context, resumes or continues execution from the preempted interruption point, ensuring task reliability and execution integrity.

[0066] Through the above steps, intelligent preemption decision-making based on a quantitative comparison of costs and benefits is realized. By saving the task context and quickly switching the adapter, the immediate response of high-priority tasks and the reliable recovery of preempted tasks are ensured.

[0067] Step S70: After the first priority coefficient task obtains computing resources, load the task adapter weights corresponding to the first priority coefficient task and the shared basic large model from the model library of the shared basic model into the allocated computing resources.

[0068] Step S80: Perform batch inference on multiple tasks that have been loaded with the same underlying large model.

[0069] In its implementation, the scheduler first determines the memory capacity and computing power required by the currently waiting tasks, and then precisely quantifies the memory capacity and computing power requirements as necessary conditions for resource matching.

[0070] The scheduler iterates through all currently idle or available computing resources in real time, then filters out all computing resources that can fully meet the task's memory capacity and computing power requirements, forming a set of available computing resources. Resources that do not meet any requirements are excluded. For each resource in the set of available computing resources, the scheduler calculates in real time its remaining memory capacity after allocating the task.

[0071] The scheduler selects the computing resource with the smallest remaining video memory capacity from the set of available computing resources according to the best-fit principle. This selection avoids using resources with ample video memory capacity to process small tasks, thus reserving more contiguous space for large tasks that may come later and have higher requirements for contiguous video memory, effectively suppressing video memory fragmentation.

[0072] The task is assigned to the selected best-matching computing resource. After resource allocation, the system loads the corresponding task adapter weights and the shared underlying large model into that computing resource, and the task begins execution.

[0073] This strategy allows the system to meet task performance requirements while optimizing the use of scarce video memory resources, thereby improving the throughput and resource utilization of the power plant monitoring system.

[0074] Step S90: Based on the output of batch processing inference, generate power plant safety monitoring instructions or alarm information, and execute the corresponding safety monitoring response.

[0075] It should be understood that by using the results obtained after processing the above reasoning to respond to power plant safety monitoring, the reliability and real-time performance of power plant safety monitoring can be guaranteed.

[0076] In this embodiment, a model library with a shared basic model is constructed. The collected and aligned operating environment features and task features are mapped to graph nodes in the topology graph of power plant equipment. Task priority coefficients obtained by extracting risk propagation features between nodes using a graph neural network are used to dynamically sort all tasks and allocate computing resources. When the resources of the first priority task are insufficient, resource preemption for the second priority task is evaluated and executed. After obtaining resources, the corresponding adapter weights and basic models are loaded from the model library of the basic model. Batch processing inference is performed on tasks with the same basic model. Finally, monitoring instructions are generated based on the inference results, which improves resource utilization, realizes intelligent scheduling decision-making based on real-time risk quantification assessment and ultra-low latency switching response, and ensures the reliability and real-time performance of power plant safety monitoring.

[0077] Furthermore, this embodiment of the invention also proposes a storage medium storing a power plant safety monitoring program based on task priority coefficients and computing power allocation. When the power plant safety monitoring program based on task priority coefficients and computing power allocation is executed by a processor, it implements the steps of the power plant safety monitoring method based on task priority coefficients and computing power allocation described above.

[0078] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the power plant safety monitoring device based on task priority coefficient and computing power allocation of the present invention.

[0079] like Figure 2 As shown, the power plant safety monitoring device based on task priority coefficient and computing power allocation proposed in this embodiment of the invention includes: The task weight allocation module 10 is used to train and generate task adapter weights based on the power plant private domain security dataset for the basic large model, so as to build a model library with shared basic models. The time series alignment module 20 is used to collect the operating environment characteristics and task characteristics of the power plant in real time and perform time series alignment. The priority coefficient calculation module 30 is used to construct a topology diagram of power plant equipment, map the aligned operating environment features and task features to the nodes of the topology diagram, and use a graph neural network to extract the risk propagation features between nodes, and output the risk propagation quantification value of each task as the task priority coefficient. The computing resource allocation module 40 is used to dynamically sort all tasks to be processed according to the task priority coefficient, and allocate computing resources to each task according to the sorting result. The resource preemption assessment module 50 is used to assess the feasibility of preempting the allocated resources of a second priority task when a first priority task needs to compute resources and the currently available resources are insufficient. The priority of the second priority task is lower than that of the first priority task. The resource preemption execution module 60 is used to pause the execution of the second priority task if the evaluation is passed, release the computing resources of the second priority task, and redistribute the released computing resources to the first priority task to achieve real-time dynamic allocation of computing power. The power plant model loading module 70 is used to load the task adapter weights and the shared basic large model corresponding to the first priority task into the allocated computing resources after the first priority task obtains computing resources; The task batch processing module 80 is used to perform batch inference on multiple tasks that have been loaded and share the same basic large model; The power plant safety monitoring module 90 is used to generate power plant safety monitoring instructions or alarm information based on the output results of batch processing inference, and to execute corresponding safety monitoring responses.

[0080] In this embodiment, a model library with a shared basic model is constructed. The collected and aligned operating environment features and task features are mapped to graph nodes in the topology graph of power plant equipment. Task priority coefficients obtained by extracting risk propagation features between nodes using a graph neural network are used to dynamically sort all tasks and allocate computing resources. When the resources of the first priority task are insufficient, resource preemption for the second priority task is evaluated and executed. After obtaining resources, the corresponding adapter weights and basic models are loaded from the model library of the basic model. Batch processing inference is performed on tasks with the same basic model. Finally, monitoring instructions are generated based on the inference results, which improves resource utilization, realizes intelligent scheduling decision-making based on real-time risk quantification assessment and ultra-low latency switching response, and ensures the reliability and real-time performance of power plant safety monitoring.

[0081] This application embodiment also provides a power plant safety monitoring device based on task priority coefficient and computing power allocation, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other through the communication bus. The memory is used to store the power plant safety monitoring program based on task priority coefficient and computing power allocation. When the processor executes the program stored in the memory, it implements the above-mentioned power plant safety monitoring method based on task priority coefficient and computing power allocation.

[0082] The communication bus mentioned in the power plant safety monitoring equipment based on task priority coefficients and computing power allocation can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0083] The communication interface is used for communication between the aforementioned power plant safety monitoring equipment based on task priority coefficients and computing power allocation and other equipment.

[0084] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0085] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0086] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0088] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0090] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0091] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0092] In addition, for technical details not described in detail in this embodiment, please refer to the power plant safety monitoring method based on task priority coefficient and computing power allocation provided in any embodiment of the present invention, which will not be repeated here.

[0093] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0094] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0095] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0096] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

[0097] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

Claims

1. A power plant safety monitoring method based on a task priority coefficient and computing power allocation, characterized in that, The power plant safety monitoring method based on task priority coefficient and computing power allocation includes: For the basic large model, task adapter weights are trained and generated based on the power plant private domain security dataset to build a model library with a shared basic model; Real-time acquisition of power plant operating environment characteristics and task-specific characteristics, followed by time series alignment; A topology graph of power plant equipment is constructed, and the aligned operating environment features and task features are mapped to the nodes of the topology graph. A graph neural network is used to extract the risk propagation features between nodes, and the risk propagation quantification value of each task is output as the task priority coefficient. All pending tasks are dynamically sorted according to the task priority coefficient, and computing resources are allocated to each task according to the sorting result. When a first priority task requires computational resources and the currently available resources are insufficient, the feasibility of preempting the allocated resources for a second priority task is evaluated, wherein the priority of the second priority task is lower than that of the first priority task. If the evaluation is passed, the execution of the second priority task is paused, the computing resources of the second priority task are released, and the released computing resources are redistributed to the first priority task to achieve real-time dynamic allocation of computing power. After the first priority task obtains computing resources, the task adapter weights corresponding to the first priority task and the shared basic large model are loaded from the model library of the shared basic model into the allocated computing resources. Perform batch inference on multiple tasks that have already loaded and share the same basic large model; Based on the output of batch processing inference, generate power plant safety monitoring instructions or alarm information, and execute the corresponding safety monitoring response.

2. The power plant safety monitoring method based on task priority coefficient and computing power allocation as described in claim 1, characterized in that, The aforementioned method involves training and generating task adapter weights based on a power plant private domain security dataset to construct a model library with a shared base model, including: Freeze all original pre-trained parameters of the base model; A low-rank adaptation technique is used to inject new trainable low-rank matrices into the preset layers of the basic large model; Based on the power plant private domain security dataset, only the new trainable low-rank matrix is ​​trained, and the trained and optimized low-rank matrix is ​​stored as task adapter weights. The task adapter weights are independent of the original pre-trained parameters of the base model and together with the base model constitute a shared model library.

3. The power plant safety monitoring method based on task priority coefficient and computing power allocation as described in claim 1, characterized in that, The time series alignment includes: Based on the real-time requirements of the task itself, a unified reference sampling frequency is set; For feature sequences with sampling frequencies lower than the reference sampling frequency, an interpolation method is used to align the operating environment features and the task's own features to the reference sampling frequency; For feature sequences with a sampling frequency higher than the benchmark sampling frequency, a time window aggregation method is used to align the operating environment features and the task's own features to the benchmark sampling frequency; Using the reference sampling frequency as the time granularity, the characteristics of the operating environment and the characteristics of the task itself are synchronized and aligned according to the timestamp.

4. The power plant safety monitoring method based on task priority coefficient and computing power allocation as described in claim 1, characterized in that, The process involves constructing a topology graph of power plant equipment, mapping aligned features onto nodes in the topology graph, and using a graph neural network to extract risk propagation features between nodes. The resulting quantified risk propagation value for each task is output as a task priority coefficient. This includes: Based on the physical connection or logical control relationship of power plant equipment, a topology graph is constructed, wherein the nodes of the topology graph represent equipment or sensors, and the edges of the topology graph represent the association and dependency between equipment. The aligned operating environment features and task features are used as node attributes and input into the graph neural network. The graph neural network aggregates the feature information of adjacent nodes through a multi-layer message passing mechanism to simulate the propagation path and attenuation law of risk between devices. The node embedding representations output by the graph neural network are mapped through a fully connected layer to obtain the risk propagation quantification value corresponding to each task, wherein the risk propagation quantification value serves as the priority coefficient of the task.

5. The power plant safety monitoring method based on task priority coefficient and computing power allocation as described in claim 1, characterized in that, The process of allocating computing resources to each task according to the sorting results includes: The required video memory and computing power for the current task are determined based on the sorting results. Traverse all available computing resources and filter out the set of computing resources that meet the memory capacity and computing power requirements of the task. From the set of computing resources, the computing resource with the smallest remaining video memory capacity is selected for allocation; The task is assigned to the selected computing resources, and the corresponding task adapter weights and basic large model are loaded.

6. The power plant safety monitoring method based on task priority coefficient and computing power allocation as described in claim 1, characterized in that, The feasibility assessment of the second priority task for seizing allocated resources includes: Calculate the delay risk loss caused by the preemption of the second priority task; Calculate the security risk reduction benefit obtained by immediately executing the task with the first priority coefficient; Determine whether the benefits of reducing the security risk outweigh the losses from the delay risk. If the benefits of reducing the security risk outweigh the losses from the delay risk, then the assessment is deemed successful.

7. The power plant safety monitoring method based on task priority coefficient and computing power allocation as described in claim 1, characterized in that, The step of pausing the execution of the second priority task, releasing the computing resources of the second priority task, and reallocating the released computing resources to the first priority task includes: Before preemption, save the current execution state and computation context of the task with the second priority coefficient; Suspend the execution of the second priority task and release the computing resources occupied by the second priority task; The freed computing resources are allocated to the task with the first priority coefficient, and a fast switch is performed; After the first priority task is completed, the second priority task is put back into the scheduling queue, and the current execution state of the second priority task is restored to the execution state according to the calculation context saved by the second priority task.

8. A power plant safety monitoring device based on task priority coefficient and computing power allocation, characterized in that, The power plant safety monitoring device based on task priority coefficient and computing power allocation is applied to the power plant safety monitoring method based on task priority coefficient and computing power allocation as described in any one of claims 1 to 7, wherein the device comprises: The task weight allocation module is used to train and generate task adapter weights based on the power plant private domain security dataset for the basic large model, so as to build a model library with a shared basic model. The time series alignment module is used to collect the operating environment characteristics and task characteristics of the power plant in real time and perform time series alignment. The priority coefficient calculation module is used to construct a topology graph of power plant equipment, map the aligned operating environment features and task features to the nodes of the topology graph, and use a graph neural network to extract the risk propagation features between nodes, and output the risk propagation quantification value of each task as the task priority coefficient. The computing resource allocation module is used to dynamically sort all pending tasks according to the task priority coefficient, and allocate computing resources to each task according to the sorting result. The resource preemption assessment module is used to assess the feasibility of preempting the allocated resources of a second priority task when a first priority task needs to compute resources and the currently available resources are insufficient. The priority of the second priority task is lower than that of the first priority task. The resource preemption execution module is used to pause the execution of the second priority task if the evaluation is passed, release the computing resources of the second priority task, and redistribute the released computing resources to the first priority task to achieve real-time dynamic allocation of computing power. The power plant model loading module is used to load the task adapter weights and the shared basic large model corresponding to the first priority task into the allocated computing resources after the first priority task obtains computing resources. The task batch processing module is used to perform batch inference on multiple tasks that have been loaded and share the same basic large model; The power plant safety monitoring module is used to generate power plant safety monitoring instructions or alarm information based on the output results of batch processing inference, and to execute corresponding safety monitoring responses.

9. A power plant safety monitoring device based on task priority coefficient and computing power allocation, characterized in that, The power plant safety monitoring device based on task priority coefficient and computing power allocation includes: a memory, a processor, and a power plant safety monitoring program based on task priority coefficient and computing power allocation stored in the memory and executable on the processor. The power plant safety monitoring program based on task priority coefficient and computing power allocation is configured to implement the steps of the power plant safety monitoring method based on task priority coefficient and computing power allocation as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a power plant safety monitoring program based on task priority coefficients and computing power allocation. When the power plant safety monitoring program based on task priority coefficients and computing power allocation is executed by the processor, it implements the steps of the power plant safety monitoring method based on task priority coefficients and computing power allocation as described in any one of claims 1 to 7.

Images