A training and deployment integrated cross-end model adaptive deployment method, system and medium

Abstract

This invention relates to the field of intelligent inspection and edge computing technology in the power industry, and particularly to a cross-terminal model adaptive deployment method, system, and medium integrating training and inference. The method includes the following steps: constructing multi-terminal deployment environment parameters for a power plant; generating edge-aware training constraint vectors; performing integrated training and inference constraint training; cross-terminal deployability analysis of the model structure; adaptive reconstruction and optimization of the model structure; cross-terminal model deployment and inference execution; and inference performance feedback and closed-loop optimization. By introducing edge controller resource constraints and business characteristics from the power plant construction and operation environment into the model training phase, automatic adaptation and optimization are achieved in the model deployment and inference phases, thereby improving the deployment efficiency, operational stability, and inference performance of the model in the multi-terminal environment of the power plant. This invention, combined with the heterogeneous deployment characteristics of multiple edge controllers in the field, realizes the collaborative design of model training and inference deployment, and has good engineering feasibility and industry application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection and edge computing technology in the power industry, and in particular to a cross-end model adaptive deployment method, system and medium that integrates training and push. Background Technology

[0002] With the expansion of power plant construction and the improvement of operational management sophistication, intelligent models based on image and data analysis are gradually being applied to business scenarios such as equipment hazard identification, construction safety supervision, and operational risk monitoring. These applications are typically deployed on-site at power plants, involving various heterogeneous computing environments, including station-side industrial control equipment, edge controllers, and central computing platforms.

[0003] In existing technologies, relevant models are mostly trained at the central site and then adapted to different types of edge controllers or station-level industrial control equipment at the power plant for inference. This approach typically requires repeated model trimming, quantization, format conversion, or recompilation for different devices, which presents the following problems: 1. The model training phase did not fully consider the constraints of the power plant's on-site edge controller in terms of computing power, memory, and real-time performance, resulting in great difficulty in deploying and adapting the model on-site; 2. The manual adaptation of different models of edge controllers involves a large workload and a long deployment cycle, making it difficult to meet the timeliness requirements of hidden danger investigation. 3. The inference performance is not compatible with the on-site operating environment of the power plant, affecting the stability and continuity of hazard identification; 4. Frequent model updates require repeated multi-platform adaptation and compilation, resulting in high maintenance costs and hindering large-scale promotion and application.

[0004] Therefore, the existing cross-end model deployment method adopts the mode of training on the central side and then adapting to various types of edge devices in the power plant to perform inference. This method has the disadvantages of high adaptation difficulty, long deployment cycle, poor stability and continuity, and high maintenance cost. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-platform model adaptive deployment method, system, and medium that integrates training and push, which can solve the technical problems of existing cross-platform model deployment methods, such as high adaptation difficulty, long deployment cycle, poor stability and continuity, and high maintenance cost.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention designs a cross-platform adaptive deployment method for training and push integration, comprising the following steps: Construct multi-terminal deployment environment parameters for power plants; collect and construct deployment environment parameters at the construction or operation sites of power plants, where the deployment terminals include the central computing platform, station-side industrial control equipment, and field edge controllers; Generate edge-aware training constraint vectors; transform the operational constraints of different deployment ends into structured constraint information that can participate in model training, and use it to guide the model to adapt to the resource conditions of the power plant field edge controller during the training phase. Perform integrated training and push constraint training; introduce edge perception training constraint vectors into the model training process to constrain the model parameter update process, so that the trained model can meet the requirements of hazard identification accuracy, adapt to the computing power, memory and real-time limitations of different field edge controllers in power plants, and form candidate models suitable for different inference scenarios. Cross-end deployability analysis of model structure; after obtaining candidate models suitable for different inference scenarios, before the model is actually deployed to the power plant field edge controller, the candidate model is subjected to model structure parsing, and combined with the operator support capabilities and runtime resource limitations of the target deployment end, the model is subjected to cross-end deployability analysis to generate model deployability evaluation results; Adaptive model structure reconstruction and optimization; Based on the model deployability evaluation results, adaptive model structure reconstruction and optimization are performed on the candidate models to generate deployment models that match the computing power, memory resources and real-time requirements of the target deployment end; Cross-end model deployment and inference execution; after generating the deployment model, the deployment model is sent to the station-end industrial control equipment or field edge controllers at the power plant site for inference execution, and the model execution strategy is dynamically adjusted according to the equipment status and environmental changes during operation, so as to realize the adaptive operation of the model in the multi-edge controller environment of the power plant and ensure the real-time performance and operational stability of the hidden danger investigation task; Inference performance feedback and closed-loop optimization: Collect inference performance data and hazard identification results of the model during its operation at the power plant site, and feed the inference performance data back to the training or deployment stage for subsequent model training updates or adaptive optimization of deployment parameters.

[0007] As a preferred approach, the generation of end-aware training constraint vectors includes the following sub-steps: Extract edge controller resource constraint parameters; extract the runtime resource parameters of each target deployment terminal from the deployment environment parameters, and construct a set of deployment terminal resource constraints using the runtime resource parameters; Resource constraint standardization: Standardize the resource constraint parameters on the deployment end so that they can participate in unified training constraint calculation; Construct an edge-aware training constraint vector; based on the standardized deployment end resource constraint parameters, construct an edge-aware training constraint vector to represent the operational constraints of the target deployment end during the model training phase; The constraint vector is bound to the training task; the edge-aware training constraint vector is bound to the training task and used as one of the input conditions in the model training process; during the training process, the edge-aware training constraint vector is used to: limit the complexity of the model structure, control the scale of the model parameters, constrain the computational load of inference, and adjust the network structure search or model parameter update strategy so that the model training results can meet the operation requirements of the power plant field edge controller. Generate a set of deployment constraints; for different types of field edge controllers, generate different edge-aware training constraint vectors and form a set of deployment constraints. This set of deployment constraints is used in the subsequent training and push-integrated constraint training stage to generate candidate models that can be adapted to different deployment ends.

[0008] As a preferred approach, integrated training and push constraint training is implemented, and the specific process includes the following sub-steps: Construct a model training input with constraints; use the generated end-aware training constraint vector and training samples as the model training input, where: the training samples are used to characterize the equipment status or potential hazards in the power plant inspection scenario, and the end-aware training constraint vector is used to characterize the resource constraints of the target deployment end; during the training process, the model dynamically adjusts the network structure selection or computation path according to different constraint vectors, so that the model can perceive the resource environment of the target deployment end during the training phase; A comprehensive training objective function for constraint awareness is constructed. During model training, a resource constraint term is introduced on the basis of the original task loss function to construct a comprehensive training objective function, so that the model can meet deployment constraints while optimizing recognition performance. By weighting and combining the resource constraint terms, the model training process can simultaneously meet the requirements of recognition accuracy and deployment resources. The model parameters are updated with constraint awareness. During training, the model parameters are iteratively updated according to the comprehensive training objective function through the backpropagation algorithm, so that the model gradually converges to the parameter state that simultaneously satisfies the following conditions: achieving the preset hazard identification accuracy on the training dataset, the computational resource limit of the target edge controller, and the real-time requirements of the target inference scenario. Generate multi-scenario candidate models; during the training process, train or sample different edge perception training constraint vectors to obtain multiple model instances that meet different deployment constraints, forming a multi-scenario candidate model set, which are adapted to different types of power plant field edge controllers, so that the model can be adapted to different deployment environments during the training stage, reducing the subsequent model compression, pruning or retraining process, and realizing cross-end model adaptation capability of training and push integration. The model structure is adaptively selected or adjusted based on constraint vectors. During model training, the model structure is adaptively selected or adjusted according to the input end-aware training constraint vectors to match the resource constraints of the target deployment end. The model completes the structural adaptation for different deployment environments during the training phase, reducing the subsequent model compression or retraining process and realizing the cross-end model adaptive capability of training and push integration.

[0009] As a preferred approach, adaptive selection or adjustment of the model structure is performed based on constraint vectors, specifically including: Construct a candidate model structure space; predefine multiple model structures or computational module combinations of different complexities to form a candidate model structure space; Perform structure search or path selection based on constraint vectors; select model structures that meet resource constraints from the candidate model structure space based on the computing power, memory and inference latency limitations contained in the end-aware training constraint vectors. Joint training and optimization are performed on the selected structure; the parameters of the selected model structure are trained to achieve better hazard identification performance while meeting resource constraints.

[0010] As a preferred approach, the cross-platform deployability analysis of the model structure includes the following sub-steps: Analyze the computational structure of the candidate model; by analyzing the computational structure of the candidate model, obtain the complete operator sequence and topology of the model, which will be used for subsequent deployment and adaptation analysis. Obtain the operator support rules for the target edge controller; based on the operating environment of different edge controllers in the power plant, classify and build the operator support rule library for the target deployment end, specifically including: operator type support list, operator implementation method, and operator operation restriction conditions; Perform operator-level deployment adaptation analysis; based on the model operator sequence and operator support rule base, perform operator-level deployment adaptation analysis on candidate models to identify operators or substructures in the model that do not meet the operating conditions of the target deployment end; Perform model substructure deployment adaptation analysis; based on operator-level deployment adaptation analysis, perform substructure analysis on the set of operators with dependencies in the model to identify model substructures that do not meet the deployment conditions; for model substructures with non-deployable operators, determine whether the substructure can be adapted by operator replacement, structure pruning, or computation path adjustment; if it cannot be adapted, mark the model substructure as a non-deployable structure. Generate model deployability evaluation results; based on the results of operator-level deployment adaptation analysis and model substructure deployment adaptation analysis, generate model deployability evaluation results for candidate models; Generate model conversion strategies; based on the model deployability assessment results, automatically generate model structure adaptation strategies, including operator replacement strategies, structure pruning strategies, and computation path adjustment strategies, for subsequent cross-platform model deployment optimization.

[0011] As a preferred approach, the model structure is adaptively reconstructed and optimized, and the specific process includes the following sub-steps: Determine the model structure optimization strategy; based on the model deployability evaluation results, determine the optimization strategy of the candidate model on the target deployment end, specifically including operator structure optimization strategy, model substructure replacement strategy, model compression strategy and inference execution optimization strategy; Perform operator-level structure optimization; based on the operator structure optimization strategy, specifically including operator fusion and operator splitting, adjust the operator structure in the candidate model to adapt to the operator types supported by the target edge controller. By fusing or splitting operators, the model operators can be correctly executed on the inference engine of the target edge controller. Perform model substructure replacement; when there is a model substructure in the candidate model that does not meet the deployment resource limit or operator support condition, perform structural replacement on the model substructure. The structural replacement includes: feature extraction module replacement, feature fusion structure replacement, detection or classification head replacement. Through model substructure replacement, the overall computational complexity of the model meets the resource limit of the target deployment end. Perform model compression optimization; while ensuring that the accuracy of hazard identification meets business requirements, perform model compression optimization on candidate models, specifically including model quantization and model pruning, to reduce the scale of model parameters, computational overhead and memory usage; Optimize the inference execution strategy; after completing the model structure optimization, generate a deployment model that matches the target deployment end for subsequent cross-end model deployment and inference execution.

[0012] As a preferred solution, cross-platform model deployment and inference execution specifically includes the following sub-steps: Deployment model distribution and loading; Distribute the deployment model according to the target deployment terminal type and load it into the corresponding device runtime environment; Cross-device collaborative reasoning execution; when performing hidden danger investigation reasoning tasks at the power plant site, different deployment terminals work together, triggering cross-device collaborative reasoning processes when necessary; Device operation status perception; during inference execution, real-time collection of operation status information of the target deployment end; The inference execution strategy is dynamically adjusted; the model inference execution method is dynamically adjusted according to the device's operating status, including: model switching adjustment, inference frequency adjustment, computation path adjustment, and cross-platform task migration. Output the reasoning results; output the hazard identification results, including: hazard category, risk level, target location or area, timestamp and equipment identification.

[0013] As a preferred approach, inference performance feedback and closed-loop optimization specifically include the following sub-steps: Multi-dimensional inference performance data collection; collecting multi-dimensional inference performance data during model operation, including inference performance metrics, inference quality metrics, and environmental context information; Data backhaul and aggregation: The collected multi-dimensional inference performance data is backhauled from the station-side industrial control equipment or field edge controller to the central computing platform for unified aggregation and storage; Inference performance evaluation and analysis; Based on the collected multi-dimensional inference performance data, the model's operation on different deployment terminals is evaluated, including model accuracy and business effectiveness, whether the inference performance meets real-time requirements, whether resource utilization is reasonable, and performance differences between different devices. Generate optimization strategies; based on the evaluation results, generate model optimization strategies, specifically including model structure optimization strategies, deployment optimization strategies, and inference optimization strategies; Closed-loop feedback is provided to the training and deployment phases; the model optimization strategy is fed back to the model training phase, the model deployment optimization phase, and the inference execution phase respectively. By constructing an inference performance feedback mechanism based on power plant field operation data, closed-loop optimization of model training, deployment, and operation is achieved, enabling the model to continuously adapt to the operating environment of different edge controllers.

[0014] This invention also designs a cross-platform model adaptive deployment system integrating training and push, comprising the following modules: The environmental parameter module is used to construct multi-terminal deployment environmental parameters for power plants; it collects and constructs deployment environmental parameters at the construction or operation sites of power plants, where the deployment terminals include the central computing platform, station-side industrial control equipment, and field edge controllers; The constraint vector module is used to generate edge-aware training constraint vectors; it transforms the operational constraints of different deployment ends into structured constraint information that can participate in model training, and guides the model to adapt to the resource conditions of the power plant's field edge controller during the training phase. The training and inference module is used to perform integrated training and inference constraint training. It introduces the edge perception training constraint vector into the model training process to constrain the model parameter update process. This enables the trained model to meet the requirements of hazard identification accuracy while adapting to the computing power, memory and real-time limitations of different field edge controllers in power plants, forming candidate models suitable for different inference scenarios. The deployment analysis module is used for cross-end deployability analysis of model structure. After obtaining candidate models suitable for different inference scenarios, before the model is actually deployed to the power plant field edge controller, the candidate model is subjected to model structure parsing. Combined with the operator support capabilities and runtime resource limitations of the target deployment end, the model is subjected to cross-end deployability analysis to generate model deployability evaluation results. The reconstruction and optimization module is used for adaptive reconstruction and optimization of the model structure. Based on the model deployability evaluation results, it performs adaptive reconstruction and optimization of the model structure on the candidate model to generate a deployment model that matches the computing power, memory resources and real-time requirements of the target deployment end. The deployment and inference module is used for cross-end model deployment and inference execution. After generating the deployment model, the deployment model is sent to the station-end industrial control equipment or field edge controllers at the power plant site for inference execution. During operation, the model execution strategy is dynamically adjusted according to the equipment status and environmental changes to achieve adaptive operation of the model in the multi-edge controller environment of the power plant, ensuring the real-time performance and operational stability of the hidden danger investigation task. The feedback and optimization module is used for inference performance feedback and closed-loop optimization; it collects inference performance data and hazard identification results of the model during its operation at the power plant site, and feeds the inference performance data back to the training or deployment stage for subsequent model training updates or adaptive optimization of deployment parameters.

[0015] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the aforementioned training and push integrated cross-end model adaptive deployment method.

[0016] The beneficial effects of this invention are: This invention provides a cross-terminal adaptive model deployment method, system, and medium integrating training and inference. By introducing edge controller resource constraints and business characteristics from the power plant construction and operation environment into the model training phase, it achieves automatic adaptation and optimization during the model deployment and inference phases, thereby improving the deployment efficiency, operational stability, and inference performance of the model in the multi-terminal environment of a power plant. This invention is designed for power plant construction and operation environments, fully considering the heterogeneous deployment characteristics of multiple edge controllers in the field, and realizes a collaborative design for model training and inference deployment, possessing good engineering feasibility and industry application value.

[0017] 1. Through an integrated training and push design, during the training phase (S3), the resource constraint term in the loss function is a dynamic value randomly sampled from the temperature trajectory in each training iteration. This allows the model to naturally learn robustness to hardware state fluctuations during training. This enables the model to be deployable across multiple power plant sites during the training phase, reducing the difficulty of on-site adaptation.

[0018] 2. Supports rapid deployment of the model on power plant station-side industrial control equipment and various edge controllers, significantly shortening the online cycle of the hidden danger investigation system.

[0019] 3. Improve the inference stability and real-time performance of the model during its operation at the power plant site to meet the requirements of continuous operation.

[0020] 4. Through the inference performance feedback mechanism, continuous optimization and evolution of the model are achieved, reducing long-term operation and maintenance costs. In the inference phase (S6), the real-time temperature data collected by the equipment status perception module directly drives the constraint vector update, triggering model switching or calculation path adjustment. The robust training in the training phase and the dynamic scheduling in the inference phase thus achieve deep consistency through the same thermodynamic model. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the deployment method of the present invention.

[0022] Figure 2 A schematic diagram of the multi-edge controller deployment architecture in the construction and operation environment of a power plant.

[0023] Figure 3 This is a schematic diagram illustrating the construction of end-sensing training constraint vectors.

[0024] Figure 4 This is a schematic diagram illustrating the cross-platform deployability analysis of candidate models.

[0025] Figure 5 This is a schematic diagram of adaptive reconstruction and optimization of the model structure.

[0026] Figure 6 A schematic diagram of collaborative reasoning between large and small models in a power plant hazard investigation scenario.

[0027] Figure 7 This is a schematic diagram of the inference performance feedback and closed-loop optimization process. Detailed Implementation

[0028] To make the technical problems solved by the present invention, the technical solutions adopted, and the technical effects achieved clearer, the technical solutions of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and not for limiting the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0029] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0030] This invention relates to the field of intelligent inspection and edge computing technology in the power industry, and particularly to an integrated training and inference cross-terminal model adaptive deployment method applied to power plant construction and operation environments. This method is suitable for the rapid deployment and stable operation of models in scenarios such as power plant equipment hazard investigation, safety risk identification, work area monitoring, and operational status analysis. The purpose of this invention is to provide an integrated training and inference cross-terminal model adaptive deployment method. By introducing edge controller resource constraints and business characteristics from the power plant construction and operation environment into the model training phase, automatic adaptation and optimization are achieved in the model deployment and inference phases. This realizes an integrated design of the training and inference phases (i.e., integrated training and inference), thereby improving the deployment efficiency, operational stability, and inference performance of the model in multi-terminal power plant environments. The integrated training and inference cross-terminal model adaptive deployment method proposed in this invention is designed for power plant construction and operation environments, fully combining the heterogeneous deployment characteristics of multiple edge controllers in the field, and realizing a collaborative design for model training and inference deployment. It has good engineering feasibility and industry application value.

[0031] This invention provides a cross-platform adaptive deployment method for training and propelling models, comprising the following steps: Step S1: Construct multi-terminal deployment environment parameters for the power plant; collect and construct deployment environment parameters at the power plant construction or operation site, where the deployment terminals include the central computing platform, station-side industrial control equipment, and field edge controllers; Step S2: Generate edge-aware training constraint vectors; transform the operational constraints of different deployment ends into structured constraint information that can participate in model training, which is used to guide the model to adapt to the resource conditions of the power plant field edge controller during the training phase. Step S3: Perform integrated training and inference constraint training; introduce the edge perception training constraint vector into the model training process to constrain the model parameter update process, so that the trained model can meet the requirements of hazard identification accuracy while adapting to the computing power, memory and real-time limitations of different field edge controllers in power plants, forming candidate models suitable for different inference scenarios; Step S4: Cross-end deployability analysis of model structure; After obtaining candidate models suitable for different inference scenarios, before the model is actually deployed to the power plant field edge controller, perform model structure parsing on the candidate model, and combine the operator support capabilities and runtime resource limitations of the target deployment end to perform cross-end deployability analysis on the model and generate model deployability evaluation results. Step S5: Adaptive Restructuring and Optimization of Model Structure; Based on the model deployability evaluation results, adaptive restructuring and optimization of the model structure are performed on the candidate models to generate a deployment model that matches the computing power, memory resources and real-time requirements of the target deployment end; Step S6: Cross-end model deployment and inference execution; After generating the deployment model, the deployment model is sent to the station-end industrial control equipment or field edge controller at the power plant site for inference execution, and the model execution strategy is dynamically adjusted according to the equipment status and environmental changes during operation to achieve adaptive operation of the model in the multi-edge controller environment of the power plant, ensuring the real-time performance and operational stability of the hidden danger investigation task; Step S7: Inference performance feedback and closed-loop optimization; collect inference performance data and hazard identification results of the model during its operation at the power plant site, and feed the inference performance data back to the training or deployment stage for subsequent model training updates or adaptive optimization of deployment parameters.

[0032] It should be understood that the specific order or hierarchy of steps in the process disclosed in this invention is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process may be rearranged without departing from the scope of this disclosure. The appended method claims provide elements of various steps in an exemplary order and are not intended to limit the specific order or hierarchy described.

[0033] This invention also provides a cross-platform adaptive deployment system for training and pushing, comprising the following modules: The environmental parameter module is used to construct multi-terminal deployment environmental parameters for power plants; it collects and constructs deployment environmental parameters at the construction or operation sites of power plants, where the deployment terminals include the central computing platform, station-side industrial control equipment, and field edge controllers; The constraint vector module is used to generate edge-aware training constraint vectors; it transforms the operational constraints of different deployment ends into structured constraint information that can participate in model training, and guides the model to adapt to the resource conditions of the power plant's field edge controller during the training phase. The training and inference module is used to perform integrated training and inference constraint training. It introduces the edge perception training constraint vector into the model training process to constrain the model parameter update process. This enables the trained model to meet the requirements of hazard identification accuracy while adapting to the computing power, memory and real-time limitations of different field edge controllers in power plants, forming candidate models suitable for different inference scenarios. The deployment analysis module is used for cross-end deployability analysis of model structure. After obtaining candidate models suitable for different inference scenarios, before the model is actually deployed to the power plant field edge controller, the candidate model is subjected to model structure parsing. Combined with the operator support capabilities and runtime resource limitations of the target deployment end, the model is subjected to cross-end deployability analysis to generate model deployability evaluation results. The reconstruction and optimization module is used for adaptive reconstruction and optimization of the model structure. Based on the model deployability evaluation results, it performs adaptive reconstruction and optimization of the model structure on the candidate model to generate a deployment model that matches the computing power, memory resources and real-time requirements of the target deployment end. The deployment and inference module is used for cross-end model deployment and inference execution. After generating the deployment model, the deployment model is sent to the station-end industrial control equipment or field edge controllers at the power plant site for inference execution. During operation, the model execution strategy is dynamically adjusted according to the equipment status and environmental changes to achieve adaptive operation of the model in the multi-edge controller environment of the power plant, ensuring the real-time performance and operational stability of the hidden danger investigation task. The feedback and optimization module is used for inference performance feedback and closed-loop optimization; it collects inference performance data and hazard identification results of the model during its operation at the power plant site, and feeds the inference performance data back to the training or deployment stage for subsequent model training updates or adaptive optimization of deployment parameters.

[0034] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-described training and push integrated cross-end model adaptive deployment method.

[0035] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce implementations of the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0037] Figure 1 This is a schematic diagram of the overall process of the cross-end model adaptive deployment method integrating training and inference according to the present invention, which is used to illustrate the overall process of model from end-to-end perceptual constraint training to cross-end deployment and inference feedback.

[0038] Figure 2This diagram illustrates the deployment architecture of multiple edge controllers in a power plant construction and operation environment, showcasing the collaborative relationship between station-side industrial control equipment, field edge controllers, and the central computing platform. The overall architecture adopts a top-down "three-layer collaborative architecture," with the core functional modules as follows: Top-level core module: Central computing platform, used for model training and model lifecycle management; The core module of the middle layer is the station-side industrial control equipment, which is used for hazard identification and collaborative reasoning, and completes data interaction with the top-level central computing platform through "model distribution / operational feedback". The core underlying module, the field edge controller, is used for real-time data monitoring and preliminary screening, and interacts with the intermediate layer station-end industrial control equipment through "data acquisition / inference collaboration".

[0039] The "three-tier collaborative architecture" is specifically divided into two links: upper and lower. 1. Downlink control link (central computing platform → station-side industrial control equipment → field edge controller) The central computing platform outputs three downlink commands to the left, which enter the corresponding functional modules: the first command enters the model distribution module, the second command enters the parameter update module, and the third command enters the strategy configuration module. The outputs of the above three modules are all fed into the station-end industrial control equipment module to the right; At the same time, the model distribution module outputs an additional downlink command, which is directly fed into the underlying field edge controller module; The station-side industrial control equipment outputs two downlink commands, which enter the corresponding functional modules respectively: the first command enters the inference strategy adjustment module, and the second command enters the model switching or downgrade command module. The outputs of the two modules mentioned above are both fed into the field edge controller module.

[0040] 2. Uplink feedback link (Field edge controller → Station-side industrial control equipment → Central computing platform) The field edge controller outputs three uplink data channels, which are then sent to the corresponding functional modules: the first channel is sent to the preliminary detection results module, the second channel is sent to the abnormal segment module, and the third channel is sent to the operating status module. The outputs of the above three modules are all fed into the station-side industrial control equipment module; At the same time, the on-site edge controller outputs a series of uplink data, which is directly fed into the top-level central computing platform module through the model operation feedback module. The station-end industrial control equipment outputs two uplink feedback data paths to the right, which enter the corresponding functional modules respectively: the first path enters the inference performance data module, and the second path enters the model operation feedback module; The outputs of both modules are fed upwards into the top-level central computing platform module.

[0041] Figure 3 This diagram illustrates the construction of the edge-sensing training constraint vector, used to explain how resource constraints of different edge controllers in a power plant participate in the model training process. The overall block diagram structure of the edge-sensing training constraint vector adopts a three-part structure of resource acquisition, constraint mapping, and training input.

[0042] Figure 4 This diagram illustrates the cross-platform deployability analysis of candidate models, serving to explain the model structure analysis and operator executability evaluation process.

[0043] Figure 5 This diagram illustrates the adaptive reconstruction and optimization of the model structure, used to explain the optimization process such as operator fusion, pruning, and quantization performed on the power plant's field edge controller.

[0044] Figure 6 This diagram illustrates the collaborative reasoning between lightweight and large models in a power plant hazard investigation scenario, serving to demonstrate the collaborative operation of lightweight and large models on different edge controllers.

[0045] Figure 7 This diagram illustrates the inference performance feedback and closed-loop optimization process, used to explain the feedback mechanism of field operation results on model training and deployment strategies.

[0046] There are two core modules: the upper-level core module, the edge controller; and the lower-level core module, the central computing platform. Within the edge controller, the core starting point is the inference data acquisition module; the inference data acquisition module outputs three parallel data streams: performance data, operational status data, and business result data; these three data streams are transmitted down to the central computing platform via data backhaul and aggregation links.

[0047] The central computing platform is divided into two core functional units, forming an evaluation-optimization closed loop; Unit 1: Inference Performance Evaluation and Analysis (marked with dashed box). This unit includes four core evaluation dimensions: model accuracy and business effectiveness evaluation, whether resource utilization is reasonable, whether inference performance meets real-time requirements, and performance difference analysis between different devices. Unit 2: Generating Optimization Strategies (marked with dashed boxes). Based on the evaluation and analysis results, three types of core optimization strategies are generated: model structure optimization strategy, deployment strategy optimization, and inference strategy optimization.

[0048] The central computing platform outputs three execution branches, which are respectively output to: model training stage step S3, model deployment and optimization stage step S5, and inference execution stage step S6.

[0049] The present invention provides a cross-platform adaptive deployment method for training and pushing models, comprising the following steps: Step S1: Construct the multi-terminal deployment environment parameters for the power plant; The deployment environment parameters at the power plant construction or operation site are collected and constructed. The deployment end includes at least a central computing platform, station-side industrial control equipment, and field edge controllers. The station-side industrial control equipment includes industrial control computers, embedded control units, etc. The deployment environment parameters include: 1. Computational capability parameters; 2. Storage and memory resource parameters; 3. Supported instruction sets and inference operator types; 4. Real-time performance and operational stability constraints, wherein the real-time performance constraints include inference latency requirements, and the operational stability constraints include continuous operation capability and fault tolerance requirements; 5. Identification of business type for power plant hazard investigation.

[0050] Step S2: Generate end-sensory training constraint vectors; After constructing the multi-terminal deployment environment parameters of the power plant in step S1, this step is used to transform the operational constraints of different deployment terminals into structured constraint information that can participate in model training, so as to guide the model to adapt to the resource conditions of the power plant field edge controller during the training phase.

[0051] Step S2 specifically includes the following sub-steps: S21: Extract edge controller resource constraint parameters; The operational resource parameters of each target deployment terminal are extracted from the deployment environment parameters. The deployment terminal includes station-level industrial control equipment and field edge controllers. The resource constraint parameters include at least: 1. Computing capability parameters: Characterize the processor type, computing capability level, or the amount of computation that can be performed per unit time; 2. Memory and storage resource parameters: Characterize the available memory space of the device for model loading and inference execution; 3. Inference delay constraint parameter: characterizes the maximum allowable inference response time for the hidden danger investigation task; 4. Energy consumption or operational stability constraint parameters: These represent the allowable resource consumption or operational stability requirements of the equipment under continuous operation.

[0052] The above parameters constitute a set of resource constraints for the deployment end.

[0053] S22: Standardization of resource constraints; Since the resource parameters of different types of edge controllers have different units and magnitudes, the resource constraint parameters are standardized so that they can participate in unified training constraint calculations.

[0054] Standardization processes include: 1. Normalize and map the computing power parameters; 2. Ratio the parameters of memory and storage resources; 3. Convert inference latency constraints into model complexity constraints; 4. Operational stability requirements are mapped to resource usage limits.

[0055] Through standardization, a unified resource constraint description vector is formed.

[0056] S23: Construct end-aware training constraint vectors; Based on the standardized resource constraint parameters and equipment thermodynamic state model, an edge-aware training constraint vector is constructed to represent the operational constraints of the target deployment end during the model training phase.

[0057] A lightweight device thermal state simulation module is established on the edge side. This module predicts the core temperature trajectory of the device within a future time window based on three input variables: device type, ambient temperature, and continuous operating duration, and accordingly calculates the decay curve of the actual available computing power. The edge-sensing training constraint vector is a time-varying function, which can be expressed as: in, The computational capability constraint at time t; Indicates the memory resource constraints at time t; Represents the inference delay constraint at time t; This represents the operational stability constraint at time t.

[0058] Computational capacity constraints There is a deterministic functional relationship between the core temperature and the processor frequency. When the core temperature exceeds the throttling trigger threshold, the dynamic frequency adjustment mechanism will actively reduce the processor frequency, thereby tightening the computing power constraint. This relationship can be expressed as: in, The nominal computing power constraint; The computing power attenuation coefficient corresponds to time t; The core temperature of the device at time t; The frequency reduction trigger temperature threshold; This is the maximum permissible operating temperature.

[0059] Memory resource constraints Similarly, memory preemption and system protection mechanisms under high temperatures will dynamically tighten. Different edge-sensing training constraint vectors can be generated for different types of field edge controllers.

[0060] S24: Constraint vectors are bound to training tasks; The end-aware training constraint vector is bound to the training task and used as one of the input conditions in the model training process. During training, the end-aware training constraint vector is used for: 1. Limit the complexity of the model structure; 2. Control the scale of model parameters; 3. The computational complexity of constraint reasoning; 4. Adjust the network structure search or model parameter update strategy.

[0061] The above methods enable the model training results to meet the operational requirements of the power plant's field edge controller.

[0062] S25: Generate the deployment constraint set; For the various edge controller types that may exist in power plant sites, multiple edge-aware training constraint vectors can be generated, forming a set of deployment constraint vectors: Q ={ Q 1, Q 2, …, Q n}; in: Q 1 corresponds to the first type of edge controller; Q 2 corresponds to the second type of edge controller; Q n Corresponding to the n Deployment equipment of this type.

[0063] This set of constraints is used to generate candidate models that can be adapted to different deployment ends during the subsequent training and push integration constraint training phase.

[0064] Step S3: Perform integrated training and push constraint training; After completing the construction of the edge perception training constraint vector in step S2, the edge perception training constraint vector is introduced into the model training process to constrain the model parameter update process, so that the trained model can meet the requirements of hazard identification accuracy while adapting to the computing power, memory and real-time limitations of different edge controllers in the power plant, thereby forming a candidate model suitable for different inference scenarios.

[0065] Step S3 specifically includes the following sub-steps: S31: Construct the training input for the model with constraints; The end-aware training constraint vector generated in step S2, together with the training samples, is used as the model training input, where: 1. Training samples are used to characterize the equipment status or potential hazards in power plant inspection scenarios; 2. The edge-aware training constraint vector is used to represent the resource constraints of the target deployment edge; During training, the model dynamically adjusts the network structure selection or computation path based on different constraint vectors, thereby enabling the model to perceive the resource environment of the target deployment end during the training phase.

[0066] S32: Construct a constraint-aware comprehensive training objective function; During model training, a constraint term corresponding to the resource conditions of the target deployment end is introduced on the basis of the original task loss function to construct a comprehensive training objective function, so that the model can meet the deployment constraints of the target edge controller while optimizing the performance of hazard identification.

[0067] The comprehensive training objective function includes: 1. Task identification loss term: used to measure the model's prediction error in the hazard identification task; 2. Model complexity constraints: used to limit the size of model parameters or the amount of computation; 3. Inference delay constraint: Used to constrain the inference delay of the model on the target edge controller; 4. Resource Constraints: These constraints limit the amount of memory or computing resources used during model execution.

[0068] By weighting and combining the above constraints, the model training process can simultaneously meet the requirements for recognition accuracy and deployment resources.

[0069] In one embodiment, the comprehensive training objective function can be expressed as: in, The overall training objective function; These are model parameters; For the training dataset; Let be the end-sensing training constraint vector corresponding to time t; The task identification loss term is used to measure the prediction error of the model on the hazard identification task; This is a model complexity constraint term used to limit the size of model parameters and / or the amount of computation. This is an inference latency constraint term, used to constrain the inference latency of the model on the target deployment end; This is a resource consumption constraint term used to constrain memory and / or computational resource consumption during model execution; These are runtime stability constraints used to constrain the stability requirements of the model in continuous operation scenarios. , , and These are the weight coefficients for the corresponding constraint terms.

[0070] To improve the model's adaptability to fluctuations in the target deployment end's operation, a constraint perturbation mechanism is further introduced based on the comprehensive training objective function. During each training round, a bounded perturbation is applied to the end-aware training constraint vector to simulate the constraint fluctuations caused by temperature rise, resource contention, load changes, or continuous operation of the target deployment end during actual operation.

[0071] The perturbated constraint vector can be expressed as: in, The constrained vector after perturbation; This represents the projection operator that projects the perturbation result onto the feasible region, ensuring that the perturbation constraint vector remains within the physical resource boundary and operational constraint boundary of the target deployment at time t. The corresponding feasible domain is jointly defined by the upper bound of resources, the upper bound of latency, and the stability boundary of the target deployment at time t; To constrain the disturbance.

[0072] In one embodiment, the constraint perturbation amount can be determined based on the gradient direction of the comprehensive training objective function with respect to the constraint vector, and its expression is as follows: in, To constrain the disturbance amount; The disturbance radius at time t is used to characterize the upper bound of resource fluctuations at the target deployment end in the current operating state; It is a symbolic function; Represents the overall training objective function; This represents the comprehensive training objective function with respect to the constraint vector. The gradient; the disturbance radius It is determined by the equipment thermal state model, continuous operating status, environmental parameters and / or resource fluctuation statistics.

[0073] Based on the above perturbation mechanism, the comprehensive training objective function for model training is further expressed as: in, It represents the infinite norm.

[0074] Through the aforementioned minimax optimization method, the model can maintain its predetermined hazard identification accuracy, latency performance, and resource adaptability even under the most unfavorable resource fluctuation conditions at the target deployment end. This improves the robustness and feasibility of the model when deploying various types of edge controllers in power plant sites. Unlike the input space perturbation in adversarial example training, the perturbation here serves to constrain the space, representing a robust optimization method for uncertainties in the deployment environment. The thermodynamic model provides the physical boundary for constraining fluctuations, while the adversarial perturbation mechanism trains within this physical boundary to find the worst-case scenario. This means that the perturbation radius is not an arbitrarily set hyperparameter, but is automatically calculated by the thermodynamic simulation module based on the actual operating characteristics of the equipment—equipment with poor heat dissipation naturally corresponds to a larger perturbation radius, while stable equipment corresponds to a smaller perturbation radius. This coupling allows the robustness strength during training to accurately match the actual operating characteristics of each deployment end, avoiding over-conservatism or insufficient protection. The coupling of thermodynamic dynamic constraints and adversarial perturbation training, with the thermodynamic model providing the physical boundary for constraining fluctuations and the adversarial perturbation training finding the worst-case scenario within this boundary, is indispensable.

[0075] S33: Perform constraint-aware model parameter updates; During training, the model parameters are iteratively updated using the backpropagation algorithm based on the comprehensive training objective function, causing the model to gradually converge to a parameter state that simultaneously satisfies the following conditions: 1. Achieve the preset hazard identification accuracy on the training dataset; 2. Meet the computational resource constraints of the target edge controller; 3. Meets the real-time requirements of the target reasoning scenario.

[0076] S34: Generate multi-scenario candidate models; During the training process, training or sampling is performed on different edge-aware training constraint vectors to obtain multiple model instances that meet different deployment constraints, forming a candidate model set.

[0077] The candidate models are adapted to different types of power plant field edge controllers, including but not limited to: high-computing-power edge servers; industrial-grade edge controllers; and low-computing-power embedded control devices.

[0078] By employing the above methods, the model can be adapted to different deployment environments during the training phase, thereby reducing the need for subsequent model compression, pruning, or retraining, and achieving cross-platform model adaptation capability that integrates training and promotion.

[0079] S35: Adaptive selection or adjustment of model structure based on constraint vectors; During model training, the model structure is adaptively selected or adjusted based on the input end-aware training constraint vector to match the resource constraints of the target deployment end. Specifically, this includes: 1. Construct candidate model structure space: Predefine multiple model structures or computational module combinations with different complexities to form candidate model structure space.

[0080] 2. Perform structure search or path selection based on constraint vectors; select model structures or computational paths that meet resource constraints from the candidate model structure space based on the computing power, memory, and inference latency limitations contained in the end-aware training constraint vectors.

[0081] 3. Perform joint training and optimization on the selected structure; train the parameters of the selected model structure to achieve better hazard identification performance while meeting resource constraints.

[0082] By employing the above methods, the model can complete structural adaptation for different deployment environments during the training phase, thereby reducing the subsequent model compression or retraining process and achieving cross-platform model adaptive capability that integrates training and push.

[0083] The constraint-driven model structure search process is as follows: 1. Input stage: Constraint vector. The initial input of the process is the constraint vector, which serves as the core decision constraint condition. It is input to the structure search module from top to bottom, providing the core basis for the subsequent model path adaptation and selection.

[0084] 2. Core processing stage: Structure search module. After receiving the constraint vector, the structure search module completes the model structure adaptability search based on the input constraints and outputs three differentiated model execution paths: lightweight model path, medium model path, and high-precision model path, which can cover model selection scenarios with different computing power, memory, real-time performance, and accuracy requirements.

[0085] 3. Implementation phase: Training and optimization. The outputs of the above three model paths are uniformly incorporated into the subsequent training and optimization phase to complete the model training and optimization iterations that adapt to the constraints.

[0086] Step S4: Cross-end deployability analysis of the model structure; After obtaining candidate models suitable for different inference scenarios in step S3, before the models are actually deployed to the power plant field edge controller, the candidate models are subjected to model structure parsing, and the cross-end deployability analysis is performed on the models in combination with the operator support capabilities and operating resource limitations of the target deployment end, thereby generating model deployability evaluation results.

[0087] Step S4 includes the following process: S41: Analyze the computational structure of the candidate model; The candidate model computation structure includes: model hierarchical structure, operator type corresponding to each computation node, data structure of operator input and output, and data dependency relationship between each computation node.

[0088] By analyzing the model's computational structure, the complete operator sequence and its topological structure of the model are obtained, which can be used for subsequent deployment and adaptation analysis.

[0089] S42: Obtain the operator support rules for the target edge controller; Based on the operating environment of different edge controllers at the power plant site, an operator support rule base is constructed for the target deployment end, as follows: 1. Operator type support list, representing the types of operators that the target edge controller supports executing; 2. Operator implementation method, which represents how the operator is implemented on the target device, including: native operator support, inference engine accelerated operators, and operators that need to be converted.

[0090] 3. Operator operation constraints, including input data format constraints, tensor size constraints, memory usage constraints, and inference delay constraints.

[0091] The operator support rule base is constructed according to different types of power plant edge devices, including industrial control server nodes, GPU edge computing nodes, embedded edge controllers, industrial control PLCs or lightweight control units.

[0092] S43: Perform operator-level deployment adaptation analysis; Based on the model operator sequence obtained in step S41 and the operator support rule base constructed in step S42, operator-level deployment adaptation analysis is performed on the candidate models. Specifically, this includes: 1. Match the operator types in the model one by one with the operator types supported by the target edge controller; 2. For operators that are fully supported, mark them as "operators that can be deployed directly"; 3. For operators that require conversion or replacement, mark them as "adaptable operators"; 4. Operators that are not supported by the target device and cannot be converted and implemented are marked as "non-deployable operators".

[0093] Based on the above analysis, operators or substructures in the model that do not meet the operating conditions of the target deployment end are identified.

[0094] S44: Perform deployment adaptation analysis of the model substructure; Based on operator-level deployment adaptation analysis, substructure analysis is performed on the set of operators with dependencies in the model to identify model substructures that do not meet the deployment conditions.

[0095] The model substructure includes: a feature extraction module, a multi-scale feature fusion module, and a detection or classification output module.

[0096] For substructures with non-deployable operators, determine whether the substructure can be adapted through the following methods: operator replacement, structure pruning, or computation path adjustment; if it cannot be adapted, mark the substructure as a non-deployable structure.

[0097] S45: Generate model deployability assessment results; Based on the analysis results of steps S43 and S44, deployability evaluation results are generated for the candidate models. The evaluation results include: 1. Model Deployability Identifier: Indicates whether the model meets the deployment conditions of the target edge controller; 2. Operator video report: including operators that can be deployed directly, operators that require adaptation, and operators that are not supported; 3. Deployment risk results: including resource consumption assessment, inference latency assessment, and memory usage assessment; 4. Deployment optimization suggestions: including model pruning suggestions, operator replacement suggestions, and inference engine adaptation suggestions.

[0098] S46: Generative model conversion strategy; Based on the model deployability assessment results, model structure adaptation strategies are automatically generated, including operator replacement strategies, structure pruning strategies, and computation path adjustment strategies, for subsequent cross-platform model deployment optimization.

[0099] Step S5: Adaptive reconstruction and optimization of model structure; Based on the model deployability evaluation results obtained in step S4, adaptive reconstruction and optimization of the model structure are performed on the candidate models to generate a deployment model that matches the computing power, memory resources and real-time requirements of the target deployment end.

[0100] Step S5 includes the following sub-steps: S51: Determine the model structure optimization strategy; Based on the model deployability evaluation results obtained in step S4, determine the optimization strategy for the candidate model on the target deployment end.

[0101] The optimization strategy generates an operator structure optimization strategy, a model substructure replacement strategy, a model compression strategy, and an inference execution optimization strategy based on existing information.

[0102] The available information includes: a list of compatible operators, a list of unsupported operators, model computational cost assessment results, inference latency assessment results, and device resource constraint parameters.

[0103] S52: Perform operator-level structure optimization; According to the operator structure optimization strategy, the operator structure in the candidate model is adjusted to adapt to the operator types supported by the target edge controller. The operator structure optimization strategy includes: Operator fusion combines multiple operators executed consecutively into a single operator to reduce data transmission overhead and computational latency. Examples include fusion of convolutional layers and batch normalization layers, and fusion of convolutional layers and activation functions.

[0104] Operator decomposition involves breaking down complex operators into multiple simpler operators to adapt to the basic operators supported by the target edge device. For example, large convolution kernels can be decomposed into combinations of multiple small convolution kernels, and complex attention modules can be decomposed into basic matrix calculation structures.

[0105] By fusing or splitting operators, the model operators can be correctly executed on the inference engine of the target edge controller.

[0106] S53: Perform model substructure replacement; When a candidate model contains a substructure that does not meet deployment resource constraints or operator support conditions, the substructure is replaced. This replacement includes: 1. Replace the feature extraction module by replacing the computationally complex feature extraction structure with a lightweight network structure. Specifically, the standard convolutional structure can be replaced with a depthwise separable convolutional structure, and the multi-layer convolutional module can be replaced with a lightweight feature extraction module.

[0107] 2. Feature fusion structure replacement: The multi-scale fusion module is simplified or replaced with a lightweight version.

[0108] 3. Replace the detection or classification head. Specifically, this can be done by compressing or simplifying the structure of the output detection or classification layer to reduce computational overhead.

[0109] By replacing the model's substructures, the overall computational complexity of the model is made to meet the resource constraints of the target deployment end.

[0110] S54: Perform model compression optimization; While ensuring that the accuracy of hazard identification meets business needs, model compression optimization is performed on candidate models to reduce the size of model parameters and computational overhead.

[0111] The model compression optimization includes model quantization and model pruning. Model quantization converts model parameters from a high-precision representation to a low-precision representation, including FP32 to FP16 and FP16 to INT8. Model pruning trims the model structure based on the importance of model weights, including channel pruning, convolutional kernel pruning, and network layer pruning. By quantizing and pruning the model, the size of the model parameters and memory usage can be reduced.

[0112] S55: Optimize the execution strategy for inference; After completing the above model structure optimization, a deployment model that matches the target deployment end is generated.

[0113] The deployment model meets the following conditions: the model structure only contains operators supported by the target edge controller; the model parameter scale meets the device memory limit; the model inference latency meets the real-time requirements of the power plant hidden danger inspection business; and the model accuracy meets the business requirements of the hidden danger identification task.

[0114] The deployment model is then used for cross-platform model deployment and inference execution in step S6.

[0115] Step S6: Cross-platform model deployment and inference execution; After generating the deployment model in step S5, the deployment model is sent to the station-side industrial control equipment or edge controller at the power plant site for inference. During operation, the model execution strategy is dynamically adjusted according to the equipment status and environmental changes to ensure the real-time performance and operational stability of the hidden danger investigation task.

[0116] S61: Deployment model distribution and loading; The deployment model is distributed according to the target deployment terminal type and loaded into the corresponding device operating environment. Specifically, this includes: distributing medium-scale or high-precision models to station-end industrial control equipment and lightweight models to field edge controllers; and matching the corresponding inference engine or operating framework according to the device type.

[0117] S62: Cross-platform collaborative reasoning execution; When performing hazard investigation and reasoning tasks at the power plant site, different deployment terminals work together, including edge controllers performing real-time detection and preliminary screening, and station-end industrial control equipment performing complex hazard identification or result verification; and triggering cross-terminal collaborative reasoning processes when necessary.

[0118] S63: Equipment operating status perception; During inference execution, real-time operational status information of the target deployment end is collected, including CPU / GPU load status, memory usage, inference task queue length, device temperature or operational stability, and network communication status.

[0119] S64: Dynamic adjustment of inference execution strategy; The model inference execution method is dynamically adjusted based on the device's operating status, and the adjustment includes: 1. Model switching adjustment: switch to a lightweight model under high load and switch to a high-precision model when resources are sufficient; 2. Adjust the inference frequency according to task priority or device load; 3. Adjust the computation path by enabling or disabling some model computation branches to reduce computational overhead; 4. Cross-device task migration: some inference tasks are migrated from the edge controller to the station-side industrial control equipment for execution.

[0120] S65: Output of reasoning results; Output hazard identification results, which may include: hazard category, risk level, target location or area, timestamp, and equipment identifier.

[0121] During the model inference phase, the model achieves adaptive operation in the power plant's multi-edge controller environment through equipment status perception and dynamic execution strategy adjustment, thereby ensuring the real-time performance and stability of the hidden danger investigation task.

[0122] Step S7: Inference performance feedback and closed-loop optimization; The inference performance data and hazard identification results of the model are collected during the operation of the model at the power plant site, and the data is fed back to the training or deployment stage for subsequent model training updates or adaptive optimization of deployment parameters.

[0123] S71 multi-dimensional performance data acquisition; Collect multidimensional data during model operation, including inference performance metrics (such as latency, throughput, and resource consumption), inference quality metrics (such as recognition accuracy, false alarm rate, and false negative rate), and environmental context information (such as equipment status, time period, and changes in operating conditions).

[0124] S72 data backhaul and aggregation; The collected data is transmitted from the station-side industrial control equipment or edge controller back to the central computing platform for unified aggregation and storage. S73 inference performance evaluation and analysis; Based on the collected data, the performance of the model on different deployment devices is evaluated, including the model accuracy and business effectiveness assessment, whether the inference performance meets the real-time requirements, whether the resource utilization is reasonable, and the performance difference analysis between different devices.

[0125] S74 generates optimization strategies; Based on the evaluation results, model optimization strategies are generated, including model structure optimization strategies, deployment optimization strategies, and inference optimization strategies. Model structure optimization strategy, fed back to step S5; deployment optimization strategy, used for model allocation and model version selection; inference optimization strategy, used for execution frequency and model switching rules.

[0126] The optimization strategies specifically include operator structure optimization strategies, model substructure replacement strategies, model compression strategies, and inference execution optimization strategies. S75 closed-loop feedback to the training and deployment phases; The optimization strategies are fed back to: 1. Model training phase (S3): Used to update training constraints or optimize model structure; 2. Model Deployment Optimization Phase (S5): Used to adjust model structure or compression strategy; 3. Inference Execution Phase (S6): Used to update the dynamic execution strategy; By constructing an inference performance feedback mechanism based on power plant field operation data, closed-loop optimization of model training, deployment, and operation is achieved, enabling the model to continuously adapt to the operating environment of different edge controllers.

[0127] Examples of specific application scenarios for this invention: This invention can be applied to the following power plant construction and operation scenarios, which are based on station-end industrial control equipment and field edge controllers: 1. Intelligent identification and monitoring of potential operational hazards in power plant equipment; 2. Safety risk monitoring in power plant construction and maintenance areas; 3. Real-time edge-side analysis of abnormal operating conditions in power plants; 4. Multi-terminal collaborative deployment and operation of the power plant on-site hidden danger investigation model.

[0128] In summary, the cross-terminal adaptive deployment method for training and inference proposed in this invention is designed for power plant construction and operation environments. It fully leverages the heterogeneous deployment characteristics of multiple edge controllers in the field, realizing the collaborative design of model training and inference deployment. It has good engineering feasibility and industry application value.

[0129] All other parts not described herein belong to the prior art. The embodiments described above are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A cross-platform adaptive deployment method for integrated training and push models, characterized in that, Includes the following steps: Collect and construct environmental parameters for the power plant's on-site deployment, which includes a central computing platform, station-side industrial control equipment, and on-site edge controllers. Generate edge-aware training constraint vectors to transform the operational constraints of different deployment ends into structured constraint information that can participate in model training, which is used to guide the model to adapt to the resource conditions of the power plant field edge controller during the training phase. The training and inference integration constraint training is carried out, and the edge perception training constraint vector is introduced into the model training process to constrain the model parameter update process. This enables the model to meet the requirements of hazard identification accuracy while adapting to different on-site edge controller requirements, forming candidate models suitable for different inference scenarios. After obtaining candidate models suitable for different inference scenarios, before the models are actually deployed to the power plant field edge controller, the candidate models are subjected to model structure parsing, and the cross-end deployability analysis is performed on the models in combination with the operator support capabilities and runtime resource limitations of the target deployment end, generating model deployability evaluation results; Based on the model deployability evaluation results, the candidate model undergoes adaptive reconstruction and optimization of its structure to generate a deployment model that matches the target deployment endpoint. After generating the deployment model, the deployment model is sent to the power plant field deployment terminal to perform cross-end model inference, and the model execution strategy is dynamically adjusted during the operation to achieve adaptive operation of the model in the power plant multi-edge controller environment; The inference performance data and hazard identification results of the model are collected during the operation of the model in the power plant and fed back to the training or deployment stage for subsequent model training updates or adaptive optimization of deployment parameters.

2. The cross-platform adaptive deployment method for training and push integration according to claim 1, characterized in that: The process of generating end-aware training constraint vectors includes the following sub-steps: Extract edge controller resource constraint parameters; extract the runtime resource parameters of each target deployment terminal from the deployment environment parameters, and construct a set of deployment terminal resource constraints using the runtime resource parameters; Resource constraint standardization: Standardize the resource constraint parameters on the deployment end so that they can participate in unified training constraint calculation; Construct an edge-aware training constraint vector; based on the standardized deployment end resource constraint parameters, construct an edge-aware training constraint vector to represent the operational constraints of the target deployment end during the model training phase; The constraint vector is bound to the training task; the edge-aware training constraint vector is bound to the training task and used as one of the input conditions in the model training process; during the training process, the edge-aware training constraint vector is used to: limit the complexity of the model structure, control the scale of the model parameters, constrain the computational load of inference, and adjust the network structure search or model parameter update strategy so that the model training results can meet the operation requirements of the power plant field edge controller. Generate a set of deployment constraints; for different types of field edge controllers, generate different edge-aware training constraint vectors and form a set of deployment constraints. This set of deployment constraints is used in the subsequent training and push-integrated constraint training stage to generate candidate models that can be adapted to different deployment ends.

3. The cross-platform adaptive deployment method for training and push integration according to claim 1, characterized in that: The process of implementing integrated training and push constraint training includes the following sub-steps: Construct a model training input with constraints; use the generated end-aware training constraint vector and training samples as the model training input, where: the training samples are used to characterize the equipment status or potential hazards in the power plant inspection scenario, and the end-aware training constraint vector is used to characterize the resource constraints of the target deployment end; during the training process, the model dynamically adjusts the network structure selection or computation path according to different constraint vectors, so that the model can perceive the resource environment of the target deployment end during the training phase; A comprehensive training objective function for constraint awareness is constructed. During model training, a resource constraint term is introduced on the basis of the original task loss function to construct a comprehensive training objective function, so that the model can meet deployment constraints while optimizing recognition performance. By weighting and combining the resource constraint terms, the model training process can simultaneously meet the requirements of recognition accuracy and deployment resources. The model parameters are updated with constraint awareness. During training, the model parameters are iteratively updated according to the comprehensive training objective function through the backpropagation algorithm, so that the model gradually converges to the parameter state that simultaneously satisfies the following conditions: achieving the preset hazard identification accuracy on the training dataset, the computational resource limit of the target edge controller, and the real-time requirements of the target inference scenario. Generate multi-scenario candidate models; during the training process, train or sample different edge perception training constraint vectors to obtain multiple model instances that meet different deployment constraints, forming a multi-scenario candidate model set, which are adapted to different types of power plant field edge controllers, so that the model can be adapted to different deployment environments during the training stage, reducing the subsequent model compression, pruning or retraining process, and realizing cross-end model adaptation capability of training and push integration. The model structure is adaptively selected or adjusted based on constraint vectors. During model training, the model structure is adaptively selected or adjusted according to the input end-aware training constraint vectors to match the resource constraints of the target deployment end. The model completes the structural adaptation for different deployment environments during the training phase, reducing the subsequent model compression or retraining process and realizing the cross-end model adaptive capability of training and push integration.

4. The cross-platform adaptive deployment method for training and push integration according to claim 3, characterized in that: Adaptive selection or adjustment of model structure based on constraint vectors, specifically including: Construct a candidate model structure space; predefine multiple model structures or computational module combinations of different complexities to form a candidate model structure space; Perform structure search or path selection based on constraint vectors; select model structures that meet resource constraints from the candidate model structure space based on the computing power, memory and inference latency limitations contained in the end-aware training constraint vectors. Joint training and optimization are performed on the selected structure; the parameters of the selected model structure are trained to achieve better hazard identification performance while meeting resource constraints.

5. The cross-platform adaptive deployment method for training and push integration according to claim 1, characterized in that: The cross-platform deployability analysis of the model structure includes the following sub-steps: Analyze the computational structure of the candidate model; by analyzing the computational structure of the candidate model, obtain the complete operator sequence and topology of the model, which will be used for subsequent deployment and adaptation analysis. Obtain the operator support rules for the target edge controller; based on the operating environment of different edge controllers in the power plant, classify and build the operator support rule library for the target deployment end, specifically including: operator type support list, operator implementation method, and operator operation restriction conditions; Perform operator-level deployment adaptation analysis; based on the model operator sequence and operator support rule base, perform operator-level deployment adaptation analysis on candidate models to identify operators or substructures in the model that do not meet the operating conditions of the target deployment end; Perform model substructure deployment adaptation analysis; based on operator-level deployment adaptation analysis, perform substructure analysis on the set of operators with dependencies in the model to identify model substructures that do not meet the deployment conditions; for model substructures with non-deployable operators, determine whether the substructure can be adapted by operator replacement, structure pruning, or computation path adjustment; if it cannot be adapted, mark the model substructure as a non-deployable structure. Generate model deployability evaluation results; based on the results of operator-level deployment adaptation analysis and model substructure deployment adaptation analysis, generate model deployability evaluation results for candidate models; Generate model conversion strategies; based on the model deployability assessment results, automatically generate model structure adaptation strategies, including operator replacement strategies, structure pruning strategies, and computation path adjustment strategies, for subsequent cross-platform model deployment optimization.

6. The cross-platform adaptive deployment method for training and push integration according to claim 1, characterized in that: The adaptive reconstruction and optimization of the model structure includes the following sub-steps: Determine the model structure optimization strategy; based on the model deployability evaluation results, determine the optimization strategy of the candidate model on the target deployment end, specifically including operator structure optimization strategy, model substructure replacement strategy, model compression strategy and inference execution optimization strategy; Perform operator-level structure optimization; based on the operator structure optimization strategy, specifically including operator fusion and operator splitting, adjust the operator structure in the candidate model to adapt to the operator types supported by the target edge controller. By fusing or splitting operators, the model operators can be correctly executed on the inference engine of the target edge controller. Perform model substructure replacement; when there is a model substructure in the candidate model that does not meet the deployment resource limit or operator support condition, perform structural replacement on the model substructure. The structural replacement includes: feature extraction module replacement, feature fusion structure replacement, detection or classification head replacement. Through model substructure replacement, the overall computational complexity of the model meets the resource limit of the target deployment end. Perform model compression optimization; while ensuring that the accuracy of hazard identification meets business requirements, perform model compression optimization on candidate models, specifically including model quantization and model pruning, to reduce the scale of model parameters, computational overhead and memory usage; Optimize the inference execution strategy; after optimizing the model structure, generate a deployment model that matches the target deployment end for subsequent cross-end model inference execution.

7. The cross-platform adaptive deployment method for training and push integration according to claim 1, characterized in that: Performing cross-platform model inference involves the following sub-steps: Deployment model distribution and loading; Distribute the deployment model according to the target deployment terminal type and load it into the corresponding device runtime environment; Cross-platform collaborative reasoning execution; When performing hazard investigation and reasoning tasks at the power plant site, different deployment terminals work together and trigger cross-terminal collaborative reasoning processes when necessary. Device operation status perception; during inference execution, real-time collection of operation status information of the target deployment end; The inference execution strategy is dynamically adjusted. Based on the device's operating status, the model inference execution method is dynamically adjusted, including: model switching adjustment, inference frequency adjustment, computation path adjustment, and cross-platform task migration. Output the reasoning results; output the hazard identification results, including: hazard category, risk level, target location or area, timestamp and equipment identification.

8. The cross-platform adaptive deployment method for training and push integration according to claim 1, characterized in that: Adaptive optimization of model training, updating, or deployment parameters includes the following sub-steps: Multi-dimensional inference performance data collection; collecting multi-dimensional inference performance data during model operation, including inference performance metrics, inference quality metrics, and environmental context information; Data backhaul and aggregation: The collected multi-dimensional inference performance data is backhauled from the station-side industrial control equipment or field edge controller to the central computing platform for unified aggregation and storage; Inference performance evaluation and analysis; Based on the collected multi-dimensional inference performance data, the model's operation on different deployment terminals is evaluated, including model accuracy and business effectiveness, whether the inference performance meets real-time requirements, whether resource utilization is reasonable, and performance differences between different devices. Generate optimization strategies; based on the evaluation results, generate model optimization strategies, specifically including model structure optimization strategies, deployment optimization strategies, and inference optimization strategies; Closed-loop feedback is provided to the training and deployment phases; the model optimization strategy is fed back to the model training phase, the model deployment optimization phase, and the inference execution phase respectively. By constructing an inference performance feedback mechanism based on power plant field operation data, closed-loop optimization of model training, deployment, and operation is achieved, enabling the model to continuously adapt to the operating environment of different edge controllers.

9. A cross-platform adaptive deployment system for training and push integration, characterized in that, Includes the following modules: The environmental parameter module is used to collect and construct environmental parameters at the power plant site deployment end. The deployment end includes a central computing platform, station-side industrial control equipment, and field edge controllers. The constraint vector module is used to generate edge-aware training constraint vectors, which transform the operational constraints of different deployment ends into structured constraint information that can participate in model training, and are used to guide the model to adapt to the resource conditions of the power plant field edge controller during the training phase. The training and inference module is used to perform integrated training and inference constraint training. It introduces the edge perception training constraint vector into the model training process and constrains the model parameter update process. This enables the model to meet the requirements of hazard identification accuracy while adapting to different on-site edge controller requirements, forming candidate models suitable for different inference scenarios. The deployment analysis module is used to obtain candidate models suitable for different inference scenarios, and before the models are actually deployed to the power plant field edge controller, to perform model structure parsing on the candidate models, and combine the operator support capabilities and runtime resource limitations of the target deployment end to perform cross-end deployability analysis on the models and generate model deployability evaluation results. The reconstruction and optimization module is used to perform adaptive reconstruction and optimization of the model structure of candidate models based on the model deployability evaluation results, and generate a deployment model that matches the target deployment end. The deployment and inference module is used to generate a deployment model, then distribute the deployment model to the power plant field deployment terminal to perform cross-terminal model inference, and dynamically adjust the model execution strategy during operation to achieve adaptive operation of the model in the power plant multi-edge controller environment; The feedback and optimization module is used to collect inference performance data and hazard identification results of the model during its operation at the power plant site, and feed them back to the training or deployment phase for subsequent model training updates or adaptive optimization of deployment parameters.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps in the cross-end model adaptive deployment method of training and push integration as described in any one of claims 1 to 8.

Images