Multi-target collaborative AI application trusted computing system of electric power Internet of Things terminal

Abstract

The invention relates to the field of electric power Internet of Things, and particularly discloses an electric power Internet of Things terminal multi-target collaborative AI application trusted computing system, which is characterized in that local operation data, neighbor collaborative states and environment contexts are collected in an omnibearing manner, and collaborative feature tensors representing real-time power grid working conditions are constructed through cleaning and heterogeneous fusion; and then the tensor is used for dynamically fitting a collaborative reference surface in a multi-dimensional target space, and a dynamic Pareto leading edge reflecting the current optimal tradeoff relation is generated. And furthermore, terminal AI actual control output is projected to the leading edge, and whether global optimal reasonable concession or malicious performance degradation is carried out in a multi-target conflict scene is precisely discriminated by analyzing a vector deviation between a convergence distance and a diversity angle of the terminal AI actual control output. And finally, performing confidence evaluation and safety judgment on the deviation based on a double-constraint mechanism.

Description

Technical Field

[0001] This application relates to the field of power Internet of Things (IoT), and more specifically, to a trusted computing system for AI applications involving multi-target collaboration in power IoT terminals. Background Technology

[0002] With the deepening construction of the ubiquitous power Internet of Things (IoT) and the vigorous development of energy internet technologies, the power grid is gradually evolving into a complex physical information system with interconnected everything and comprehensive state awareness. Against this backdrop, a large number of intelligent terminals deployed at the power grid edge are beginning to integrate artificial intelligence algorithms to achieve localized autonomous decision-making and efficient control. However, the operating environment of power IoT terminals is open and complex, and terminals face the risk of malicious tampering or adversarial attacks when performing distributed control tasks. Therefore, constructing a trusted computing scheme for multi-objective collaborative AI applications of power IoT terminals is of paramount importance for ensuring the security of power grid control commands, ensuring the predictability of multi-terminal collaborative behavior, and maintaining the overall stable operation of the power system.

[0003] Although some existing trusted computing technologies have been attempted to be applied to the security protection of power terminals, they still face severe technical bottlenecks when dealing with complex multi-objective collaborative scenarios such as voltage control and economic dispatch. First, existing collaborative detection mechanisms struggle to accurately distinguish between reasonable collaborative concessions and malicious performance degradation attacks. When a terminal sacrifices local performance indicators to achieve global optimization, traditional monitoring methods often misjudge this as abnormal behavior, leading to a high false alarm rate and severely impacting the practical implementation of collaborative control strategies. Second, in calculating the convergence distance for quantifying the trustworthiness of terminal behavior, existing technologies mostly use standard Euclidean distance to indiscriminately measure the deviation between the actual output and the optimal benchmark, mathematically treating all optimization objectives as equally weighted. However, in actual power system operation, physical risks across different objective dimensions exhibit significant heterogeneity and dynamic evolution. For example, the cascading failure risk caused by small deviations in safety-critical dimensions such as voltage stability is far greater than the larger fluctuations in economic cost dimensions, and this risk weight changes drastically with grid emergency states. Existing technologies lack the ability to perceive the heterogeneity of physical risks to multidimensional targets. They are prone to losing sensitivity to covert attacks in the security dimension and may overreact to normal market economic fluctuations, making it difficult to meet the high reliability security requirements of the power Internet of Things.

[0004] Therefore, an optimized power Internet of Things (IoT) terminal system is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a trusted computing system for multi-objective collaborative AI applications in power IoT terminals.

[0006] According to one aspect of this application, a trusted computing system for AI applications involving multi-objective collaboration in power Internet of Things (IoT) terminals is provided, comprising: The multi-source collaborative sensing module is used to acquire local terminal operation data, collaborative neighbor terminal status data, and environmental context data; The heterogeneous data fusion module is used to perform data cleaning, heterogeneous feature extraction, and collaborative situational concatenation on local terminal operation data, collaborative neighbor terminal status data, and environmental context data to obtain a collaborative feature tensor. The dynamic fitting module for the reference surface is used to perform multi-objective dynamic fitting of the collaborative feature tensor to obtain the dynamic Pareto front. The Behavior Mapping and Trust Domain Projection Module is used to perform cooperative behavior vector mapping and trust domain projection on the actual control output of terminal AI applications based on the dynamic Pareto front to obtain convergence distance and diversity angle. The confidence assessment module is used to perform a dual-constraint confidence assessment on the convergence distance and diversity perspectives to obtain a comprehensive trust score. The Trust Decision module is used to make trust decisions based on the comprehensive trust score according to the security threshold to obtain a trust decision result.

[0007] Compared with existing technologies, this application provides a trusted computing system for multi-objective collaborative AI applications in the power IoT terminal. First, it comprehensively collects local operating data, neighbor collaborative states, and environmental context. After cleaning and heterogeneous fusion, it constructs a collaborative feature tensor representing the real-time power grid operating conditions. Then, it uses this tensor to dynamically fit a collaborative reference surface in a multi-dimensional objective space, generating a dynamic Pareto front reflecting the current optimal trade-off. Next, it projects the actual control output of the terminal AI onto this front, and by analyzing its convergence distance and vector deviations from diverse angles, it accurately distinguishes between reasonable concessions for global optimization and malicious performance degradation in multi-objective conflict scenarios. Finally, it performs confidence assessment and security judgment on the deviations based on a dual constraint mechanism. This effectively overcomes the limitations of traditional isotropic metrics in perceiving the heterogeneity of physical risks, achieving precise and dynamic protection for the collaborative behavior of power IoT terminal AI. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1This is a system block diagram of a trusted computing system for multi-objective collaborative AI applications in the power Internet of Things (IoT) terminal, according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of data flow in a trusted computing system for multi-objective collaborative AI applications of power Internet of Things terminals according to an embodiment of this application.

[0011] Figure 3 This is a block diagram of the heterogeneous data fusion module in a trusted computing system for multi-objective collaborative AI applications of power Internet of Things terminals according to an embodiment of this application.

[0012] Figure 4 This is a block diagram of a multi-dimensional collaborative situation stitching unit in a trusted computing system for multi-target collaborative AI applications of power Internet of Things terminals according to an embodiment of this application.

[0013] Figure 5 This is a block diagram of the reference surface dynamic fitting module in the trusted computing system for multi-objective collaborative AI applications of power Internet of Things terminals according to an embodiment of this application.

[0014] Figure 6 This is a block diagram of the behavior mapping and trust domain projection module in the trusted computing system for multi-objective collaborative AI applications of power Internet of Things terminals according to an embodiment of this application.

[0015] Figure 7 This is a block diagram of a geometric reliability decomposition unit in a trusted computing system for multi-objective collaborative AI applications of power IoT terminals according to an embodiment of this application.

[0016] Figure 8 This is a block diagram of the confidence assessment module in a trusted computing system for multi-objective collaborative AI applications of power Internet of Things terminals according to an embodiment of this application. Detailed Implementation

[0017] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0018] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0019] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0020] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0021] To address the problems mentioned above in the background technology, this application proposes a trusted computing system for AI applications involving multi-objective collaboration in power Internet of Things (IoT) terminals. Figure 1 This is a system block diagram of a trusted computing system for multi-objective collaborative AI applications in the power Internet of Things (IoT) terminal, according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in a trusted computing system for multi-objective collaborative AI applications in the power Internet of Things (IoT) terminal, according to an embodiment of this application. Figure 1 and Figure 2 As shown, the trusted computing system 100 for multi-objective collaborative AI applications of power Internet of Things terminals according to an embodiment of this application includes: a multi-source collaborative sensing module 110, used to acquire local terminal operation data, collaborative neighbor terminal status data, and environmental context data; a heterogeneous data fusion module 120, used to perform data cleaning, heterogeneous feature extraction, and collaborative situational splicing on the local terminal operation data, collaborative neighbor terminal status data, and environmental context data to obtain a collaborative feature tensor; a reference surface dynamic fitting module 130, used to perform multi-objective collaborative reference surface dynamic fitting on the collaborative feature tensor to obtain a dynamic Pareto front; a behavior mapping and trust domain projection module 140, used to perform collaborative behavior vector mapping and trust domain projection on the actual control output of the terminal AI application based on the dynamic Pareto front to obtain convergence distance and diversity angle; a confidence evaluation module 150, used to perform confidence evaluation based on dual constraints on the convergence distance and diversity angle to obtain a comprehensive trust score; and a trusted decision module 160, used to perform trusted decision on the comprehensive trust score based on a security threshold to obtain a trusted decision result.

[0022] In the aforementioned trusted computing system 100 for multi-objective collaboration of power IoT terminals using AI applications, the multi-source collaborative sensing module 110 is used to acquire local terminal operating data, collaborative neighbor terminal status data, and environmental context data. It should be understood that in multi-objective collaboration scenarios of power IoT terminals, AI application decisions rely on comprehensive and real-time operating condition information. Single-dimensional data cannot reflect the collaborative relationship between multiple objectives such as voltage stability and economic dispatch, easily leading to trusted detection bias. Therefore, this application collects local operating status, neighbor interaction data, and environmental constraint information through multi-source collaboration to provide complete data support for subsequent heterogeneous data fusion and dynamic Pareto front fitting. This ensures that subsequent trusted detection is based on real operating conditions, avoiding misjudgments caused by missing or incomplete data, and laying a data foundation for behavior identification in multi-objective collaborative scenarios.

[0023] Specifically, in one example of this application, local terminal operation data is collected through the terminal's built-in sensing unit and control module, covering core operational information such as real-time electrical parameters and AI control command execution status. Collaborative neighbor terminal status data is acquired through a dedicated communication protocol for the power IoT. To address the risks of privacy leaks and poisoning attacks during this data acquisition process, this system constructs a secure aggregation channel based on secret sharing and random vectors in federated learning. Specifically, the execution of this channel includes five stages: First, the initialization stage, where a key generation center distributes key pairs; second, the secret sharing stage, where neighboring terminals exchange public key information, execute a secret sharing algorithm to divide their private keys into secret fragments, and encrypt and distribute them to other nodes; next, the random vector stage, where the terminal uses the negotiated key to calculate random vectors with other nodes, adds them as mask noise to the collaborative status data (such as power deficit), and performs homomorphic encryption to ensure that single-point data presents a high-entropy scrambled state during transmission; next, the authentication stage, verifying the authenticity of node identities; and finally, the secure aggregation stage, utilizing aggregation characteristics to cancel out the random vectors between different nodes, thereby restoring the global collaborative situation without exposing individual privacy data. This mechanism transmits only masked state data with added noise and homomorphic encryption, enabling secure aggregation of usable but invisible data under centralized coordination or peer-to-peer interaction. This achieves collaborative optimization while reducing the risk of data leakage at the source. Furthermore, environmental context data is collected through environmental monitoring equipment and topology sensing modules deployed at key nodes of the power grid. This data includes constraint information such as power grid topology connection status, real-time weather conditions, and timestamps, comprehensively capturing external environmental factors affecting collaborative decision-making.

[0024] In the aforementioned trusted computing system 100 for multi-objective collaboration of power IoT terminals using AI applications, the heterogeneous data fusion module 120 is used to perform data cleaning, heterogeneous feature extraction, and collaborative situational awareness splicing on local terminal operation data, collaborative neighbor terminal status data, and environmental context data to obtain a collaborative feature tensor. It should be understood that because multi-source data comes from different terminals and environments, there are quality issues such as noise and missing values, and the data types are heterogeneous (electrical quantities, status flags, environmental parameters), lacking a unified collaborative representation, and cannot directly support subsequent baseline fitting. Therefore, this application further performs cleaning and denoising, heterogeneous feature extraction, and collaborative situational awareness splicing on local terminal operation data, collaborative neighbor terminal status data, and environmental context data to eliminate data interference, mine key features, and fuse collaborative context. This module follows the concept of multimodal learning, similar to the ability of a biological brain to seamlessly integrate multi-sensory information. By fusing data from different modalities such as electrical quantities, status quantities, and environmental quantities, it aims to provide a more comprehensive analytical perspective than a single data source, thereby constructing a more intelligent and robust collaborative situational awareness system. This enables consistency and coordination of multi-source data, forming a tensor structure that comprehensively reflects the real-time operating conditions of the power grid, providing high-quality input for dynamic Pareto front fitting.

[0025] Figure 3 This is a block diagram of a heterogeneous data fusion module in a trusted computing system for multi-objective collaborative AI applications of power IoT terminals according to an embodiment of this application. Figure 3 As shown in the embodiments of this application, the heterogeneous data fusion module 120 includes: a spatiotemporal synchronization preprocessing unit 121, used to perform spatiotemporal alignment and normalization preprocessing on local terminal running data, cooperative neighbor terminal state data and environmental context data to obtain a synchronization state vector; a temporal feature extraction unit 122, used to perform temporal feature extraction on the synchronization state vector based on lightweight CNN-1D to obtain an abstract temporal feature matrix; and a multidimensional cooperative situation splicing unit 123, used to perform multidimensional cooperative situation splicing on the abstract temporal feature matrix to obtain the cooperative feature tensor.

[0026] Specifically, the spatiotemporal synchronization preprocessing unit 121 is used to perform spatiotemporal alignment and normalization preprocessing on local terminal operating data, collaborative neighbor terminal state data, and environmental context data to obtain a synchronized state vector. It should be understood that due to the different sampling frequencies of the local terminal and neighbor terminals, and the differences in environmental data acquisition times, the data is spatiotemporally asynchronous, and various data units have different dimensions, such as voltage (kV), current (A), and power (kW). Direct fusion will cause serious interference. Therefore, this application further carries out data spatiotemporal benchmark unification and dimensional standardization processing to eliminate spatiotemporal deviations and dimensional differences, ensuring data comparability and temporal consistency. In this way, a spatiotemporally synchronized and dimensionally unified state vector can be obtained, providing interference-free basic data for subsequent feature extraction and avoiding feature extraction distortion caused by data heterogeneity.

[0027] More specifically, in a concrete example of this application, during the spatiotemporal alignment stage, a fixed time window is set based on the sampling time of the local terminal. Linear interpolation is used to complete the time mapping of neighboring terminal data, and spatial location alignment is achieved by combining power grid topology information, ensuring that multi-source data within the same time window correspond to the same power grid operating condition. It is worth noting that, considering that edge and device layer data are directly related to the power grid operating status and user privacy, the preprocessing unit also integrates a real-time anonymization mechanism at the device end when processing the aforementioned data. The system can perform millisecond-level dynamic desensitization processing on sensitive information such as user identifiers uploaded by terminals such as smart meters, ensuring data availability while preventing the leakage of sensitive geographical locations or user privacy data. During the normalization stage, statistical standardization methods are used to map data to the same numerical range for physical quantities of different dimensions, eliminating the differences in magnitude of parameters such as voltage, current, and power, and ultimately outputting a state vector with unified dimensions and spatiotemporal synchronization.

[0028] Specifically, the temporal feature extraction unit 122 is used to extract temporal features from the synchronization state vector based on lightweight CNN-1D to obtain an abstract temporal feature matrix. It should be understood that since the synchronization state vector is only raw temporal data, it does not explore deep temporal dependencies and local correlations between data, and the computing power of edge terminals is limited, complex models cannot be deployed. Therefore, this application further employs lightweight CNN-1D to extract temporal features from the synchronization state vector, thereby efficiently capturing the dynamic changing trends of electrical parameters, the temporal patterns of neighbor collaboration requests, and the influence of environmental factors. In this way, a high-dimensional abstract temporal feature matrix can be obtained while adapting to edge computing power, enhancing the discriminative power of the data and providing accurate feature support for subsequent collaborative behavior mapping.

[0029] More specifically, in a concrete example of this application, the temporal feature extraction network undergoes a rigorous pre-training process before deployment. First, self-supervised learning is performed based on massive amounts of unlabeled historical data from the edge, enabling the network to learn the inherent temporal dependencies of power parameters. Then, an edge adversarial training mechanism is used to fine-tune the network; that is, adversarial samples containing harmonic current waveforms and electromagnetic interference noise are actively injected during training, forcing the network to learn robust feature representations, thereby preventing misjudgments caused by environmental noise. During actual inference, the synchronization state vector is first expanded along the time dimension to construct a temporal matrix, clarifying the sequence length and feature dimension. Then, a small-sized convolutional kernel optimized by the aforementioned adversarial training is used to perform sliding convolution operations on the temporal matrix, extracting local temporal features such as voltage fluctuations and power change trends in parallel through multi-channel convolutional kernels. Finally, the ReLU activation function enhances the nonlinear representation capability of the features, and pooling operations compress the feature dimension, reducing computational overhead while ensuring feature effectiveness, ultimately outputting an abstract temporal feature matrix containing deep temporal correlations and noise-resistant properties.

[0030] Specifically, the multi-dimensional collaborative situation stitching unit 123 is used to perform multi-dimensional collaborative situation stitching on the abstract temporal feature matrix to obtain the collaborative feature tensor. It should be understood that, since the abstract temporal feature matrix only integrates the temporal correlation features of multi-source data and does not clearly distinguish core dimensions such as local operation, neighbor collaboration, and environmental constraints, it lacks a comprehensive characterization of the collaborative scenario of the power Internet of Things and cannot support accurate fitting of multi-objective collaborative benchmarks. Therefore, this application further implements dimensional decoupling, feature fusion, and channel stitching on the abstract temporal feature matrix to mine collaborative correlation information of different dimensions and construct a multi-dimensional representation that combines its own state, interaction relationships, and environmental constraints. In this way, a feature tensor that comprehensively reflects the real-time collaborative operating conditions of the power grid can be formed, providing complete and accurate input data for dynamic Pareto front fitting and ensuring the effectiveness of subsequent reliable detection.

[0031] Figure 4 This is a block diagram of a multi-dimensional collaborative situation stitching unit in a trusted computing system for multi-target collaborative AI applications of power IoT terminals according to an embodiment of this application. Figure 4 As shown in the embodiments of this application, the multi-dimensional collaborative situation stitching unit 123 includes: a feature decoupling subunit 1231, used to decouple the abstract temporal feature matrix to obtain local self-features, neighbor interaction features and environmental constraint features; and a feature stitching subunit 1232, used to stitch the local self-features, neighbor interaction features and environmental constraint features based on the topological embedding vector to obtain a collaborative feature tensor.

[0032] More specifically, the feature decoupling subunit 1231 is used to decouple the abstract temporal feature matrix to obtain local self-features, neighbor interaction features, and environmental constraint features. It should be understood that since the abstract temporal feature matrix is ​​a mixed collection of temporal features from multiple sources, features from different sources and with different physical meanings are intertwined, making it impossible to distinguish the influence of local operating state, neighbor cooperative behavior, and environmental constraints. This leads to ambiguity in the cooperative situation representation, affecting the logic and effectiveness of subsequent feature splicing. Therefore, this application further decouples the abstract temporal feature matrix precisely based on the source attributes and physical meaning of the features, thereby separating feature components of different dimensions and clarifying the physical connotation of each feature. This allows for the acquisition of pure, independent local self-features, neighbor interaction features, and environmental constraint features, laying the foundation for subsequent feature splicing incorporating topological information and ensuring the structural rationality and information integrity of the cooperative feature tensor.

[0033] Specifically, in a concrete example of this application, a feature attribute label library is first constructed to label each feature dimension in the abstract temporal feature matrix with attributes such as source (local / neighbor / environment) and physical meaning (electrical parameters / interaction commands / constraints). Then, based on the attribute label library, a feature filtering and separation algorithm is used to extract features labeled as originating locally and representing the terminal's own operating state, forming local features. Next, features labeled as originating from neighbors and reflecting collaborative interaction behavior are filtered to form neighbor interaction features. Finally, features labeled as originating from the environment and reflecting constraints are extracted to obtain environmental constraint features, ensuring that the three types of features are non-overlapping, non-omitted, and fully cover the effective information of the abstract temporal feature matrix.

[0034] More specifically, the feature splicing subunit 1232 is used to splice local self-features, neighbor interaction features, and environmental constraint features based on topology embedding vectors to obtain a cooperative feature tensor. It should be understood that since local self-features, neighbor interaction features, and environmental constraint features each represent only a single dimension of information, lacking characterization of the terminal's physical location and cooperative correlation strength within the power grid, the spliced ​​features cannot reflect the influence of spatial topology on cooperative behavior, making it difficult to accurately represent complex power grid cooperative operating conditions. Therefore, this application further introduces topology embedding vectors to fuse and splice the three types of features, thereby incorporating spatial topology information and strengthening the cooperative correlation of features. In a specific example of this application, the local self-features, neighbor interaction features, and environmental constraint features are spliced ​​using the following formula to obtain the cooperative feature tensor: in, For the co-feature tensor, This is a splicing operation along the feature channel dimension. Due to local characteristics, For neighbor interaction features, Due to environmental constraints, This is a topological embedding vector. In this way, a collaborative feature tensor with time series features, spatial topological features, and multi-dimensional correlation features can be constructed to comprehensively and accurately characterize the collaborative situation of power Internet of Things terminals, providing high-quality input for subsequent dynamic Pareto front fitting.

[0035] In the aforementioned trusted computing system 100 for multi-objective collaborative AI applications in the power IoT terminal, the dynamic fitting module 130 is used to dynamically fit the collaborative feature tensor to a multi-objective collaborative benchmark to obtain a dynamic Pareto front. It should be understood that in the multi-objective collaborative scenario of the power IoT, objectives such as voltage stability and economic dispatch are mutually exclusive, and the power grid operating conditions (load, topology, environment) change dynamically in real time. Traditional static benchmarks or single-objective thresholds cannot reflect the optimal trade-offs under different operating conditions, resulting in a lack of accurate basis for judging AI collaborative behavior. Therefore, this application further constructs a multi-objective collaborative benchmark based on the collaborative feature tensor to capture the optimal trade-off boundaries of each objective under the current operating conditions. This forms a dynamic benchmark adapted to the real-time power grid state, providing an accurate reference for distinguishing reasonable concessions from malicious attacks by AI, and solving the problems of poor adaptability and high misjudgment rate of traditional static benchmarks.

[0036] Figure 5 This is a block diagram of the reference surface dynamic fitting module in the trusted computing system for multi-objective collaborative AI applications of power Internet of Things terminals according to an embodiment of this application. Figure 5 As shown in the embodiments of this application, the reference surface dynamic fitting module 130 includes: a multi-objective reference vector prediction unit 131, used to perform multi-objective reference vector prediction based on a meta-optimizer on the co-functional tensor based on a predefined set of optimization objective functions to obtain a discrete Pareto optimal reference vector set; and a front fitting unit 132, used to perform front fitting based on weighted Chebyshev scalarization on the discrete Pareto optimal reference vector set to obtain a dynamic Pareto front.

[0037] Specifically, the multi-objective reference vector prediction unit 131 is used to perform multi-objective reference vector prediction on the cooperative feature tensor based on a predefined set of optimization objective functions to obtain a discrete Pareto optimal reference vector set. It should be understood that, since the cooperative feature tensor contains complex multi-dimensional information about the power grid operating conditions, directly fitting a continuous front surface involves large computational loads, insufficient real-time performance, and lacks a clear theoretical optimal anchor point, resulting in low efficiency and poor accuracy in benchmark construction. Therefore, this application further combines a predefined set of optimization objective functions with a meta-optimizer to perform multi-objective reference vector prediction on the cooperative feature tensor, thereby quickly obtaining representative discrete optimal anchor points. This provides high-quality core support points for subsequent front surface fitting while ensuring real-time performance, ensuring the accuracy and efficiency of the dynamic benchmark surface, and adapting to the computing power constraints of edge terminals.

[0038] More specifically, in a concrete example of this application, a predefined set of optimization objective functions is first loaded to clarify the computational logic and constraint boundaries of each objective, such as voltage stability and economic benefits. Then, the collaborative feature tensor is input into a meta-optimizer built on a federated learning framework. This model is not simply an offline training product, but rather, through a federated learning mechanism, it aggregates multi-source heterogeneous optimization experience data from the entire network while protecting the privacy of the original topology and load data of each participant. This meta-optimizer has learned the high-dimensional mapping relationship between complex power grid operating conditions and Pareto optimal trade-off vectors, possessing the ability to infer local optimal solutions from a global perspective. Finally, the meta-optimizer quickly outputs a set of discrete Pareto optimal reference vectors, each vector corresponding to a theoretical optimal solution under a specific objective trade-off preference, comprehensively covering the core trade-off scenarios under the current operating conditions, forming a set of discrete reference vectors.

[0039] Specifically, the front fitting unit 132 is used to perform front fitting based on weighted Chebyshev scalarization on the discrete Pareto optimal reference vector set to obtain the dynamic Pareto front. It should be understood that since the discrete Pareto optimal reference vectors can only cover a limited set of objective trade-off scenarios, they cannot form a continuous benchmark boundary, making it difficult to comprehensively evaluate all possible collaborative strategies of AI, resulting in evaluation blind spots in the benchmark surface. Therefore, this application further employs the weighted Chebyshev scalarization method to fit the discrete reference vectors, thereby constructing a continuous and complete multi-objective collaborative benchmark boundary. This enables full-coverage evaluation of any multi-objective trade-off strategy, ensuring that the dynamic Pareto front can accurately match various collaborative behaviors of AI, providing blind-spot-free benchmark support for subsequent behavior credibility determination.

[0040] More specifically, in a concrete example of this application, the discrete Pareto optimal reference vector set is first normalized to eliminate the influence of differences in the dimensions of different objectives. Then, a weighted Chebyshev scalarization method is used to transform the multi-objective optimization problem into a series of single-objective aggregation problems, traversing all objective trade-off directions by adjusting the weight vectors. Finally, using Kriging interpolation, based on the discrete reference vectors and the scalarization results, the gaps between the trade-off directions are filled, constructing a continuous and smooth dynamic Pareto front that fully characterizes all possible optimal objective trade-offs under the current operating condition.

[0041] In the aforementioned trusted computing system 100 for multi-objective collaborative AI applications in the power IoT terminal, the behavior mapping and trust domain projection module 140 is used to perform collaborative behavior vector mapping and trust domain projection on the actual control output of the terminal AI application based on the dynamic Pareto front to obtain convergence distance and diversity angle. It should be understood that since the actual control output of the terminal AI is an abstract operational instruction, it cannot be directly compared with the optimal benchmark of the multi-objective dynamic Pareto front. Furthermore, in multi-objective collaborative scenarios, it is necessary to distinguish between capability deficiencies and trade-off intentions, and a single indicator is insufficient to comprehensively characterize behavioral credibility. Therefore, this application further maps the control output to a multi-objective effect vector and then decomposes the geometric features through trust domain projection to quantify the degree of deviation from the optimal benchmark and objective trade-off preferences. In this way, abstract instructions can be transformed into measurable two-dimensional indicators, accurately distinguishing between reasonable collaborative concessions and malicious attacks, providing core data support for subsequent confidence assessment, and solving the problem that traditional detection methods cannot accurately characterize collaborative behavior.

[0042] Figure 6 This is a block diagram of the behavior mapping and trust domain projection module in a trusted computing system for multi-objective collaborative AI applications of power IoT terminals according to an embodiment of this application. Figure 6 As shown in the embodiments of this application, the behavior mapping and trust domain projection module 140 includes: a behavior effect mapping unit 141, used to perform behavior effect mapping based on a physical agent model on the actual control output of the terminal AI application to obtain an actual effect vector; a Pareto projection unit 142, used to perform Pareto projection based on Euclidean distance minimization on the dynamic Pareto front and the actual effect vector to obtain a Pareto projection point; and a geometric credibility decomposition unit 143, used to perform geometric feature decomposition and credibility metric calculation on the Pareto projection point and the actual effect vector to obtain a convergence distance and a diversity angle.

[0043] Specifically, the behavior effect mapping unit 141 is used to perform behavior effect mapping on the actual control output of the terminal AI application based on a physical proxy model to obtain an actual effect vector. It should be understood that since the control output of the terminal AI is a device-oriented operation command, such as power adjustment or switching action commands, it lacks a direct representation of the multi-objective optimization effect and cannot be directly compared with the target space benchmark of the dynamic Pareto front. Therefore, this application further simulates the execution effect of control commands through a physical proxy model, transforming abstract commands into quantized vectors in a multi-dimensional target space, thereby establishing a correlation between control behavior and multi-objective optimization effect. In this way, incomparable operation commands can be transformed into measurable target effect representations, providing a unified dimension for subsequent projection calculations with the Pareto front, ensuring the effectiveness of the benchmark comparison.

[0044] More specifically, in a concrete example of this application, the AI ​​control output is first parsed to extract key information such as operation type, adjustment parameters, and execution time, which is then transformed into input parameters recognizable by the proxy model. These parameters are then injected into a lightweight physical proxy model constructed using robustness-preserving distillation technology. This model is derived from a high-fidelity digital twin model in the cloud through knowledge distillation compression. While adapting to the limited computing power at the edge and device layers, it retains the model's robustness against attacks, resisting potential attacks targeting model vulnerabilities. The model rapidly simulates the specific impact values ​​of each optimization objective, such as voltage stability and economic benefits, after instruction execution. Finally, the impact values ​​of each objective are combined in a predefined order to form a multi-dimensional actual effect vector representing the multi-objective effect, ensuring that the vector dimension is consistent with the target space dimension of the Pareto front.

[0045] Specifically, the Pareto projection unit 142 is used to perform Pareto projection on the dynamic Pareto front and the actual effect vector based on Euclidean distance minimization to obtain the Pareto projection point. It should be understood that since the actual effect vector usually does not fall exactly on the dynamic Pareto front, there is a lack of a direct optimal benchmark for comparison, making it impossible to accurately measure its deviation from the optimal state. Therefore, this application further uses an Euclidean distance minimization algorithm to find the point on the Pareto front that is closest to the actual effect vector, thereby establishing a theoretically optimal matching benchmark point. This provides a unique and most closely fitting optimal benchmark for the actual effect vector, transforming the abstract deviation from the optimal state into a concrete geometric distance, laying the foundation for subsequent convergence distance calculations and ensuring the accuracy of deviation assessment.

[0046] More specifically, in a concrete example of this application, the continuous dynamic Pareto front is first discretized to generate a high-density set of discrete reference points covering all objective trade-off directions. Then, the Euclidean distance between the actual effect vector and each discrete reference point is calculated, establishing a distance matrix. Finally, the distance matrix is ​​traversed, and the discrete reference point with the smallest distance is selected as the initial projection point. If a continuous region exists near the initial point, linear interpolation is used to optimize and obtain the final Pareto projection point, ensuring that the projection point is both located on the front surface and has the shortest distance to the actual effect vector.

[0047] Specifically, the geometric credibility decomposition unit 143 is used to perform geometric feature decomposition and credibility metric calculation on the Pareto projection point and the actual effect vector to obtain the convergence distance and diversity angle. It should be understood that since a single geometric distance cannot distinguish whether AI behavior is a deviation caused by insufficient capability or an active goal trade-off preference, relying solely on distance indicators can lead to misjudgment. Therefore, this application further decomposes the geometric relationship between the two in two dimensions to calculate the convergence distance and diversity angle, thereby quantifying the capability level and collaborative intent of the behavior respectively. In this way, a dual-dimensional characterization of AI behavior can be achieved, reflecting both the degree of deviation from the optimal benchmark (capability) and the direction of goal trade-offs (intent), providing comprehensive and accurate indicator support for subsequent dual-constraint evaluation and avoiding the blind spots of single-indicator evaluation.

[0048] More specifically, in a concrete example of this application, the calculation of the convergent distance first involves determining dynamic risk weights. Based on the current power grid operating conditions (normal state, emergency state), weights are assigned to each objective dimension in the actual effect vector, with safety-critical objectives receiving higher weights than economic objectives. Then, the standard Euclidean distance is calculated, using the Pareto projection point as a reference. The deviation between the actual effect vector and the reference point is quantified using the Euclidean distance formula to obtain the original distance value. Finally, a credibility metric is performed, extracting the distance distribution benchmark from historical credibility data. The original distance value is then normalized to eliminate numerical fluctuations caused by differences in operating conditions. Simultaneously, the rationality of the distance is verified in conjunction with power grid safety rules, ultimately yielding the standardized convergent distance.

[0049] It should be understood that the convergence distance calculation mechanism in the above embodiments has a fundamental technical reason: its loss measurement follows the isotropic assumption. Specifically, this mechanism uses standard Euclidean distance to quantify the deviation between the actual effect vector and the Pareto optimal projection point, which means that mathematically, deviations across all optimization objective dimensions are treated equally without distinction. However, in the specific application scenario of power systems, different optimization objectives naturally have significantly different criticalities or risk weights. A typical example is the comparison between safety-critical objectives and economic optimization objectives. A small deviation in the voltage stability dimension carries a potential risk far greater than a large deviation in the economic cost dimension. The former may directly touch the grid safety red line, or even trigger a chain reaction of instability, while the latter only affects short-term economic benefits. More complexly, this criticality is not static but evolves dynamically with the grid's operating conditions. For example, during normal grid operation, the economic weight may be relatively high, but when the system suffers disturbances or enters an emergency, the weight of safety objectives, such as voltage and frequency stability, will instantly surge to absolute dominance. Therefore, the original mechanism fails to consider the heterogeneity of physical risks carried by different dimensions in a multi-objective space, measuring them in a geometrically equal-weighted manner. This conflicts with the fundamental reality of unequal weighting of physical risks in the domain, easily leading to insufficient sensitivity to minor shifts in security-critical dimensions, potentially overlooking dangerous slow poisoning attacks, while being overly sensitive to normal market fluctuations in the economic dimension, resulting in frequent false alarms of normal cooperative behavior. To address these technical shortcomings, a dynamic criticality-weighted anisotropic distance measurement mechanism is proposed. This mechanism abandons isotropic Euclidean distance and instead constructs a novel distance measurement that dynamically reflects the current operating state of the power grid and the risk level of each objective.

[0050] In particular, Figure 7 This is a block diagram of a geometric reliability decomposition unit in a trusted computing system for multi-objective collaborative AI applications of power IoT terminals according to an embodiment of this application. Figure 7 As shown in the embodiments of this application, the geometric confidence decomposition unit 143 includes: a risk quantification subunit 1431, used to quantify the risk of the actual effect vector based on the power grid operating condition indicator to obtain a dynamic critical weight vector; an anisotropic distance calculation subunit 1432, used to calculate the anisotropic distance between the Pareto projection point and the actual effect vector based on the dynamic critical weight vector to obtain the original anisotropic distance; and a distance normalization subunit 1433, used to normalize the original anisotropic distance based on historical benchmarks to obtain the convergence distance.

[0051] More specifically, the risk quantification subunit 1431 is used to quantify the risk of the actual effect vector based on the power grid operating condition indicators to obtain a dynamic critical weight vector. It should be understood that since the risk level of each objective dimension in the actual effect vector changes dynamically with the power grid operating condition—for example, the risk weight of voltage stability is much higher than the economic benefit under emergency conditions, while the weights are relatively balanced under normal conditions—equal weighting would lead to distorted risk perception. Therefore, it is necessary to dynamically generate a critical weight vector corresponding one-to-one with each objective dimension based on the actual effect vector of the AI ​​application and the current power grid operating condition. That is, the abstract risk level must be quantified into specific mathematical weights to effectively guide subsequent distance calculations. Specifically, this is accomplished through a risk quantification function, the core purpose of which is: as the state value of a certain objective approaches its safety boundary, its corresponding weight will increase exponentially. This process can be expressed as: Wherein, the weight of the i-th objective The calculation method is as follows: in, This represents the final generated dynamic critical weight vector; This represents a risk quantification function; This is the actual effect vector, and its components are... This represents the actual effect of the AI ​​decision on the i-th objective; Indicates the operating status of the power grid, such as normal, alarm, or emergency. , and Representing vectors respectively The first, second, and m-th components, i.e., the dynamic weights of the first, second, and m-th targets; This represents the total number of dimensions of the optimization objective. Representing vectors The i-th component, i.e., the dynamic weight of the i-th target; Indicates the relationship with power grid operating conditions The relevant risk gain coefficient, for example, will increase significantly in an emergency, thereby increasing the overall weight of safety-related dimensions; Represents the natural exponential function; Indicates taking the absolute value; This represents the safety center value or desired operating point of the i-th target; This represents the safety limit boundary value of the i-th target.

[0052] More specifically, in a particular example of this application, a specific numerical calculation embodiment is given to aid in the illustration: the power grid operating condition is set as an emergency state (risk gain coefficient). =2), the i-th objective is the voltage stability objective, and its actual effect value is =0.7kV, safety center value =1.0kV, safety limit boundary value =0.6kV. First, substitute the weight formula of the i-th target: calculate the numerator of the exponent term |0.7-1.0|=0.3, the denominator |0.6-1.0|=0.4, and the result of the exponent term is... (0.3 / 0.4)= (0.75)≈2.117, risk gain term 2×2.117≈4.234, final weight =1+4.234≈5.234. Therefore, it can be seen that the voltage stability target, due to its emergency operating condition and actual value approaching the safety limit, has a dynamic weight of approximately 5.234. This reflects the high weight allocation of key targets under risk heterogeneity, providing a reasonable risk weight basis for subsequent anisotropic distance calculations.

[0053] In this way, dynamic adaptation of the weights is achieved, meaning that the weights are no longer fixed, but rather adapt to the actual effect of the AI ​​decision-making. Strong correlation exists; as the state deviates from the safety center and approaches the limit, the risk weight is amplified non-linearly. Simultaneously, by introducing a working condition coupling coefficient... This enables the entire weighting system to respond to the global power grid security situation, ultimately outputting a dynamic critical weight vector. It accurately reflects the real-time risk distribution.

[0054] More specifically, the anisotropic distance calculation subunit 1432 is used to calculate the anisotropic distance between the Pareto projection point and the actual effect vector based on the dynamic critical weight vector to obtain the original anisotropic distance. It should be understood that because traditional Euclidean distance applies equal weights to each target dimension, it cannot reflect the risk heterogeneity of each target under different operating conditions. For example, in an emergency, a small deviation in the voltage dimension is equivalent to a large deviation in the economic dimension, resulting in the distance calculation failing to reflect the true safety status. Therefore, a diagonal weight matrix is ​​constructed using the critical weight vector generated in the previous step, and the weighted distance between the actual effect vector and the Pareto projection point is calculated using this matrix. In particular, it integrates risk weights into the core of the distance measurement, aiming to replace the original Euclidean distance calculation. Specifically, a non-uniform stretch is applied to the original target space geometrically. In the key risk dimensions, small geometric deviations will be significantly amplified, as shown below: in, This represents the calculated original anisotropic distance; This represents the projection point of the actual effect vector onto the Pareto front; Represents the transpose operation of a matrix or vector; This is a diagonal weight matrix composed of critical weights; This indicates an operation to construct a diagonal matrix whose diagonal elements are the inputs within the parentheses.

[0055] More specifically, in a particular example of this application, a specific numerical calculation embodiment is given to aid in the illustration: setting a dynamic critical weight vector. =[5.234, 1.862] (corresponding to voltage stability and economic benefit targets respectively), actual effect vector =[0.7,85], Pareto projection point =[0.9,90]. First, construct the diagonal weight matrix. = (5.234, 1.862). The vector difference is then calculated. =[-0.2,-5], its transpose is [-0.2,-5]. Then calculate the quadratic form (-0.2,-5) × (5.234, 1.862) × (-0.2, -5) =(-0.2)²×5.234+(-5)²×1.862=0.20936+46.55=46.75936. Finally, taking the square root yields the original anisotropic distance. = ≈6.838. This demonstrates that the high weighting of voltage stability amplifies the impact of small deviations, enabling the distance results to accurately reflect the risk distribution and providing a risk-adjusted distance basis for subsequent normalization processing.

[0056] The calculation in this step It is no longer a purely geometric distance, but a risk-adjusted distance. For example, if the weight of the voltage stability dimension... The contribution of voltage components to the final distance calculation is amplified even if the actual difference is small, thus ensuring the effective capture of key safety risks and achieving the core objective of risk-oriented measurement.

[0057] More specifically, the distance normalization subunit 1433 is used to normalize the original anisotropic distance based on a historical benchmark to obtain the convergent distance. It should be understood that since the numerical range of the original anisotropic distance fluctuates with changes in operating conditions, the distance values ​​under different operating conditions lack comparability and are not benchmarked against the deviation range of historical reliable behavior, they cannot be directly used as a basis for capability reliability assessment. To ensure that the newly calculated distance has a unified scale and cross-scenario comparability, this application needs to normalize it using a statistical benchmark of historical normal behavior. That is, considering that the absolute value of the original anisotropic distance will drift with changes in operating conditions, it is not conducive to setting a robust decision threshold. Specifically, the Z-Score normalization method is used to transform the absolute measure of distance into a relative measure of how many standard deviations it deviates from normal behavior, expressed as: in, This represents the final output, normalized critical weighted convergence distance; The original anisotropic distance representing historical reliable behavior samples. The statistical mean; The original anisotropic distance representing historical reliable behavior samples. The statistical standard deviation.

[0058] More specifically, in a particular example of this application, a specific numerical calculation embodiment is given to aid in the illustration: setting the original anisotropic distance ≈6.838, the statistical mean of the original anisotropic distance of historical reliable behavior samples. =4.5, statistical standard deviation =1.2. Substituting into the normalization formula: Critical weighted convergence distance =(6.838-4.5) / 1.2≈2.338 / 1.2≈1.948. Therefore, after normalization to historical benchmarks, this distance is transformed into a relative measure of approximately 1.95 standard deviations from normal behavior, clearly reflecting the degree of deviation between current behavior and historically reliable behavior, providing a standardized quantitative basis for subsequent AI capability and efficiency assessments.

[0059] Accordingly, by comparing with historical benchmarks, an adaptive scaling of the distance metric was achieved, enabling it to adapt to the normal fluctuation range of different transformer areas and time periods, thus enhancing the robustness of subsequent threshold decisions. This produces a stable, reliable, and statistically significant risk metric, providing high-quality input for subsequent confidence assessments.

[0060] Specifically, the geometric credibility decomposition unit 143 further includes: calculating the angle between the Pareto projection point and the reference vector as the diversity angle. It should be understood that since convergence distance only quantifies the degree to which AI behavior deviates from the Pareto optimal surface, it cannot distinguish whether this deviation is due to inefficient output caused by capability deficiencies or a proactive trade-off by the AI ​​for global optimality. A single distance metric is insufficient to characterize collaborative intent. Therefore, this application further quantifies the AI's trade-off preference direction among multiple objectives by calculating the angle between the Pareto projection point and the reference vector. This allows for accurate identification of whether the AI ​​prioritizes safety and stability, economic benefits, or other objectives, clarifying whether its collaborative intent aligns with the current power grid dispatch strategy, providing a core dimension for subsequent intent compliance assessment, and avoiding system operation interference caused by misjudging reasonable trade-off behaviors.

[0061] Specifically, in a concrete example of this application, a reference vector is first established. Based on the current power grid operating conditions (normal state, emergency state, etc.), the target priorities issued by the dispatch center (such as voltage stability priority in emergency state) are transformed into unit reference vectors in the target space. Then, the vectors corresponding to the Pareto projection points are extracted, and the angle between this vector and the reference vector is calculated through vector dot product operations to clarify the directional differences between the two. Finally, the angle value is calibrated by combining power grid topology constraints and historical trade-off data to eliminate calculation biases caused by differences in vector dimensions, ultimately obtaining diverse angles that can accurately represent target trade-off preferences.

[0062] In the aforementioned trusted computing system 100 for multi-objective collaboration of AI applications in the power Internet of Things (IoT) terminals, the confidence assessment module 150 is used to perform a confidence assessment based on dual constraints on convergence distance and diversity perspectives to obtain a comprehensive trust score. It should be understood that since convergence distance only quantifies the degree to which AI behavior deviates from the optimal benchmark (capability dimension), and diversity perspective only represents objective trade-offs (intent dimension), a single-dimensional assessment cannot comprehensively reflect the credibility of collaborative behavior and is prone to misjudgment due to overlooking capability deficiencies or illegal intentions. Therefore, this application further transforms the two-dimensional indicators into credibility scores and integrates them through a dual-constraint mechanism to achieve a comprehensive assessment of AI capability efficiency and intent compliance. This allows for the formation of a comprehensive trust criterion that considers both capability and intent, accurately distinguishing between different scenarios such as capable and compliant, capable but illegal, and inadequate but compliant, providing a scientific and comprehensive quantitative basis for subsequent trustworthy judgments.

[0063] Figure 8 This is a block diagram of the confidence assessment module in a trusted computing system for multi-objective collaborative AI applications of power IoT terminals according to an embodiment of this application. Figure 8As shown in the embodiments of this application, the confidence assessment module 150 includes: an AI capability efficiency assessment unit 151, used to perform AI capability efficiency assessment based on Gaussian kernel function on convergence distance based on historical confidence distribution benchmark to obtain capability confidence score; an AI intent compliance assessment unit 152, used to perform AI intent compliance assessment based on sector constraints on diverse angles based on current power grid dispatching strategy instructions and legal angle domain to obtain intent confidence score; and a weighted fusion unit 153, used to perform weighted fusion of capability confidence score and intent confidence score to obtain comprehensive trust score.

[0064] Specifically, the AI ​​capability efficiency evaluation unit 151 is used to evaluate the AI ​​capability efficiency based on the convergence distance using a Gaussian kernel function, based on a historical reliable distribution benchmark, to obtain a capability reliability score. It should be understood that since the convergence distance is a raw geometric distance value, without normalization and risk mapping based on historical operating patterns, it cannot directly reflect the quality of AI capabilities. Furthermore, the reasonable distance range varies under different operating conditions, and a single threshold judgment is prone to bias. Therefore, this application further introduces a historical reliable distribution benchmark, using a Gaussian kernel function to transform the distance into a standardized reliability score, thereby quantifying the AI's capability efficiency in approaching the optimal benchmark. This maps the raw distance value to an intuitive score in the 0-1 range, highlighting the difference between normal fluctuations and abnormal deviations, accurately identifying inefficient outputs caused by insufficient capability or malicious attacks, and providing reliable capability dimension support for comprehensive evaluation.

[0065] Specifically, in one example of this application, statistical features of the convergent distance are first extracted from historical reliable data to construct a historical reliable distribution benchmark and determine key parameters such as the distance standard deviation under normal scenarios. Then, the current convergent distance is input into a Gaussian kernel function. Utilizing the nonlinear decay characteristics of this function, the distance value is converted into a preliminary capability score. The closer the distance, the closer the score is to 1; when the distance exceeds a reasonable range, the score drops rapidly. Finally, the preliminary score is calibrated based on the current power grid operating conditions. For example, the score sensitivity is appropriately increased under emergency conditions, and the tolerance is relaxed under normal conditions, ensuring that the score accurately matches the capability assessment requirements under different scenarios.

[0066] Specifically, the AI ​​intent compliance assessment unit 152 is used to perform sector-constraint-based compliance assessments of diverse perspectives based on current power grid dispatching strategy instructions and legal perspective domains to obtain an intent credibility score. It should be understood that since diverse perspectives only reflect the AI's objective trade-off direction and are not associated with power grid dispatching requirements and safety rules, it is impossible to determine whether the trade-off intent meets current operational needs. For example, a trade-off biased towards economic objectives in an emergency may violate safety rules. Therefore, this application further combines dispatching strategy instructions and legal perspective domains, using sector constraint verification and similarity calculation to quantify the compliance of AI intents. This effectively distinguishes between compliant objective trade-offs and non-compliant intent tendencies, ensuring that AI collaborative behavior meets core requirements such as power grid safety and dispatching priority, and avoiding power grid operation risks caused by intent violations.

[0067] Specifically, in a concrete example of this application, firstly, based on the current power grid dispatching strategy instructions, such as priority voltage stabilization and priority economy, and in conjunction with safe operation rules, the allowed legal angle domain, i.e., the compliant sector, is defined. Then, it is verified whether the diverse angles fall within the legal sector. If they exceed this range, the intent is directly determined to be non-compliant, and the score is 0. If they are within the legal sector, the cosine similarity between the angle and the angle guided by the dispatching strategy is calculated. The higher the similarity, the more the intent aligns with dispatching requirements. Finally, the cosine similarity value is used as the intent credibility score, ensuring that the score directly reflects the compliance and fit of the intent.

[0068] Specifically, the weighted fusion unit 153 is used to weight and fuse the capability credibility score and the intent credibility score to obtain a comprehensive trust score. It should be understood that the importance of AI capability efficiency and intent compliance varies under different power grid operating conditions. For example, in an emergency, intent compliance (such as prioritizing safety) is more critical than capability efficiency, while in a normal state, both are equally important. Simply adding them together would lead to an imbalance in the evaluation focus. Therefore, this application further determines dynamic weights based on real-time operating conditions and weights and fuses the two scores to balance the evaluation weights of capability and intent. This results in a comprehensive trust score adapted to real-time power grid operating conditions, avoiding inefficient collaboration due to neglecting capability deficiencies and preventing security risks caused by intent violations, thus ensuring the scientific rigor and relevance of the comprehensive evaluation.

[0069] Specifically, in a concrete example of this application, the current operating state (normal, emergency, recovery, etc.) is first determined by the real-time power grid operating condition identification module. Based on preset rules, the dynamic weights of the capability score and intent score are determined; for example, the intent weight is 0.7 and the capability weight is 0.3 in the emergency state, while both are 0.5 in the normal state. Subsequently, the two credibility scores are linearly weighted according to their weights to obtain a preliminary fusion score. Finally, the preliminary score is normalized and mapped to the 0-1 range to ensure a consistent and intuitive score range, facilitating subsequent comparison and judgment with safety thresholds.

[0070] In the aforementioned trusted computing system 100 for multi-objective collaboration of AI applications in the power IoT terminal, the trusted decision module 160 is used to make a trusted decision based on the comprehensive trust score according to a security threshold to obtain a trusted decision result. It should be understood that since the comprehensive trust score is only a quantitative evaluation result in the 0-1 range, it lacks a clear judgment boundary and cannot be directly converted into a practical decision on whether to allow the execution of AI control commands. Furthermore, the power system has stringent requirements for the security and continuity of control commands, and ambiguous evaluation conclusions can lead to security risks or hinder normal collaboration. Therefore, this application further establishes a binary decision logic through a quantitative comparison of the comprehensive trust score and the security threshold to clearly distinguish between trusted collaborative behavior and untrustworthy abnormal behavior. This provides a clear decision basis for AI command execution, accurately intercepts dangerous commands caused by malicious attacks or capability defects, and avoids misjudging normal collaborative behavior, ensuring the safety, stability, and continuity of power grid operation.

[0071] Specifically, in a concrete example of this application, the system first adapts the corresponding safety thresholds based on the real-time operating conditions of the power grid (normal, emergency, and recovery states). Then, it performs a precise numerical comparison between the comprehensive trust score and the safety thresholds for the current operating conditions. Finally, it executes a decision based on the comparison results: if the score is below the threshold, the command is immediately intercepted and switched to a preset backup traditional controller, such as PID control. If the score is above the threshold, the command is deemed trustworthy, but when actually issued to power infrastructure such as relay protection, the system strictly adheres to the principle of least privilege, restricting the AI ​​application to only calling the API permissions necessary to execute the current task. Simultaneously, for critical operations that may trigger power outages, the system enforces a dual authorization mechanism: after local approval, it must undergo secondary verification and approval by an upper-level management node (such as the cloud or gateway) before the final control signal can be issued to the underlying infrastructure. Furthermore, regardless of the judgment result, the system generates a judgment flag and records operating condition information, storing trustworthy samples in a positive sample library for model iteration.

[0072] In summary, the trusted computing system for multi-objective collaborative AI applications in the power IoT terminal, according to embodiments of this application, is explained. First, it comprehensively collects local operating data, neighbor collaborative states, and environmental context. After cleaning and heterogeneous fusion, a collaborative feature tensor representing the real-time power grid operating conditions is constructed. Then, this tensor is used to dynamically fit a collaborative reference surface in a multi-dimensional objective space, generating a dynamic Pareto front reflecting the current optimal trade-off. Next, the actual control output of the terminal AI is projected onto this front. By analyzing its convergence distance and vector deviations from diverse angles, it accurately distinguishes between reasonable concessions for global optimization and malicious performance degradation in multi-objective conflict scenarios. Finally, a confidence assessment and security decision are made based on a dual-constraint mechanism for the deviations. This effectively overcomes the limitations of traditional isotropic measurements in perceiving physical risk heterogeneity, achieving precise and dynamic protection for the collaborative behavior of power IoT terminal AI.

[0073] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An AI application trusted computing system for power Internet of Things terminal multi-target cooperation, characterized in that, The method comprises the following steps: A multi-source collaborative perception module is used to obtain local terminal operation data, collaborative neighbor terminal state data, and environmental context data; A heterogeneous data fusion module is used to perform data cleaning, heterogeneous feature extraction, and collaborative situation splicing on the local terminal operation data, collaborative neighbor terminal state data, and environmental context data to obtain a collaborative feature tensor; A reference surface dynamic fitting module is used to perform multi-objective collaborative reference surface dynamic fitting on the collaborative feature tensor to obtain a dynamic Pareto front; An action mapping and trust domain projection module is used to perform collaborative action vector mapping and trust domain projection on the actual control output of the terminal AI application based on the dynamic Pareto front to obtain a convergence distance and a diversity angle; A confidence evaluation module is used to perform double-constraint-based confidence evaluation on the convergence distance and the diversity angle to obtain a comprehensive trust score; A trusted decision module is used to perform trusted decision on the comprehensive trust score based on a security threshold to obtain a trusted decision result.

2. The AI application trusted computing system for power internet of things terminal multi-target cooperation according to claim 1, characterized in that, The heterogeneous data fusion module comprises: A space-time synchronization preprocessing unit is used to perform space-time alignment and normalization preprocessing on the local terminal operation data, collaborative neighbor terminal state data, and environmental context data to obtain a synchronization state vector; A time sequence feature extraction unit is used to perform time sequence feature extraction on the synchronization state vector based on a lightweight CNN-1D to obtain an abstract time sequence feature matrix; A multi-dimensional collaborative situation splicing unit is used to perform multi-dimensional collaborative situation splicing on the abstract time sequence feature matrix to obtain the collaborative feature tensor.

3. The AI application trusted computing system for power internet of things terminal multi-target cooperation according to claim 2, characterized in that, The multi-dimensional collaborative situation splicing unit comprises: A feature decoupling subunit is used to perform feature decoupling on the abstract time sequence feature matrix to obtain local self features, neighbor interaction features, and environmental constraint features; A feature splicing subunit is used to perform feature splicing on the local self features, neighbor interaction features, and environmental constraint features based on a topological embedding vector to obtain the collaborative feature tensor according to the following formula: wherein, is a co-feature tensor, is a concatenation operation along the feature channel dimension, is a local self-feature, is a neighbor interaction feature, is an environment constraint feature, is a topological embedding vector.

4. The AI application trusted computing system for power internet of things terminal multi-target cooperation of claim 1, wherein The reference surface dynamic fitting module comprises: A multi-objective reference vector prediction unit is used to perform multi-objective reference vector prediction on the collaborative feature tensor based on a meta-optimizer based on a pre-defined optimization objective function set to obtain a discrete Pareto optimal reference vector set; A front surface fitting unit is used to perform front surface fitting on the discrete Pareto optimal reference vector set based on weighted Chebyshev quantization to obtain a dynamic Pareto front.

5. The power Internet of Things terminal multi-target cooperative AI application trusted computing system according to claim 1, characterized in that, The action mapping and trust domain projection module comprises: An action effect mapping unit is used to perform action effect mapping on the actual control output of the terminal AI application based on a physical agent model to obtain an actual effect vector; A Pareto projection unit is used to perform Pareto projection on the dynamic Pareto front and the actual effect vector based on Euclidean distance minimization to obtain a Pareto projection point; A geometric trustworthiness decomposition unit is used to perform geometric feature decomposition and trustworthiness calculation on the Pareto projection point and the actual effect vector to obtain a convergence distance and a diversity angle.

6. The power Internet of Things terminal multi-target cooperative AI application trusted computing system according to claim 5, characterized in that, The geometric trustworthiness decomposition unit comprises: A risk quantization subunit is used to perform risk quantization on the actual effect vector based on a power grid operation condition flag to obtain a dynamic criticality weight vector; The anisotropic distance calculation subunit is configured to perform anisotropic distance calculation on the Pareto projection point and the actual effect vector based on the dynamic criticality weight vector to obtain an original anisotropic distance. The distance normalization subunit is configured to perform history reference based distance normalization on the original anisotropic distance to obtain the convergence distance.

7. The power Internet of Things terminal multi-target cooperative AI application trusted computing system according to claim 5, characterized in that, The geometric credibility decomposition unit comprises: calculating an angle value between the Pareto projection point and a reference reference vector as the diversity angle.

8. The power Internet of Things terminal multi-target cooperative AI application trusted computing system according to claim 1, characterized in that, The credibility evaluation module comprises: The AI capability efficiency evaluation unit is configured to perform Gaussian kernel function based AI capability efficiency evaluation on the convergence distance based on a history credible distribution reference to obtain a capability credibility score. The AI intent compliance evaluation unit is configured to perform sector constraint based AI intent compliance evaluation on the diversity angle based on a current power grid dispatching strategy instruction and a legal angle domain to obtain an intent credibility score. The weighted fusion unit is configured to perform weighted fusion on the capability credibility score and the intent credibility score to obtain a comprehensive trust score.

Images