A transformer holographic intelligent test method and system

Abstract

The application provides a transformer holographic intelligent test method and system, belongs to the field of transformer testing and state evaluation, and is used for solving the problems of test system module fragmentation, information island, low data reliability and low scheduling efficiency in the related art. The application unifies the representation of test resource state, transformer physical behavior, safety risk and health degradation trend by constructing a physical information neural random field global model; realizes model collaborative updating under multi-station data privacy protection by using a federal security multi-party computing protocol; realizes trusted evidence preservation of key states and token economy closed loop by using a trusted execution environment and a directed acyclic graph-based distributed ledger; realizes adaptive optimization of resource scheduling by coupling Bayesian optimization and token price dynamics. The application can significantly improve the intelligent level, data reliability and resource utilization efficiency of transformer testing.

Description

Technical Field

[0001] This application relates to the technical field of transformer testing and condition assessment, and in particular to a transformer holographic intelligent testing method and system. Background Technology

[0002] Currently, the field of transformer testing is developing towards intelligence and automation, with various testing equipment and technologies emerging to improve testing efficiency, ensure testing safety, and enhance diagnostic accuracy.

[0003] Existing transformer testing systems typically employ a modular design, distributing functions such as scheduling management, safety monitoring, performance testing, and health diagnosis across different subsystems or independent devices. The scheduling system is responsible for resource allocation and task sequencing, the safety system for real-time monitoring and early warning, the testing system for measuring various electrical parameters, and the diagnostic system for condition assessment and lifespan prediction. These systems interact via data interfaces, forming a relatively loosely coupled relationship.

[0004] However, this distributed technical architecture has inherent flaws. Scheduling decisions are based solely on the current task queue and device status, failing to detect changes in the security situation during the experiment; security warnings rely on preset thresholds, making it difficult to match with the dynamic characteristics of the experiment process; health diagnostics are often performed offline after the experiment, and the diagnostic results cannot be fed back in real time, affecting the current experimental strategy; the fusion and processing of multi-source heterogeneous data lacks a unified probabilistic framework, leading to information loss and accumulated uncertainty. These problems make it difficult for existing systems to achieve global optimization and dynamic coordination throughout the entire experiment, hindering further improvements in experimental efficiency and diagnostic accuracy. Summary of the Invention

[0005] This application provides a holographic intelligent testing method and system for transformers, which can achieve integrated characterization and collaborative optimization of test resource scheduling, safety situation awareness, physical behavior modeling and health degradation diagnosis using a unified probabilistic graphical model, fundamentally solving the problems of information loss and suboptimal decision-making caused by the fragmentation of multiple modules.

[0006] Firstly, this application provides a holographic intelligent testing method for transformers. A physical information neural random field global model is constructed. This global model is a unified probabilistic graphical model used to jointly characterize the state of test resources, transformer physical behavior, safety risks, and health degradation trends. The model describes the dependencies between variables through an energy function. These variables include input observation data, a first latent variable representing the interaction graph structure between test resources and tasks, a second latent variable representing the spatiotemporal distribution of the transformer's physical field state, and output variables. The output variables include scheduling decisions, safety risk indices, health indices, and remaining lifetime distribution. A unique identification code for the transformer under test is obtained. Based on this unique identification code, the corresponding test task book is retrieved from the database, and current environmental parameters are collected as initial input observation data. This initial input observation data is input into the physical information neural random field global model. Output variables are obtained through model inference, and corresponding test operations are executed based on these output variables. During the execution of the test operations, real-time observation data is continuously collected, and the physical information neural random field global model is updated based on this real-time observation data. The output variables are updated through model inference to achieve dynamic optimization and control of the test process.

[0007] By adopting the above technical solutions, scheduling, safety, physical and health information that were originally scattered in different subsystems are unified under the same probabilistic graphical model framework. The dependency relationship between variables is established through energy functions, which enables scheduling decisions to perceive the safety situation and health status, safety early warning to be dynamically adjusted in combination with the test process, and health diagnosis to provide real-time feedback to affect the test strategy. This achieves multi-objective global optimization and dynamic closed-loop control, significantly improving the intelligence level and collaborative efficiency of the test system.

[0008] Furthermore, the energy function is composed of a weighted sum of a graph energy term, a physical energy term, and a coupled energy term; wherein the graph energy term models the experimental resource topology through a graph neural network to characterize the interaction relationships between resource nodes; the physical energy term embeds partial differential equation constraints through a physical information neural network to characterize the evolution law of the transformer's physical field; and the coupled energy term connects the graph structure and the physical field to output the final scheduling decision, safety risk index, health index, and remaining lifetime distribution.

[0009] By adopting the above technical solution, the energy function is decomposed into three components with clear physical meaning. The graph energy term captures the topological dependence between resources, the physical energy term embeds the electromagnetic and thermodynamic multi-field coupling law, and the coupling energy term establishes the correlation mapping between resource state and physical field. This not only ensures the interpretability of the model, but also realizes the organic integration of multi-source heterogeneous information.

[0010] Furthermore, the graph energy term is constructed based on a graph neural network structure and includes node potential functions and edge potential functions. The node potential function is used to describe the state characteristics of a single resource node, and the edge potential function is used to describe the interaction relationship between resource nodes. Both the node potential function and the edge potential function are parameterized by the neural network.

[0011] By adopting the above technical solution, graph neural networks are used to perform structured modeling of experimental resources. The node potential function captures individual features such as equipment health, location, and busy / idle status, while the edge potential function learns the collaborative relationships and conflict constraints between resources, thereby achieving adaptive representation and dynamic updating of complex experimental resource topology.

[0012] Furthermore, the physical energy term is constructed based on physical operators, which are used to calculate the residuals of the partial differential equations of the transformer's electromagnetic field, temperature field, or stress field, and to embed physical laws into the model by minimizing the residuals.

[0013] By adopting the above technical solution, physical priors such as Maxwell's equations and heat conduction equations are embedded into the model training process in the form of soft constraints. This allows the model to maintain physical consistency in sparse data regions while allowing a certain degree of physical deviation to fit the actual measurement data, thus achieving a dynamic balance between physical laws and data-driven approaches.

[0014] Furthermore, the update of the physical information neural random field global model adopts a federated secure multi-party computation protocol, specifically including: each experimental station calculates the model gradient in its local trusted execution environment, and uses secret sharing to split the gradient into multiple parts and distribute them to the trusted execution environments of other stations; the trusted execution environments of each station securely aggregate the received gradient parts and send the aggregation results to the central cluster; the central cluster reconstructs the global gradient in the trusted execution environment, updates the model parameters, and broadcasts the updated model encrypted to the trusted execution environments of each station.

[0015] By adopting the above technical solution, the gradient is split into multiple parts using secret sharing, ensuring that no single node can obtain the complete gradient information; combined with the hardware isolation and remote proof capabilities of the trusted execution environment, multi-party collaborative training is achieved while ensuring data privacy, thus resolving the contradiction between data silos and privacy protection among multiple test stations.

[0016] Furthermore, the scheduling decision is generated using a Bayesian optimization method, specifically including: defining a collection function that balances expected return and prediction uncertainty, wherein the expected return and prediction uncertainty originate from the output of the physical information neural random field model; and selecting the optimal action as the scheduling decision by maximizing the collection function.

[0017] By adopting the above technical solution, the uncertainty of the model output prediction guides the exploration direction of scheduling decisions. It actively explores areas with unknown benefits to reduce model uncertainty, achieves a dynamic balance between exploration and utilization, and avoids the shortcomings of traditional reinforcement learning that requires a large number of trial and error samples.

[0018] Furthermore, the method for generating the health index and remaining lifespan distribution includes: constructing a stochastic differential equation to describe the health degradation trajectory based on the posterior distribution of the health index output by the physical information neural random field model, wherein the stochastic differential equation includes a drift coefficient and a diffusion coefficient; solving the first arrival time of the stochastic differential equation to obtain the probability density distribution of the remaining lifespan, wherein the failure threshold is used to define the critical value of health index failure.

[0019] By adopting the above technical solution, the posterior distribution of the health index is used as the initial condition of the stochastic differential equation. The drift coefficient characterizes the deterministic degradation trend, the diffusion coefficient represents the random fluctuation, and the first arrival time solution gives the full probability distribution of the remaining lifespan rather than a point estimate, providing confidence information for risk decision-making.

[0020] Furthermore, it also includes a critical state notarization step: generating a verifiable computational proof for the critical state within a local trusted execution environment, the verifiable computational proof being used to prove that the critical state is correctly calculated by the current model under the input observation data; and submitting the verifiable computational proof and the hash value of the critical state to a distributed ledger for notarization.

[0021] By adopting the above technical solution, the remote authentication capability of the trusted execution environment is used to generate proof of the computation process. Combined with the immutability of blockchain, the trusted storage and post-audit of key states are realized, thus solving the problems of the credibility and traceability of experimental data.

[0022] Furthermore, it also includes: generating verifiable computational proofs for key states within a local trusted execution environment, and submitting the verifiable computational proofs and the hash values ​​of the key states to a distributed ledger for storage; the distributed ledger adopts a consensus mechanism based on a directed acyclic graph and manages the issuance and transfer of experimental resource tokens through smart contracts; the tokens are automatically triggered for transfer based on the actual resource consumption during the experiment, and the token price is fed back to the acquisition function through an oracle.

[0023] By adopting the above technical solution, a token economy mechanism is introduced to quantify resource consumption into tradable tokens. The token price is fed back to the Bayesian-optimized collection function in real time through an oracle, enabling scheduling decisions to perceive changes in the supply and demand relationship of resources, and realizing closed-loop coupling and self-organizing optimization of resource pricing and scheduling.

[0024] Secondly, this application provides a transformer holographic intelligent testing system. It includes: a distributed intelligent testing unit comprising a main control console, a mobile characteristic testing vehicle, and a mobile temperature rise testing vehicle; each unit has a built-in edge computing module supporting a trusted execution environment for local data acquisition, model inference, and local federated learning computation; a central collaborative computing cluster comprising a high-performance server for training a global model of a physical information neural random field, a federated learning aggregation server, blockchain nodes, and a trusted execution environment verification service; a global perception unit comprising infrared sensors, video surveillance equipment, environmental sensors, equipment status monitoring terminals, and a wireless emergency stop device; a trusted execution environment network composed of the trusted execution environments of all computing nodes, establishing trusted channels through remote proof; and a blockchain and token network employing a distributed ledger based on a directed acyclic graph for recording key state hashes and token transactions; the system is used to execute the transformer holographic intelligent testing method as described in any one of the first aspects above.

[0025] By adopting the above technical solutions, a complete hardware support system has been constructed, from edge awareness, local computing, federated aggregation to blockchain notarization. The trusted execution environment network ensures data security and computational trustworthiness, and the DAG blockchain provides high-throughput notarization capabilities, providing a reliable physical foundation for the aforementioned methods.

[0026] In summary, this application has at least the following beneficial effects:

[0027] A holographic intelligent testing method and system for transformers is provided, which realizes unified representation and collaborative optimization of multi-source information through physical information neural random fields;

[0028] The federally secure multi-party computation and trusted execution environment ensure data privacy and computational trustworthiness during multi-site collaborative training.

[0029] By using a token economy mechanism, a closed-loop coupling between resource pricing and scheduling decisions is achieved, driving the system's self-organizing optimization.

[0030] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 A schematic diagram of a transformer holographic intelligent testing system according to an embodiment of this application is shown.

[0033] Figure 2 A flowchart of a transformer holographic intelligent testing method according to an embodiment of this application is shown. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0036] This application provides a holographic intelligent testing method and system for transformers, which realizes unified representation and collaborative optimization of multi-source information through physical information neural random fields. While ensuring data privacy and security, it achieves trusted evidence storage and self-organizing resource scheduling, significantly improving testing efficiency and diagnostic accuracy.

[0037] In one aspect, embodiments of this application disclose a transformer holographic intelligent testing system.

[0038] Figure 1 This illustration shows a schematic diagram of a transformer holographic intelligent testing system according to an embodiment of this application. In a typical transformer manufacturing plant or power maintenance center, multiple testing stations are usually set up to simultaneously conduct routine and type tests on multiple transformers. Each testing station is equipped with corresponding testing equipment and wiring devices. The testing area is typically divided into a test preparation area, a test execution area, and a central control area, with physical isolation and signal shielding measures ensuring safety between these areas. This embodiment of the application is deployed and implemented in this real-world application scenario.

[0039] Reference Figure 1 The system includes a distributed intelligent experimental unit, a central collaborative computing cluster, a global perception unit, a trusted execution environment network, and a blockchain and token network, which together constitute an experimental support system integrating data acquisition, edge computing, collaborative training, trusted storage, and resource scheduling.

[0040] The distributed intelligent test unit comprises a main control console, a mobile characteristic testing vehicle, and a mobile temperature rise testing vehicle. Each unit integrates an edge computing module supporting a trusted execution environment. The main control console is fixedly installed in the central control area, equipped with an industrial-grade touchscreen display and operating console, serving as the human-machine interface for test personnel. The mobile characteristic testing vehicle and the mobile temperature rise testing vehicle utilize an electric remote-controlled chassis, allowing flexible movement between the test preparation area and the test execution area. The vehicles integrate various testing instruments, such as DC resistance testers, turns ratio testers, power analyzers, and temperature monitoring instruments. The edge computing module employs an industrial-grade embedded computer, such as an Intel Xeon D-series processor or an NVIDIA Jetson AGX Orin module, with integrated Intel SGX or ARM TrustZone trusted execution environment technology. This edge computing module is responsible for performing local computational tasks required for data acquisition, model inference, and federated learning. The trusted execution environment ensures the confidentiality and integrity of the computation process and data, guaranteeing that even if the physical device is accessed without authorization, the model parameters and sensitive data stored within the module will not be leaked. The main control console serves as the test command center, establishing a real-time communication connection with the mobile test vehicle via a 5G industrial private network or WiFi 6 wireless network. The bandwidth configuration is no less than 100Mbps, and the latency is controlled within 20ms to ensure the issuance of real-time control commands and the transmission of test data, and to coordinate the collaborative work of various units.

[0041] The central collaborative computing cluster is deployed in an enterprise data center or a dedicated edge computing node room, connected to the network equipment in the test site via a dedicated fiber optic line. This cluster includes high-performance servers, federated learning aggregation servers, blockchain nodes, and a trusted execution environment verification service. The high-performance servers are equipped with at least four NVIDIA A100 or equivalent GPUs, with at least 512GB of system memory, used to train a global model of a physical information neural random field. The training cycle can be set weekly or monthly depending on the data scale, or incremental training can be triggered based on data from newly connected test stations. The federated learning aggregation server is responsible for collecting gradient updates uploaded by each test station and performing secure aggregation. It employs a secret sharing scheme from a secure multi-party computation protocol, splitting the gradient into multiple parts and aggregating them separately, ensuring that no single node can obtain the complete gradient information. Blockchain nodes participate in the consensus and maintenance of the distributed ledger, implemented using a consensus algorithm written in Rust. A single node with a 4-core CPU and 16GB of memory is sufficient for operation. The Trusted Execution Environment (TEE) verification service verifies the remote proof reports of the participants. The verification reports adopt the IETF RFC 9334 standard format to ensure that the TEE of each participating node is in the expected, tamper-free state. The central cluster is connected to the distributed intelligent experimental unit via a 10 Gigabit fiber optic network, receives local processing results, and distributes updated global models. The model distribution process uses end-to-end encryption, and the key is generated by the recipient's TEE and securely exchanged through the remote proof channel.

[0042] The comprehensive sensing unit includes infrared sensors, video surveillance equipment, environmental sensors, equipment status monitoring terminals, and a wireless emergency stop device. The infrared sensors employ dual-element pyroelectric infrared detection technology, with a detection distance of at least 15 meters and a detection angle of 110 degrees. They are deployed at the boundary fences and key passageways of the test area, spaced 8 to 10 meters apart to form a comprehensive coverage without blind spots, used to monitor unauthorized personnel entry. The video surveillance equipment uses 4-megapixel high-definition network cameras, supporting H.265 encoding and the ONVIF protocol. Deployed directly above and to the side of the test stations, they capture real-time footage and support edge AI analysis, capable of identifying wiring status, personnel actions, and abnormal events. The environmental sensors integrate temperature, humidity, and barometric pressure sensors. The temperature measurement range is -40°C to 85°C with an accuracy of ±0.3°C; the humidity measurement range is 0% to 100%RH with an accuracy of ±2%RH. They are deployed near the test stations. The equipment status monitoring terminal includes a vibration acceleration sensor and an ultra-high frequency partial discharge sensor. The vibration sensor has a range of ±16g and a frequency response range of 0.5Hz to 1kHz. It is magnetically mounted on the surface of the transformer tank. The partial discharge sensor uses an external antenna and a detection frequency band of 300MHz to 1.5GHz to collect the operating status of the test equipment. The wireless emergency stop device uses 2.4GHz industrial-grade wireless communication with a response time of less than 100ms. Each test station is equipped with two handheld emergency stop buttons and one wall-mounted emergency stop button, providing remote emergency power-off functionality. All sensing data is transmitted in real time to the edge computing module of the distributed intelligent test unit via Modbus TCP or MQTT protocol. The transmission period can be configured from 100ms to 1s, serving as input for model inference.

[0043] The Trusted Execution Environment (TEE) network consists of the TEEs of all computing nodes, including edge computing modules in the distributed intelligent experimental unit, various servers in the central collaborative computing cluster, and TEEs in the blockchain nodes. Each node verifies its identity and the trustworthiness of its computing environment through a remote verification protocol. The verification process employs either EPID group signatures or DAA direct anonymity authentication. After successful verification, a TLS-encrypted trusted communication channel is established. The channel session key is negotiated and generated by both TEEs and automatically rotates every 24 hours or when the transmitted data volume reaches 1GB, ensuring the security of cross-node data transmission and collaborative computing.

[0044] The blockchain and token network employs a distributed ledger based on a directed acyclic graph (DAG), utilizing solutions such as IOTA's coordinator scheme or Avalanche's consensus protocol. Deployed on blockchain nodes within a central collaborative computing cluster, it consists of at least seven nodes forming a consortium blockchain. Employing a practical Byzantine fault-tolerant variant consensus algorithm, the network ensures a single block confirmation time of no more than 2 seconds and can process at least 1000 transactions per second. This network records the hash values ​​and verifiable computational proofs of critical states. The hash algorithm used is either SHA256 or SM3 (China's national cryptographic algorithm), and the verifiable computational proofs are generated using zk-SNARK or zk-STARK technology, with a proof size limited to less than 1KB. Simultaneously, the network manages the issuance, transfer, and settlement of experimental resource tokens. These tokens utilize an ERC20-compatible interface, with each token representing the right to use 1 kilowatt-hour of electricity or 1 minute of test vehicle man-hours. The smart contracts are written in Solidity or Rust and automatically execute token transfer rules. For example, after a mobile test vehicle completes a DC resistance test task, the smart contract automatically transfers a predetermined number of tokens from the task issuer's account to the test vehicle's account. The oracles utilize a decentralized oracle network such as Chainlink or Band Protocol to provide real-time feedback of on-chain token prices to the scheduling decision module. The price update frequency is no less than once per minute, forming a closed-loop interaction between on-chain and off-chain operations.

[0045] The aforementioned units work collaboratively: the distributed intelligent testing unit is responsible for on-site data acquisition and local computation; the omni-channel perception unit provides environmental and safety monitoring data; the trusted execution environment network ensures the trustworthiness of data transmission and computation; the central collaborative computing cluster completes global model training and federated aggregation; and the blockchain and token network realize key state notarization and resource token management, thus providing a complete hardware support environment for the transformer holographic intelligent testing method. In actual deployment, the equipment configuration and number of nodes can be flexibly adjusted according to the scale of the test and the investment budget. For example, small test stations can reduce the number of mobile test vehicles and deploy the central cluster using virtualization, while large test bases can expand the sensor deployment density and increase the number of blockchain nodes, but the core architecture and functional modules remain unchanged.

[0046] With the aforementioned hardware support, in a second aspect, this application discloses a holographic intelligent testing method for transformers.

[0047] Figure 2 A flowchart of a transformer holographic intelligent testing method according to an embodiment of this application is shown.

[0048] Reference Figure 2 The method specifically includes the following steps:

[0049] S1: Construct a global model of a physical information neural random field.

[0050] The physical information neural random field global model is a unified probabilistic graphical model used to jointly characterize the state of test resources, transformer physical behavior, safety risks, and health degradation trends. The model describes the dependencies between variables through an energy function. The variables include input observation data, a first latent variable representing the interaction graph structure between test resources and tasks, a second latent variable representing the spatiotemporal distribution of the transformer physical field state, and output variables. The output variables include scheduling decisions, safety risk index, health index, and remaining lifetime distribution.

[0051] In this step of the method, the core of the physical information neural random field global model lies in incorporating the originally scattered scheduling, physical, security, and health information into the same probabilistic framework through a unified energy function. This model first defines the input observation dataset. This dataset originates from real-time data collected by the global sensing unit and the distributed intelligent experimental unit. Specifically, It includes data on the status of test resources, such as the current location, busy / idle status, and equipment health of the mobile test vehicle, as well as environmental parameters such as temperature and humidity, and electrical quantity data such as voltage, current, and partial discharge collected in real time by sensors. These data form a sequence in the time dimension and are associated with different resource nodes and measurement points in the spatial dimension.

[0052] The model introduces the first latent variable. This graph structure is used to characterize the interaction relationships between experimental resources and tasks. It uses experimental resources as nodes and collaborative relationships and task dependencies between resources as edges, and is dynamically constructed using a graph neural network. Node feature vectors include resource type, capability parameters, and real-time status, while edge feature vectors include communication latency, physical distance, and task coupling degree. The graph structure is not statically preset but evolves in real-time based on current task requirements and resource availability, enabling scheduling decisions to adapt to dynamically changing experimental conditions.

[0053] The model introduces a second latent variable. This variable is used to characterize the spatiotemporal distribution of the physical field state of a transformer. It encompasses multidimensional distribution information of the electromagnetic field, temperature field, and stress field within the transformer, providing a deep characterization of the transformer's physical behavior. The evolution of the physical field follows the physical laws described by partial differential equations such as Maxwell's equations, the heat conduction equation, and the elasticity equation. The dimension depends on the spatial discretization accuracy and the time sampling frequency. Finite element meshes or regular meshes are usually used to divide the three-dimensional structure of the transformer, with each mesh point corresponding to multiple physical field components.

[0054] Model output variables It comprises four core components: scheduling decision This refers to the allocation plan of experimental resources and the order of task execution at the next moment; safety risk index. This represents the real-time risk level during the current trial process; health index Quantify the current health status of the transformer; remaining lifetime distribution. The expected time for the transformer to continue operating safely and its probability distribution are given. These four output variables together constitute the decision-making basis and state assessment results of the test system.

[0055] The dependencies between all the above variables are expressed through the energy function. A unified model is used. The smaller the value of the energy function, the higher the joint probability of the variable combinations. The joint probability distribution of the model is expressed as...

[0056]

[0057] in The partition function is normalized for all possible combinations of variables. This probabilistic graphical model framework enables the system to perform operations on given observation data. Under these conditions, by minimizing energy or through probabilistic reasoning, we can simultaneously obtain optimal scheduling decisions, accurate safety assessments, and reliable health diagnoses.

[0058] In this step, the energy function is constructed by a weighted sum of the graph energy term, the physical energy term, and the coupling energy term, i.e.

[0059]

[0060] The weighting coefficients before each energy term can be adjusted according to the actual application scenario. For example, when the data noise is large, the weight of the physical energy term can be appropriately increased, and when resources are scarce, the weight of the graph energy term can be increased.

[0061] The graph energy term The topology of experimental resources is modeled using graph neural networks to characterize the interactions between resource nodes. The graph energy term, constructed based on the graph neural network structure, includes node potential functions and edge potential functions, and its mathematical form is...

[0062]

[0063] in It is a collection of resource nodes, each node corresponding to a test device such as a mobile test vehicle, a main control console or a power cabinet; This is the set of edges between resource nodes, representing communication links, physical location relationships, or task dependencies between devices. Node potential function. Used to describe the state characteristics of a single resource node, the input is the first... The potential states of each node and observation features , It is a low-dimensional embedding vector. Includes device health, busy / idle status, location coordinates, etc. The parameterization of the multilayer perceptron network outputs a scalar energy value, reflecting the degree of deviation between the current state of the node and the ideal state. (Boundary potential function) Used to describe the interaction relationships between resource nodes, the input is the potential state of two connected nodes. and and the corresponding observation features and It outputs a scalar energy value, reflecting the degree of cooperation or conflict between the two nodes. It is also parameterized by a neural network. By minimizing the graph energy term, the model tends to select a scheduling scheme that ensures good cooperation between resources and smooth task execution.

[0064] The physical energy term A physical information neural network embedding partial differential equation constraints is used to characterize the evolution of the physical field of a transformer. The physical energy term is constructed based on physical operators, and its mathematical form is as follows:

[0065]

[0066] in For a moment The physical field state tensor For a moment The observational data includes boundary conditions and excitation sources. Physical operators. This is used to calculate the residuals of partial differential equations for the electromagnetic, temperature, or stress fields of a transformer. Taking the electromagnetic field as an example, the physical operator is constructed based on Maxwell's equations, specifically in the form of...

[0067]

[0068] in The magnetic field strength, For current density, For electric displacement vectors, these physical quantities are all derived from... It is obtained through neural network decoding. For the temperature field, the physical operator is based on the heat conduction equation.

[0069]

[0070] in For temperature, For density, For specific heat capacity, Thermal conductivity, As a heat source, these parameters are also determined by Decode and combine The material properties were calculated. By minimizing the physical energy term, the model forces... The evolution of the model satisfies the constraints of physical laws, ensuring that the physical fields predicted by the model remain physically consistent even in regions with sparse observation data. The residual calculation of the physical energy term employs automatic differentiation techniques, utilizing the differentiability of spatial and temporal coordinates in neural networks to directly calculate partial derivatives, eliminating the need for grid discretization and difference approximation in traditional numerical methods.

[0071] The coupling energy term The connection between the graph structure and the physical field is used to output the final scheduling decision, safety risk index, health index, and remaining lifetime distribution. The coupled energy term takes the form of...

[0072]

[0073] in For scheduling decision decoders, the latent states of the graph structure are... This is mapped to specific scheduling actions, such as test vehicle allocation schemes and task execution order; For safety risk assessors, the input is the physical field state. and real-time observation It outputs a continuous safety risk index, with the risk index range normalized to between 0 and 1. The higher the value, the greater the risk. For health index evaluators, the input is the physical field state. Output a health index scalar, also normalized to between 0 and 1, with higher values ​​indicating better health. Remaining lifespan distribution. The health index is not directly included in the norm loss of the coupled energy term, but is obtained through time-series extrapolation of the health index. By minimizing the coupled energy term, the model ensures that the output scheduling decision is consistent with the graph structure state, the safety risk index is consistent with the physical field state, and the health index is consistent with the physical field state, thereby achieving consistent fusion of multi-source information and decision output.

[0074] The training objective of the entire Physical Information Neural Random Field global model is to minimize the energy function, i.e., to find the optimal model parameters that minimize the average energy value on the training data. The training process employs a variational Bayesian method, approximating the posterior distribution by maximizing the lower bound of evidence, while considering the balance between graph energy, physical energy, and coupling energy. The pre-training phase initializes the model using historical experimental data, and subsequent updates and optimizations are achieved through federated learning in practical applications. Once the model is built, it is stored on a high-performance server in a central collaborative computing cluster and encrypted and distributed to the edge computing modules of various distributed intelligent experimental units via a trusted execution environment network, providing a foundation for subsequent real-time inference and decision-making.

[0075] S2: Obtain the unique identification code of the transformer under test, retrieve the corresponding test task book from the database based on the unique identification code, and collect the current environmental parameters as the initial input observation data.

[0076] In this step, after the transformer under test is transported to the testing station, the system first obtains the transformer's unique identification code using an RFID reader or QR code scanner deployed at the station entrance. This identification code is typically the transformer's factory serial number or asset number, stored in UUID format, such as a 32-bit hexadecimal string, ensuring that each transformer has unique identification globally. Obtaining the unique identification code is a prerequisite for triggering all subsequent operations. The system uses this identification code to establish a link between this test and the transformer's historical data, achieving traceability of the test data.

[0077] The system retrieves the corresponding test task book from the enterprise resource planning system or local test database based on the unique identification code. The task book is stored in JSON or XML format, and after parsing, a set of test task parameters is obtained. The test specification sheet includes the transformer's model and specifications, such as rated capacity. Rated voltage Rated current It includes connection group, short-circuit impedance percentage, etc.; it also includes a preset test item sequence, such as no-load test, load test, power frequency withstand voltage test, induced withstand voltage test, partial discharge test, temperature rise test, etc.; each test item corresponds to a set of standard threshold vectors. For example, the standard value of no-load loss Standard value of load loss Impedance voltage standard value Standard value of partial discharge These standard threshold vectors will serve as the benchmark for subsequent health diagnosis and compliance determination. The test schedule also includes operator information, test date, test standard such as GB 1094 or IEC 60076, and special test requirements such as whether specific load conditions or environmental conditions need to be simulated.

[0078] Meanwhile, the system collects current environmental parameters in real time through environmental sensors deployed at the test station. These sensors include temperature, humidity, and air pressure sensors. The temperature sensor uses a PT100 platinum resistance thermometer or a digital temperature sensing chip, the humidity sensor uses a capacitive humidity sensor, and the air pressure sensor uses a MEMS piezoresistive sensor. These sensors are connected to the main control console via an RS485 bus or a wireless ZigBee network, with a sampling frequency set to 1Hz, meaning one sample is collected every second. The collected environmental parameters include ambient temperature. The unit is Celsius, and the measurement range covers the common temperature range in industrial environments; relative humidity. Atmospheric pressure, expressed as a percentage. The unit is hectopascals. The environmental parameters were collected at the same time as the transformer's entry into the work station to ensure that the environmental data used in subsequent model inference corresponded to the test start time.

[0079] The system integrates the unique identification code, the set of experimental task parameters parsed from the task book, and the environmental parameters collected in real time to form the initial input observation data. The initial input observation data is mathematically represented as a tuple. ,in This is the transformer's unique identification code. This is a set of test parameters, including the transformer's rated parameters, a preset test item sequence, and a standard threshold vector. The environmental parameter triplet collected at the initial moment These data will serve as the first frame input to the physical information neural random field global model, used to initialize the model's state and generate the prior information needed for the first inference.

[0080] To ensure data integrity and reliability, the system immediately calculates the hash value of the initial input observation data after it is generated. The hash value, where timestamp is the current UNIX timestamp accurate to milliseconds, is broadcast to blockchain nodes via the trusted execution environment network for preservation, serving as immutable evidence of the start of the experiment's lifecycle. Simultaneously, the initial input observation data is encrypted and stored in the cache of the local edge computing module, and synchronously backed up to the historical database of the central collaborative computing cluster for subsequent querying and analysis. The encryption process employs the AES-256-GCM algorithm, with the key generated and managed by the trusted execution environment of the edge computing module, ensuring that the data content cannot be deciphered even if the storage medium is illegally accessed.

[0081] The preparation of initial input observation data lays the data foundation for subsequent model inference and experimental execution. The standard threshold vector in the experimental task specification will be compared with the measured values ​​in the subsequent health diagnosis steps. Environmental parameters will serve as boundary condition inputs for the physical field simulation. A unique identification code will be used throughout the entire experimental process to ensure that all subsequently generated data and decisions can be associated with a specific transformer. Through this step, the system establishes a one-to-one mapping relationship between the physical transformer entity and the model representation in the digital space, providing a data entry point and identity anchor for subsequent holographic intelligent experiments.

[0082] S3: Input the initial input observation data into the physical information neural random field global model, obtain the output variables through model inference, and perform corresponding experimental operations based on the output variables.

[0083] In this step, the system uses the initial input observation data obtained in step S2. Input is given to the physical information neural random field global model constructed in step S1. The model receives... Then, the original observation data is first mapped to the initial states of the latent variables through an encoder network. The encoder network adopts a multilayer perceptron structure, with the input layer dimension being [missing information]. The feature dimensions are consistent, and after three hidden layers for dimensionality reduction, the first latent variable is finally output. Initial embeddings and second latent variables The initial field distribution. First latent variable. The initial state of the experimental resources is represented by graph node embedding vectors, with each resource node corresponding to a fixed-dimensional vector, such as 64-dimensional. Second latent variable. The initial distribution of the physical field of the transformer is represented in tensor form. Its spatial dimension depends on the meshing accuracy of the three-dimensional transformer model. For example, it can be divided into a 32×32×32 mesh. Each mesh point corresponds to multiple physical quantities of electromagnetic field, temperature field and stress field, forming a multi-channel three-dimensional tensor.

[0084] After obtaining the initial state of the latent variables, the model solves for the optimal output variable under the current conditions using energy minimization or probabilistic inference algorithms. The inference process employs iterative optimization methods, such as gradient descent or Hamiltonian Monte Carlo sampling, to progressively adjust the latent variables. and The value of makes the energy function The process reaches a local minimum. After iterative convergence, the output variable is obtained by decoding the latent variables corresponding to the minimum energy state. The output variable contains four core components: scheduling decision Safety risk index Health Index and remaining lifetime distribution .

[0085] In this step, the scheduling decision is generated using a Bayesian optimization method, specifically including: defining a collection function that balances expected return and prediction uncertainty, wherein the expected return and prediction uncertainty originate from the output of the physical information neural random field model; and selecting the optimal action as the scheduling decision by maximizing the collection function.

[0086] Bayesian optimization is a global optimization method, particularly suitable for scenarios where the objective function evaluation is costly, noisy, and lacks gradient information. This aligns well with the complex resource allocation and physical process simulation involved in each scheduling decision during transformer testing. In this embodiment, the objective of Bayesian optimization is to optimize the experimental resource state space. and action space The optimal scheduling action is found to maximize the cumulative trial reward. State space. The current state information is composed of the output of the physical information neural random field model, including the first latent variable. Encoded resource state, second latent variable Encoded physical field states and real-time observation data Action space It includes all feasible scheduling actions, such as assigning a mobile test vehicle to a test task, adjusting power supply capacity, switching current transformer taps, switching compensation capacitor banks, etc. Each action is represented as a discrete or continuous parameter vector.

[0087] The core of Bayesian optimization is to construct a probabilistic surrogate model to approximate an unknown objective function. The objective function represents the state. Take action below The cumulative benefit that can be obtained. In this embodiment, the physical information neural random field model itself acts as this probabilistic proxy model because it can output the expected benefit under a given state and action. and the corresponding prediction standard deviation Expected returns The scheduling decision component originates from the model's output variables. Specifically, the model couples the scheduling decision decoder in the energy term. This maps the latent states of a graph structure to the expected reward value of a scheduling action. This reward value is a scalar indicator that comprehensively considers factors such as task completion time, energy consumption, and equipment wear and tear. (Prediction standard deviation) This comes from the model's uncertainty estimate of its own predictions, obtained through stochastic forward propagation of a Bayesian neural network or multi-model ensemble methods, and reflects the model's confidence in the profit prediction under the current state and action.

[0088] Define the acquisition function based on expected return and predicted standard deviation.

[0089]

[0090] in This is the exploration coefficient, used to balance exploration and utilization. When... When the value is large, the acquisition function tends to select actions with high prediction uncertainty, i.e., exploring unknown areas; when When the value is small, the acquisition function tends to select actions with high expected returns, i.e., utilizing known advantageous regions. (Exploration coefficient) It can be dynamically adjusted according to the experimental stage, for example, setting it in the early stage of the experiment. To encourage exploration, the level was gradually reduced in the later stages of the experiment. The focus is on utilization. The higher the value of the acquisition function, the greater the overall value of the action in the current state.

[0091] The system selects the optimal action by maximizing the acquisition function.

[0092]

[0093] This serves as the scheduling decision for the current moment. Since the action space can be large and continuous, the maximization process employs efficient global optimization algorithms, such as L-BFGS-B or CMA-ES, to search for the optimal action within a finite number of iterations. The search process can incorporate domain knowledge for pruning, such as eliminating obviously infeasible action combinations, to improve search efficiency. The optimal action is then obtained. The system then parses these instructions into specific scheduling commands and sends them to each execution unit, including the mobile test vehicle's path planning, the test instrument's parameter settings, and the power supply's start / stop control.

[0094] After executing the scheduling action, the system observes the actual benefit value. For example, the actual completion time and energy consumption of the task are recorded, and these observations are added to the training set as new data points for subsequent updates to the physical information neural random field model. This Bayesian optimization framework enables scheduling decisions to continuously improve through the accumulation of experience, while maintaining the ability to explore unknown situations. It avoids the rigidity of traditional rule-based scheduling methods when facing new situations and the shortcomings of reinforcement learning methods that require a large number of trial and error samples.

[0095] In this step, the method for generating the health index and remaining lifespan distribution includes: constructing a stochastic differential equation to describe the health degradation trajectory based on the posterior distribution of the health index output by the physical information neural random field model, wherein the stochastic differential equation includes a drift coefficient and a diffusion coefficient; solving the first arrival time of the stochastic differential equation to obtain the probability density distribution of the remaining lifespan, wherein the failure threshold is used to define the critical value of health index failure.

[0096] Health Index As one of the output variables of the physical information neural random field model, it originates from the model's input to the second latent variable. Decoding. Specifically, the model decodes by coupling a health index evaluator into the energy term. This maps the physical field state tensor to a health index scalar. Health Index The values ​​are normalized to between 0 and 1, where 1 represents a brand new, intact state and 0 represents a completely failed state. Due to the Bayesian property of the physical information neural random field model, its output health index is not a single point estimate, but rather a posterior probability distribution, meaning that for a given input observation data... The model outputs a health index. probability density function This posterior distribution reflects the uncertainty in health index estimation caused by factors such as measurement noise and model uncertainty.

[0097] Based on the posterior distribution of the health index at the current moment, the system constructs a stochastic differential equation to describe the degradation trajectory of the health index over time. The general form of the Itō stochastic differential equation is adopted:

[0098]

[0099] in For health index stochastic process, The drift coefficient function characterizes the deterministic trend of health index degradation over time; The diffusion coefficient function characterizes the effect of random fluctuations or external disturbances on the degradation process. The increment of the Wiener process represents the standard Brownian motion. Drift coefficient. and diffusion coefficient The functional form can be learned from the physical information neural random field model. Specifically, the model parameters are estimated using historical health index sequence data through maximum likelihood estimation or Bayesian inference methods. In this embodiment, a simplified form of linear drift and power-law diffusion is adopted, namely...

[0100]

[0101] in For degradation rate parameters, For fluctuation amplitude parameters, These are non-linear exponents, and the parameters are learned from historical data and stored securely through a trusted execution environment.

[0102] The initial condition of the stochastic differential equation is the posterior distribution of the health index at the current time, i.e. ,in This represents all accumulated observation data up to the current moment. By solving this stochastic differential equation, the probability distribution of the health index at any future time can be obtained. Specifically, a failure threshold is defined. For example, according to industry standards This represents the critical value at which a transformer requires maintenance or replacement. Remaining lifespan. Defined as a stochastic process First crossing of the failure threshold The time, that is

[0103]

[0104] This is a typical first-pass time problem, and its probability density distribution is... It can be obtained by solving the corresponding Fock-Planck equation or by using Monte Carlo simulation.

[0105] Numerical methods can be used to determine the first-arrival time (FAT). For example, the health index space can be discretized, and the backward Kolmogorov equation can be solved using the finite difference method to obtain the probability density function of FAT. Another more efficient method is to use Monte Carlo simulation. A large number of samples are drawn from the current posterior distribution of the health index, and trajectory simulations are performed for each sample according to stochastic differential equations. The time when each trajectory first crosses the threshold is recorded, and the histogram of these times is used as an approximation of the remaining lifetime probability density distribution. To improve computational efficiency, tens of thousands of simulated trajectories can be executed in parallel on the GPU of the edge computing module, obtaining a high-precision estimate of the remaining lifetime distribution within seconds.

[0106] The obtained remaining lifetime probability density distribution It provides a complete quantification of the uncertainty of the transformer's future operating time, including statistics such as the expected value, median, and confidence interval of the remaining lifetime. For example, the system can output a 90% confidence interval for the remaining lifetime. The range indicates a 90% certainty that the transformer's remaining lifespan falls within this range. This information is more valuable for decision-making than a simple estimate of the remaining lifespan. Maintenance personnel can judge the reliability of the prediction based on the width of the confidence interval; a narrow interval indicates strong prediction certainty, while a wide interval indicates the need for more data or a more cautious decision.

[0107] Scheduling decisions derived from model reasoning Safety risk index Health Index and remaining lifetime distribution The system executes corresponding test operations. Scheduling decisions drive the mobile test vehicle to the designated location and configure test parameters. The safety risk index is displayed in real-time on the main control panel interface, triggering audible and visual alarms when it exceeds a threshold. The health index and remaining lifespan distribution are stored in the database as part of the transformer's health profile, and also used to guide adjustments to the current test strategy. For example, if the health index falls below a preset warning line, the system can automatically reduce the test voltage or shorten the test time to avoid additional damage to the transformer. In this way, model inference results are directly translated into actual test actions, forming a closed-loop control from data to decision-making to execution.

[0108] S4: During the execution of the experimental operation, real-time observation data is continuously collected, and the physical information neural random field global model is updated based on the real-time observation data. The output variables are updated through model inference to achieve dynamic optimization and control of the experimental process.

[0109] In this step of the method, after the scheduling decision generated in step S3 begins execution, the distributed intelligent test unit and the global perception unit continuously collect real-time observation data during the test. This data includes electrical quantity measurements uploaded by the mobile test vehicle, such as voltage, current, power, DC resistance, and turns ratio, with a sampling frequency ranging from 10 to 1000 times per second depending on the test item. It also includes environmental parameters, infrared sensor status, video monitoring frames, equipment vibration, and partial discharge signals collected by the global perception unit. Environmental parameters are updated once per second, two frames of video are collected per second for edge AI analysis, and vibration and partial discharge signals are collected at a high sampling rate of 10,000 points per second. All real-time observation data undergoes preliminary preprocessing within the edge computing module, including denoising, normalization, and feature extraction, forming a new observation data tensor. The subscript 't' represents the time index. These new observations are temporarily stored in the cache of the edge computing module, awaiting use for model updates.

[0110] To learn from the experience accumulated across multiple test sites while protecting data privacy at each site, this step employs a federated secure multi-party computation protocol to update the global model of the Physical Information Neural Random Field. The core idea of ​​this protocol is that each test site computes the model gradient locally, then splits the gradient into multiple parts using a secret sharing technique. These parts are then securely aggregated across trusted execution environments at multiple sites, and finally, a central cluster reconstructs the global gradient and updates the model parameters. Throughout this process, no single node can access the original gradient information from other sites, and the central cluster cannot infer data from individual sites from the aggregation results.

[0111] First, each experimental station calculates the model gradient based on its locally accumulated real-time observation data. Let the observation dataset stored locally at the k-th experimental station be... ,in This refers to the actual output values ​​corresponding to the observed data, such as the actual measured safety event labels or true values ​​of health indices. The model loss function is defined as the weighted sum of the data fitting loss and the physical constraint loss, i.e.

[0112]

[0113] in Indicates the current global model parameters. This is the mean squared error or cross-entropy loss. For the physical energy term defined in S1, This is the balancing coefficient. Within the trusted execution environment, the edge computing module uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters. The gradient is a tensor with the same dimension as the model parameters.

[0114] To protect gradient privacy, the k-th test station performs gradient... Implement (t, n) threshold secret sharing. Specifically, adopt the Shamir secret sharing scheme: select a random polynomial of degree t-1.

[0115]

[0116] The constant term The remaining coefficients Randomly select from a finite field. For each of the n experimental stations (including itself) participating in federated learning, calculate the share corresponding to each station. Here, j is the station index, j=1,...,n. Thus, the gradient is split into n parts; any t parts can reconstruct the original gradient, while fewer than t parts yield no information. The k-th station will receive a share... The encrypted data is sent to the trusted execution environment at station j via a trusted execution environment network. Since the share is a point of random multinomial, even if an attacker intercepts a portion of the share, the original gradient cannot be recovered unless more than t shares are collected.

[0117] Each test station j receives a share from other stations. After (k=1,...,n), secure aggregation is performed within its local trusted execution environment. Because secret sharing exhibits additive homomorphism—that is, the share of the sum of multiple secrets equals the sum of their individual shares—site j adds up all received shares to obtain the aggregated share.

[0118]

[0119] This aggregation process is performed in encrypted form. Station j is unaware of the original gradient corresponding to each share, but the resulting aggregated share is actually the share of the sum of all gradients. Each station will aggregate shares. It is sent along with its own share to the trusted execution environment of the central cluster.

[0120] After the central cluster collects the aggregated shares from at least t different sites, it reconstructs the sum of all gradients using Lagrange interpolation. Specifically, for t shares Construct the Lagrange basis functions

[0121]

[0122] The reconstruction result is

[0123]

[0124] This result is the sum of the gradients of all participating sites. The central cluster further calculates the average gradient.

[0125]

[0126] Then, the global model parameters are updated using stochastic gradient descent.

[0127]

[0128] in The learning rate can be dynamically adjusted based on convergence. The entire gradient reconstruction and parameter update process is completed in a trusted execution environment of the central cluster, ensuring the confidentiality and integrity of parameter updates.

[0129] Updated global model parameters The encrypted broadcast is sent to the trusted execution environment of all test stations. Upon receiving the broadcast, each station decrypts it using its private key and loads the new model, replacing the old model cached locally. This new model is then used for subsequent real-time inference and the next round of local gradient calculation. This completes one round of federated learning updates.

[0130] Through this federated security multi-party computation protocol, multiple test stations can collaboratively optimize the global model without disclosing their individual original data. The introduction of a trusted execution environment further ensures the trustworthiness of the computation process. A remote proof mechanism ensures that the execution environment of each participating node has not been tampered with, effectively preventing attacks from malicious nodes. This update method allows the model to continuously learn from newly generated test data, adapting to factors such as equipment aging and environmental changes, achieving dynamic optimization and control of the test process. For example, when a test station finds that the no-load loss of a certain type of transformer is generally high, the model parameters will automatically adjust, making subsequent scheduling decisions more inclined to arrange longer test times or more precise measurements. Simultaneously, the assessments of safety risk indices and health indices will be revised accordingly, thereby improving the adaptability and accuracy of the entire test system.

[0131] S5: Generate verifiable computational proofs for critical states within the local trusted execution environment, and submit the verifiable computational proofs and the hash values ​​of the critical states to the distributed ledger for storage.

[0132] In this step of the method, critical states refer to core data with significant evidentiary value throughout the experiment, including but not limited to the initial input observation data and its hash value generated in step S2, each scheduling decision output in step S3 and its corresponding acquisition function value, event records where the safety risk index exceeds a preset threshold, the health index and its key quantiles of remaining lifespan distribution, the global model parameter hash values ​​before and after each model update in step S4, and the gradient aggregation proof in the federated aggregation process. Specifically, each generated scheduling decision... At that time, the system records the timestamp of the decision and the corresponding status. Collect function values and the actual benefits after implementation Each safety risk index Exceeding the warning threshold or danger threshold At that time, the system records a complete perception data frame at the moment the event occurs. and risk index values; after each health diagnosis, let the mean of the posterior distribution of the health index be . The standard deviation is Record health index and its 90% confidence interval The remaining lifetime distribution's 10th, 50th, and 90th quantiles are also recorded. After each model update cycle, the hash value of the global model before and after the update, as well as the number of sites participating in this federated aggregation and the proof of the gradient share contributed by each site, are recorded.

[0133] When a critical state occurs, the system immediately generates a verifiable computational proof for it within the local trusted execution environment that generated the state. The verifiable computational proof is generated using zk-SNARK zero-knowledge proof technology, and the content of the proof is: given input data... and output results There is a calculation process. and a set of model parameters , making Established, and and It is a correct program that executes within the current trusted execution environment. The mathematical representation of the proof is as follows:

[0134]

[0135] in The verification key is generated during system initialization and shared among all trusted execution environments; The evidence includes hash values ​​of intermediate variables and model parameters during computation. The proof generation process is completed within a trusted execution environment enclave, utilizing GPU-accelerated elliptic curve computation, with generation time controlled within 500 milliseconds, and the proof size fixed at 192 bytes. This proof can demonstrate the correctness of the computation results to any verifier without exposing the specific input data and model parameters.

[0136] Simultaneously, the system calculates the hash value of critical states. The hash algorithm uses either SHA256 or SM3 (Chinese national cryptographic algorithm), serializing the complete data of the critical states into a byte stream before calculating the hash digest. For scheduling decisions, the hash input includes a timestamp, the serialized value of the state vector, the serialized value of the action vector, and the byte representation of the floating-point value of the acquisition function. For security events, the hash input includes the event type code, timestamp, risk index value, and abbreviated features of the key frames of the perceived data. For health diagnosis results, the hash input includes a timestamp, the mean of the health index and its upper and lower confidence intervals, and the quantiles of remaining lifespan. The hash value calculation is also completed within a trusted execution environment, ensuring that the entire process from raw data to hash digest remains tamper-proof.

[0137] After generating verifiable computational proofs and hash values, the system packages them into a notarized transaction and submits it to the distributed ledger via a trusted execution environment network. The data structure of the notarized transaction is as follows:

[0138]

[0139] Where timestamp is the UNIX timestamp of the transaction generation time, accurate to milliseconds; nodeID is the unique identifier of the test station that generated this critical state; dataType is the critical state type code, for example, 0x01 represents scheduling decision, 0x02 represents security event, 0x03 represents health diagnosis, and 0x04 represents model update; hash is the hash value of the critical state. It is a byte stream that provides verifiable computational proof; the signature is a digital signature generated by the trusted execution environment using the private key for the entire transaction, used for authenticating the transaction's origin.

[0140] In this step, the distributed ledger employs a consensus mechanism based on a directed acyclic graph (DAG), and smart contracts manage the issuance and transfer of experimental resource tokens. The distributed ledger uses a consensus protocol based on a DAG structure, such as IOTA's coordinator scheme or Avalanche's DAG consensus. Each transaction, upon submission, needs to verify and reference two previous transactions, forming a network of reference relationships between transactions. Transaction confirmation time is independent of the overall network transaction load and can be stably kept within 1 second. Ledger nodes are deployed on the blockchain server of a central collaborative computing cluster, initially configured with 7 nodes forming a consortium blockchain, with the number of nodes dynamically increased as needed. Each node maintains a complete copy of the DAG ledger and synchronizes new transactions via a gossip protocol. The consensus mechanism uses a variant based on proof-of-stake, where each node receives a corresponding voting weight based on the number of tokens it holds. A transaction is considered final after being confirmed by more than two-thirds of the nodes with that weight.

[0141] The smart contract is deployed at the execution layer of the DAG ledger, written in Rust and compiled into WebAssembly bytecode, running within the trusted execution environment of the blockchain nodes. The main functions of the smart contract include managing the issuance and transfer of test resource tokens, maintaining the token account balances of each participant, and executing the automatic triggering logic for token transfers. The token is named Test Resource Token (TRT), and adopts a fixed-supply model with an initial total issuance of 100 million tokens, which cannot be increased. Each TRT represents one unit of test resource, specifically: 1 TRT is equivalent to 1 kilowatt-hour of electricity consumption, 10 minutes of mobile test vehicle man-hours, or one standard test project execution right. The smart contract maintains an account mapping table. It records the number of TRTs held by each address, which is generated by hashing the public key of the trusted execution environment of the participating nodes.

[0142] The token is automatically transferred based on actual resource consumption during the experiment. Upon completion of a scheduling decision, the distributed intelligent experimental unit submits a resource consumption proof to the smart contract; the proof format is as follows:

[0143]

[0144] The taskID is the unique identifier for this experiment task, resourceType is the resource type code, amount is the consumption quantity (accurate to two decimal places), executorID is the resource provider address, requesterID is the task initiator address, and signature is the signature of the executor's trusted execution environment on the proof. The smart contract verifies the signature validity and checks whether the consumption quantity is within a reasonable range (e.g., the deviation from the standard consumption quantity corresponding to the task type does not exceed 10). After successful verification, the account update is executed. , The entire transfer process is completed automatically without human intervention.

[0145] The oracle feeds the token price back to the acquisition function. The oracle uses a decentralized oracle network, such as Chainlink's decentralized oracle node cluster, deployed on an independent server within a central collaborative computing cluster. Each oracle node periodically obtains real-time trading prices of TRT and other benchmark tokens (such as USDT) from decentralized exchanges or on-chain liquidity pools based on order books. The oracle nodes aggregate the collected prices, remove outliers, calculate the median price, and submit the aggregation result and aggregation proof to the smart contract. The smart contract updates the price storage variable. This variable is stored in fixed-point format, retaining 4 decimal places of precision. The price update frequency is set to once per minute, or to trigger an immediate update when the price fluctuates by more than 5.

[0146] Bayesian optimization of the acquisition function in step S3 Exploration coefficient in It is no longer a fixed value, but is dynamically adjusted based on the token price. The specific adjustment function is as follows:

[0147]

[0148] in The basic exploration coefficient is preset to 1.0; This is the price sensitivity coefficient, preset to 0.5; The reference price can be set to the average price over the past 24 hours or the initial offering price. When the token price rises, it indicates that experimental resources are becoming scarce, the exploration coefficient increases, and scheduling decisions tend to explore new resource combinations to find more efficient utilization methods. When the token price falls, it indicates that resources are abundant, the exploration coefficient decreases, and scheduling decisions focus more on utilizing known efficient solutions. The token price is fed back to the edge computing module in real time via an oracle. The edge computing module obtains the latest price and updates the collection function before each decision, enabling scheduling decisions to respond in real time to changes in resource supply and demand, achieving a deep integration of market mechanisms and algorithmic decision-making.

[0149] Through this complete trusted evidence storage and token economic mechanism, the system not only ensures the immutability and auditability of key test data, but also builds a self-organized resource scheduling market, enabling test resources to be efficiently allocated under the guidance of economic incentives. This provides a trusted foundation and an intrinsic driving force for continuous optimization for the transformer holographic intelligent test system.

[0150] In summary, the embodiments of this application construct a global model of physical information neural random fields, unifying scheduling decisions, security assessments, and health diagnoses within the same probabilistic framework to achieve deep fusion and collaborative optimization of multi-source information; through a federated secure multi-party computation protocol, continuous model updates are achieved while protecting the data privacy of each test station; through a trusted execution environment and blockchain technology, trusted storage of critical states and the orderly operation of the token economy are ensured; and through the dynamic coupling of Bayesian optimization and token pricing, real-time response of resource scheduling to market supply and demand is achieved. The methods and systems provided by the embodiments of this application can significantly improve the intelligence level, data reliability, and resource utilization efficiency of transformer testing.

[0151] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0152] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A holographic intelligent testing method for transformers, characterized in that, Includes the following steps: A physical information neural random field global model is constructed. The physical information neural random field global model is a unified probabilistic graphical model, which is used to jointly characterize the state of test resources, transformer physical behavior, safety risks and health degradation trends. The model describes the dependencies between variables through an energy function. The variables include input observation data, a first latent variable representing the interaction graph structure between test resources and tasks, a second latent variable representing the spatiotemporal distribution of the transformer physical field state, and output variables. The output variables include scheduling decisions, safety risk index, health index and remaining lifetime distribution. Obtain the unique identification code of the transformer under test, retrieve the corresponding test task book from the database based on the unique identification code, and collect the current environmental parameters as the initial input observation data; The initial input observation data is input into the physical information neural random field global model, the output variables are obtained through model inference, and the corresponding experimental operations are performed based on the output variables. During the experimental operation, real-time observation data is continuously collected, and the physical information neural random field global model is updated based on the real-time observation data. The output variables are updated through model inference to achieve dynamic optimization and control of the experimental process.

2. The transformer holographic intelligent testing method according to claim 1, characterized in that, The energy function is composed of a weighted sum of graph energy terms, physical energy terms, and coupling energy terms; The graph energy term is used to model the topology of the experimental resources through a graph neural network, and is used to characterize the interaction relationships between resource nodes. The physical energy term is constrained by partial differential equations embedded in a physical information neural network and is used to characterize the evolution of the physical field of the transformer. The coupled energy term connection diagram structure and physical field are used to output the final scheduling decision, safety risk index, health index and remaining lifetime distribution.

3. The transformer holographic intelligent testing method according to claim 2, characterized in that, The graph energy term is constructed based on a graph neural network structure and includes node potential functions and edge potential functions. The node potential function is used to describe the state characteristics of a single resource node, and the edge potential function is used to describe the interaction relationship between resource nodes. Both the node potential function and the edge potential function are parameterized by the neural network.

4. The transformer holographic intelligent testing method according to claim 2, characterized in that, The physical energy term is constructed based on physical operators, which are used to calculate the residuals of the partial differential equations of the electromagnetic field, temperature field, or stress field of the transformer. Physical laws are embedded into the model by minimizing the residuals.

5. The transformer holographic intelligent testing method according to claim 1, characterized in that, The updated physical information neural random field global model employs a federated secure multi-party computation protocol, specifically including: Each test site calculates the model gradient within its local trusted execution environment and uses secret sharing to split the gradient into multiple parts and distribute them to the trusted execution environments of other sites. The trusted execution environment at each site securely aggregates the received gradient shares and sends the aggregation results to the central cluster. The central cluster reconstructs the global gradient within the trusted execution environment, updates the model parameters, and broadcasts the updated model to the trusted execution environments of each site in an encrypted manner.

6. The transformer holographic intelligent testing method according to claim 1, characterized in that, The scheduling decision is generated using a Bayesian optimization method, specifically including: Define an acquisition function that balances expected return and prediction uncertainty, wherein the expected return and prediction uncertainty are derived from the output of the physical information neural random field model; The optimal action is selected as the scheduling decision by maximizing the acquisition function.

7. The transformer holographic intelligent testing method according to claim 1, characterized in that, The methods for generating the health index and remaining life expectancy distribution include: Based on the posterior distribution of the health index output by the physical information neural random field model, a stochastic differential equation is constructed to describe the trajectory of health degradation. The stochastic differential equation includes a drift coefficient and a diffusion coefficient. Solving the stochastic differential equation for the first arrival time yields the probability density distribution of the remaining lifetime, where the failure threshold is used to define the critical value for health index failure.

8. The transformer holographic intelligent testing method according to claim 1, characterized in that, It also includes the critical state evidence preservation step: Generate verifiable computational proofs for critical states within a local trusted execution environment. These verifiable computational proofs are used to demonstrate that the critical state is correctly calculated by the current model using the input observation data. The verifiable computational proof and the hash value of the key state are submitted to the distributed ledger for storage.

9. The transformer holographic intelligent testing method according to claim 6, characterized in that, Also includes: A verifiable computational proof is generated for the critical state within a local trusted execution environment, and the verifiable computational proof and the hash value of the critical state are submitted to a distributed ledger for storage. The distributed ledger adopts a consensus mechanism based on a directed acyclic graph and manages the issuance and transfer of experimental resource tokens through smart contracts. The token is automatically transferred based on the actual resource consumption during the trial, and the token price is fed back to the acquisition function through an oracle.

10. A transformer holographic intelligent testing system, characterized in that, include: The distributed intelligent test unit includes a main control console, a mobile characteristic test vehicle, and a mobile temperature rise test vehicle. Each unit has a built-in edge computing module that supports a trusted execution environment for local data acquisition, model inference, and local computing for federated learning. The central collaborative computing cluster includes high-performance servers for training global models of physical information neural random fields, federated learning aggregation servers, blockchain nodes, and trusted execution environment verification services. The global sensing unit includes infrared sensors, video surveillance equipment, environmental sensors, equipment status monitoring terminals, and wireless emergency stop devices. The Trusted Execution Environment Network consists of trusted execution environments for all computing nodes, and establishes trusted channels through remote proof. Blockchain and token networks employ a distributed ledger based on a directed acyclic graph to record key state hashes and token transactions. The system is used to perform the transformer holographic intelligent testing method according to any one of claims 1 to 9.

Images