Industrial equipment intelligent training evaluation and on-chain certification traceability system

Abstract

This invention discloses an intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment, relating to the fields of intelligent training for industrial equipment and data asset certification. The system includes: constructing a causal directed graph of the target industrial equipment based on a causal discovery algorithm; optimizing the structural causal model; generating a digital twin model; collecting multimodal behavioral data generated during interactions; generating a user experience level profile; calculating adaptive guidance strength based on the user experience level profile and a preset threshold; outputting virtual prompts matching the operator's experience level; packaging the operational data and multimodal behavioral data into a data package to be uploaded to the blockchain; performing zero-knowledge proof calculations on the data package to be uploaded, generating a zero-knowledge proof package, and uploading it to the blockchain network for storage, thus generating an on-chain certification record. This invention achieves the integration of digital twin causal consistency, AI-driven personalized guidance, and low-overhead, high-privacy blockchain certification of data assets in industrial equipment training and evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent training and data asset authentication technology for industrial equipment, and in particular to an intelligent training, evaluation and on-chain authentication and traceability system for industrial equipment. Background Technology

[0002] In the operation and maintenance training of high-value, high-risk industrial equipment in power, chemical, and metallurgical industries, traditional methods mainly rely on video operation demonstrations and VR model displays. However, these methods can only present fixed scenarios and cannot realize real interactive operations between employees and equipment failure scenarios, resulting in the common problem of "easy to understand at first glance, but confusing when actually using it." In addition, equipment failure troubleshooting experience is mainly passed down through "experienced mentors," which is not only inefficient but also fails to cover all failure types, making it difficult to systematically retain experience. Traditional operation and maintenance logs only record photos and text, which cannot fully reproduce the operation process and troubleshooting ideas, resulting in poor experience transfer. For employees' creative contributions to solving failures, companies often only provide one-time rewards, making it impossible to quantify their value throughout the equipment's entire life cycle and to establish a fair competency certification system.

[0003] In the application of digital twin technology, existing digital twin models lack causal consistency with real devices, failing to accurately reflect the true operational logic and fault evolution process of the devices. Meanwhile, traditional intelligent assistants use the same model output for all users, failing to consider the differences in operating habits among users with varying experience levels, resulting in insufficient guidance for novice users and redundant guidance for expert users. Regarding data assetization, existing solutions generally suffer from high on-chain overhead and insufficient privacy protection when storing device operation data on the blockchain, making it difficult to achieve efficient value exchange and traceability while ensuring data security. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment. It aims to solve the problems of causal inconsistency between digital twin models and physical equipment in existing industrial equipment training, evaluation, and maintenance assistance systems; the inability of intelligent assistants to provide personalized guidance based on users' experience levels; and the high overhead and insufficient privacy protection when storing equipment operation data on the blockchain.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment, comprising the following steps: Model training module: used to acquire the physical parameters and historical operating data of the target industrial equipment, construct a causal directed graph of the target industrial equipment based on the causal discovery algorithm, optimize the structural causal model through joint training loss function, and generate a digital twin model; Data acquisition module: used to collect multimodal behavioral data generated when operators interact with the digital twin model; The profile generation module is used to extract the skill level vector, operation speed features, error pattern features and interaction preference features of operators based on multimodal behavioral data, and to generate a user experience level profile of the operator through feature fusion via a multilayer perceptron network. Strategy guidance module: used to calculate adaptive guidance strength based on user experience level profiles and preset experience level thresholds through logical functions; Information prompt module: It is used to output virtual prompt information that matches the operator's current experience level based on the adaptive guidance strength, and to collect the operation data performed by the operator based on the virtual prompt information; Data upload module: Used to package the operation data performed by the operator and the multimodal behavior data to generate a data package to be uploaded to the blockchain; The authentication record module is used to perform zero-knowledge proof calculations on the data packets to be uploaded to the blockchain, generate zero-knowledge proof packets, upload the zero-knowledge proof packets to the blockchain network for storage, and generate on-chain authentication records.

[0006] Furthermore, the process of acquiring the physical parameters and historical operating data of the target industrial equipment, constructing a causal directed graph of the target industrial equipment based on a causal discovery algorithm, optimizing the structural causal model through joint training of the loss function, and generating a digital twin model includes: Collect the physical parameters and historical operating data of the target industrial equipment, perform time alignment and normalization on the physical parameters and historical operating data, and construct an observation dataset; Based on the observation dataset, a scoring function and a sparse regularization term are used to learn the causal structure. The optimal causal structure is found by minimizing the weighted sum of the data fitting score and the number of edges in the causal directed graph. The optimal causal directed graph is generated using the following formula: ; In the formula, To form an optimal causal directed graph, For the observation dataset, Score the data fit. The regularization coefficient is . Let be the number of edges in the causal directed graph; A joint training loss function is constructed, which includes a pre-defined spatiotemporal graph neural network prediction loss, causal discovery loss, structural causal model fitting loss, and sparse regularization loss. The joint training loss function is defined by the following formula: ; In the formula, For the spatiotemporal graph neural network prediction loss, To discover loss in causal discovery, For the fitting loss of the structural causal model, For sparsity regularization loss, to Weights for each loss item; Using the optimal causal directed graph as the initial graph structure, and aiming to minimize the joint training loss function, the parameters of the spatiotemporal graph neural network and the structural causal model are iteratively updated until the joint training loss function converges. The converged structural causal model is then used as a digital twin model.

[0007] Furthermore, the multimodal behavioral data generated when operators interact with the digital twin model includes: The device collects data on the operator's hand skeletal movement trajectory, eye focus, and voice control commands using mixed reality. Record the response time, operation type, operation parameters, and operation completion flag for each interaction between the operator and the virtual control panel in the digital twin model. Arrange the response time, operation type, operation parameters, and operation completion flag in the order of operation execution to generate an operation sequence. The hand skeletal movement trajectory, eye gaze focus data, voice control commands, and operation sequences are combined into multimodal behavioral data.

[0008] Furthermore, the step of extracting the operator's skill level vector, operation speed features, error pattern features, and interaction preference features based on multimodal behavioral data, and then fusing these features through a multilayer perceptron network to generate a user experience level profile of the operator, includes: The trajectory smoothness of hand skeletal motion trajectories and the proportion of successful operation completion in the operation sequence are extracted from multimodal behavioral data. The trajectory smoothness and the success ratio are concatenated into a two-dimensional vector as a skill level vector. Extract the arithmetic mean of the reciprocals of the response times of each interaction operation in the operation sequence from the multimodal behavioral data, and use it as the operation speed feature; The number of times an operation completion marker is marked as a failure is counted in the operation sequence from multimodal behavioral data, and the operation type distribution corresponding to the number of failures is extracted. The number of failures and the operation type distribution are encoded as error pattern features. The frequency of voice control commands and the percentage of time spent on each area of ​​the virtual control panel are extracted from multimodal behavioral data. The frequency of use and the percentage of time spent on each area are then combined to form interaction preference features. The skill level vector, operation speed feature, error pattern feature, and interaction preference feature are concatenated to generate a fused feature vector. This fused feature vector is then input into a multilayer perceptron network, and the user experience level profile is calculated using the following formula: ; In the formula, To profile user experience levels, MLP stands for Multilayer Perceptron Network. For skill level vectors, For operating speed characteristics, For error mode characteristics, For interaction preference features, This is a vector concatenation operation.

[0009] Furthermore, the step of calculating the adaptive guidance strength based on the user's experience level profile and a preset experience level threshold using a logical function includes: Read the user's experience level profile, obtain the preset experience level threshold, calculate the difference between the user's experience level profile and the experience level threshold, and obtain the experience deviation. Based on the preset adjustment sensitivity parameters and the maximum guidance intensity value, the adaptive guidance intensity is calculated using the following formula: ; In the formula, For adaptive guidance strength, For logical functions, For maximum guiding strength, To adjust sensitivity, The experience level threshold is the threshold when the user's experience level is... When below the threshold, the guiding strength is close to When the value is above the threshold, the guiding strength approaches 0.

[0010] Furthermore, the step of outputting virtual prompts that match the operator's current experience level based on the adaptive guidance strength, and collecting operation data performed by the operator based on the virtual prompts, includes: Obtain adaptive guidance strength, a pre-defined large language model, and a computer vision model; The large language model is used to generate text prompts and voice prompts of different levels of detail based on the adaptive guidance intensity. The computer vision model is used to generate three-dimensional graphic annotation parameters of different complexities based on the adaptive guidance intensity. The text prompts, voice prompts, and three-dimensional graphic annotation parameters are combined to generate virtual prompt information. Display virtual prompts in the mixed reality device, collect the operator's hand movements and eye gaze points in the mixed reality device in real time, generate the operation type, operation parameters and operation completion mark performed by the operator, and obtain the original operation record; The state change sequence is obtained from the digital twin model. The state change sequence includes the device parameter change curve and alarm event log. The original operation record is associated with the state change sequence and stored to obtain operation data.

[0011] Furthermore, the step of packaging the operation data performed by the operator with multimodal behavioral data to generate a data packet to be uploaded to the blockchain includes: According to preset time intervals or preset specific search conditions, all operation data and all multimodal behavior data collected in the current period are aggregated to generate aggregated data records; Add operation timestamps and operator identification to aggregated data records to generate aggregated data records with identification; The tagged aggregated data records are serialized into a structured data format to obtain the data packet to be uploaded to the blockchain.

[0012] Furthermore, the step of performing zero-knowledge proof computation on the data packet to be uploaded to the blockchain to generate a zero-knowledge proof package, uploading the zero-knowledge proof package to the blockchain network for storage, and generating an on-chain authentication record includes: Obtain the data packet to be uploaded to the blockchain, and perform proof calculation on the data packet using a zero-knowledge concise non-interactive knowledge proof algorithm to generate a zero-knowledge proof packet. The proof size, verification time, and gas consumption of the zero-knowledge proof packet satisfy the following formula: ; ; ; In the formula, For the size of the zero-knowledge proof packet, , It is an elliptic curve group. To verify the time, For bilinear pairing operations, For scalar multiplication operations, For the input quantity, The gas consumption on the Ethereum network; the information-theoretic security and privacy leakage probability of this zero-knowledge proof package satisfies the formula: ; In the formula, To reduce the probability of privacy breaches for zero-knowledge proof schemes, The probability of privacy breaches in encryption schemes. This represents the probability that the key will be cracked. The zero-knowledge proof package is uploaded to the blockchain network, triggering a smart contract in the blockchain network to perform a verification operation on the zero-knowledge proof package using a preset verification key. After successful verification, the transaction hash value and block height returned by the blockchain network are obtained. The transaction hash value and block height are combined as key-value pairs to generate an on-chain authentication record.

[0013] The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment provided by this invention has the following main advantages: 1. This invention utilizes causal consistency digital twin modeling to ensure that the causal relationship between the digital twin model and the physical system remains consistent, achieving real-time synchronization between the state changes of virtual devices and real devices. Employees can repeatedly practice various fault scenarios in a virtual environment, improving the speed of equipment fault diagnosis and extending equipment lifespan by approximately 20%. Through a personalized AI-assisted system, the system can dynamically adjust guidance strategies based on the user's operating skills, speed, error patterns, and interaction preferences, providing a personalized experience of detailed guidance for beginners and concise prompts from experts. This accelerates the onboarding speed for new employees, shortens training time by approximately 30%, and increases employee work enthusiasm by approximately 30%.

[0014] 2. This invention employs a low-overhead, highly private blockchain authentication and traceability scheme, utilizing zk-SNARK zero-knowledge proof technology. The proof size is only approximately 192 bytes, and the verification time is constant, ensuring the engineering feasibility of blockchain deployment. Enterprises can transform equipment operation experience, troubleshooting methods, and other knowledge into traceable and verifiable digital assets, increasing the value of their knowledge assets and enhancing the ability to mine the value of equipment data. Through an information-theory-secure privacy protection mechanism, hierarchical protection of data assets is achieved, raising the probability of privacy leakage from computational security to information-theory security, effectively safeguarding the security of enterprise data assets. Attached Figure Description

[0015] Figure 1 This is a functional module diagram of the intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment of the present invention; Figure 2 This is a system architecture diagram of the intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-2 This invention provides an intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment, comprising: Model training module: used to acquire the physical parameters and historical operating data of the target industrial equipment, construct a causal directed graph of the target industrial equipment based on the causal discovery algorithm, optimize the structural causal model through joint training loss function, and generate a digital twin model; In this embodiment, the steps of acquiring the physical parameters and historical operating data of the target industrial equipment, constructing a causal directed graph of the target industrial equipment based on a causal discovery algorithm, optimizing the structural causal model through joint training of the loss function, and generating a digital twin model include: Collect the physical parameters and historical operating data of the target industrial equipment, perform time alignment and normalization on the physical parameters and historical operating data, and construct an observation dataset.

[0018] Specifically, physical parameters refer to the inherent characteristics of the equipment, such as its geometric dimensions, material properties, and rated power, while historical operating data refers to the time-series signals from the sensors recorded by the equipment in the past. Time alignment associates the two types of data at the same sampling time, and normalization maps values ​​with different dimensions to a unified interval. The observation dataset is a structured table after alignment and normalization, used for subsequent causal discovery.

[0019] Based on the observation dataset, a scoring function and a sparse regularization term are used to learn the causal structure. The optimal causal structure is found by minimizing the weighted sum of the data fitting score and the number of edges in the causal directed graph. The optimal causal directed graph is generated using the following formula: ; In the formula, To form an optimal causal directed graph, For the observation dataset, Score the data fit. The regularization coefficient is . Let be the number of edges in the causal directed graph.

[0020] Specifically, the scoring function uses the Bayesian information criterion to measure the degree of matching between the candidate causal graph and the observed data. The sparse regularization term is composed of the number of edges multiplied by a preset regularization coefficient, which is used to penalize overly complex graph structures. The regularization coefficient α is set empirically by the modelers based on the ratio of the variable dimension of the target industrial equipment to the sample size. The principle for its value is: when the sample size is small relative to the number of variables, a larger α is used to enhance sparsity; conversely, a smaller α is used to retain weak causal edges. The specific value can be determined through sensitivity analysis of the pre-experiment.

[0021] A joint training loss function is constructed, which includes a pre-defined spatiotemporal graph neural network prediction loss, causal discovery loss, structural causal model fitting loss, and sparse regularization loss. The joint training loss function is defined by the following formula: ; In the formula, For the prediction loss of the spatiotemporal graph neural network, To discover loss in causal discovery, For the fitting loss of the structural causal model, For sparse regularization loss, to Weights for each loss item; Specifically, the prediction loss of the spatiotemporal graph neural network is predefined as the mean square error between the predicted output and the true value, and its network structure is set according to the temporal characteristics of the device. The causal discovery loss forces the learned graph structure to be consistent with the optimal causal directed graph, and the structural causal model fitting loss constrains the model output to approximate the observed data.

[0022] Sparse regularization loss further limits model complexity. The four weight coefficients are preset by the modelers to balance the contribution of each loss term. During the system construction phase, the observation dataset is divided into training set and validation set according to the proportion. In the preset weight combination candidate space, the model is trained on the training set for each set of candidate weight coefficients, and the comprehensive performance index predicted by the model is calculated using the validation set. The weight combination that makes the comprehensive performance on the validation set is selected as the preset weight coefficient of each loss term. This set of weight coefficients remains fixed in the subsequent training process of the digital twin model and does not change with iterative updates.

[0023] Using the optimal causal directed graph as the initial graph structure, and aiming to minimize the joint training loss function, the parameters of the spatiotemporal graph neural network and the structural causal model are iteratively updated until the joint training loss function converges. The converged structural causal model is then used as a digital twin model.

[0024] Specifically, the initial graph structure uses the directed edges of the optimal causal directed graph as the adjacency matrix of the spatiotemporal graph neural network. Iterative updates employ backpropagation and gradient descent to repeatedly adjust network weights and model parameters. Convergence is determined when the joint training loss function stops decreasing after multiple rounds. At this point, the structural causal model is the digital twin model, used to reproduce the causal logic of real devices in a virtual environment.

[0025] Data acquisition module: used to collect multimodal behavioral data generated when operators interact with the digital twin model; In this embodiment, the method for collecting multimodal behavioral data generated when operators interact with the digital twin model includes: The mixed reality device collects data on the operator's hand skeletal movement trajectory, eye focus, and voice control commands.

[0026] Specifically, mixed reality devices are head-mounted devices that can track user body movements and overlay virtual images. Hand skeletal movement trajectory refers to the sequence of changes in the three-dimensional spatial position of finger joints and palms. Eye gaze focus data refers to the coordinates of the point where the gaze lands in the virtual scene. Voice control commands refer to the semantic content of the recognized speech.

[0027] Record the response time, operation type, operation parameters, and operation completion flag for each interaction between the operator and the virtual control panel in the digital twin model. Arrange the response time, operation type, operation parameters, and operation completion flag in the order of operation execution to generate an operation sequence.

[0028] Specifically, the virtual control panel is a three-dimensional area in the digital twin model that simulates the operating interface of a real device. Response time refers to the time consumed from the start to the completion of the operation. Operation type refers to the type of interactive action such as clicking or rotating. Operation parameters refer to values ​​such as rotation angle or sliding distance associated with the operation type. Operation completion flag refers to the indicator value of success or failure. Arranging the above four pieces of information in chronological order yields the operation sequence.

[0029] The hand skeletal movement trajectory, eye gaze focus data, voice control commands, and operation sequences are combined into multimodal behavioral data.

[0030] Specifically, the combination refers to synchronously associating the above four types of data according to the same operation timestamp and aggregating them into a unified data structure. Hand trajectories reflect the smoothness of operation, eye focus reflects the allocation of attention, voice commands reflect the behavior of actively seeking help, and operation sequences reflect the progress and correctness of task completion. Multimodal behavioral data is the fusion of the above four types of information.

[0031] The profile generation module is used to extract the skill level vector, operation speed features, error pattern features and interaction preference features of operators based on multimodal behavioral data, and to generate a user experience level profile of the operator through feature fusion via a multilayer perceptron network. In this embodiment, the step of extracting the operator's skill level vector, operation speed features, error pattern features, and interaction preference features based on multimodal behavioral data, and then fusing these features through a multilayer perceptron network to generate a user experience level profile of the operator includes: The trajectory smoothness of hand skeletal motion trajectories and the proportion of successful operation completion in the operation sequence are extracted from multimodal behavioral data. The trajectory smoothness and the success proportion are concatenated into a two-dimensional vector as a skill level vector.

[0032] Specifically, trajectory smoothness refers to the reciprocal of the variance of the change in direction angle between adjacent sampling points in the hand skeletal movement trajectory; a larger value indicates smoother movement. Success rate refers to the ratio of the number of successful operations to the total number of operations in the sequence. Combining trajectory smoothness and success rate sequentially into a two-dimensional vector yields the skill level vector.

[0033] Extract the arithmetic mean of the reciprocals of the response times of each interaction operation in the operation sequence from the multimodal behavioral data, and use it as the operation speed feature.

[0034] Specifically, the reciprocal of the response time measures the speed of each operation, and its arithmetic mean is the average of the speeds of all operations. This average is the operation speed characteristic, reflecting how quickly the operator performs interactions.

[0035] The number of times an operation was marked as a failure in the operation sequence was counted from the multimodal behavioral data, and the operation type distribution corresponding to the number of failures was extracted. The number of failures and the operation type distribution were encoded into error pattern features.

[0036] Specifically, the number of failures refers to the cumulative number of operations marked as failed. The operation type distribution refers to the proportion of various actions such as clicks, rotations, and swipes among the failed operations. Combining the number of failures with these proportions into a multi-dimensional vector yields the error pattern features.

[0037] The frequency of voice control commands and the percentage of time spent by eye gaze focus data in different areas of the virtual control panel are extracted from multimodal behavioral data. The frequency of use and the percentage of time spent are then combined to form interaction preference features.

[0038] Specifically, the frequency of voice control commands refers to the ratio of the number of voice commands issued to the total time within the current operation cycle. The dwell time percentage refers to the percentage obtained by dividing the cumulative dwell time of the eye focus in each panel area by the total gaze time. Combining the frequency of use and the percentage of each area in sequence into a vector yields the interaction preference features.

[0039] The skill level vector, operation speed feature, error pattern feature, and interaction preference feature are concatenated to generate a fused feature vector. This fused feature vector is then input into a multilayer perceptron network, and the user experience level profile is calculated using the following formula: ; In the formula, To profile user experience levels, MLP stands for Multilayer Perceptron Network. As a skill level vector, For operating speed characteristics, For error mode characteristics, For interaction preference features, This is a vector concatenation operation.

[0040] Specifically, vector concatenation refers to joining the four feature vectors sequentially end-to-end to form a single vector, i.e., a fused feature vector. The multilayer perceptron network is a pre-defined feedforward neural network. Its structure is determined by the modeler based on experience, consisting of an input layer, a single hidden layer, and an output layer. The input layer dimension equals the dimension of the fused feature vector, and the output layer dimension is 1. After non-linear mapping by the network, the fused feature vector outputs a scalar, which represents the user's experience level profile.

[0041] Strategy guidance module: used to calculate adaptive guidance strength based on user experience level profiles and preset experience level thresholds through logical functions; In this embodiment, the step of calculating the adaptive guidance strength based on the user's experience level profile and a preset experience level threshold using a logical function includes: Read the user's experience level profile, obtain the preset experience level threshold, calculate the difference between the user's experience level profile and the experience level threshold, and obtain the experience deviation.

[0042] Specifically, the experience level threshold is a preset scalar benchmark value, derived from the arithmetic mean of user experience level profiles collected from multiple certified skilled operators. The experience deviation is equal to the user's experience level profile minus the experience level threshold; a positive value indicates that the user's level is higher than the benchmark, and a negative value indicates that it is lower than the benchmark.

[0043] Based on the preset adjustment sensitivity parameters and the maximum guidance intensity value, the adaptive guidance intensity is calculated using the following formula: ; In the formula, For adaptive guidance strength, For logical functions, For maximum guiding strength, To adjust sensitivity, The experience level threshold is the threshold when the user's experience level is... When below the threshold, the guiding strength is close to When the value is above the threshold, the guiding strength approaches 0.

[0044] Specifically, the maximum guidance intensity is a preset maximum output value for prompts, set to 100% by the system designers based on the display capabilities of the mixed reality device. The sensitivity adjustment is a preset positive coefficient, selected through cross-validation to achieve the optimal differentiation of guidance intensity across different experience levels. A logic function maps the negative values ​​of the experience deviation to the zero-to-one range, multiplies it by the maximum guidance intensity to obtain the adaptive guidance intensity, which controls the level of detail in subsequent virtual prompts.

[0045] Information prompt module: It is used to output virtual prompt information that matches the operator's current experience level based on the adaptive guidance strength, and to collect the operation data performed by the operator based on the virtual prompt information; In this embodiment, the step of outputting virtual prompts that match the operator's current experience level based on the adaptive guidance strength, and collecting operation data performed by the operator based on the virtual prompts, includes: Obtain adaptive guidance strength, preset large language model, and computer vision model.

[0046] Specifically, the adaptive guidance strength is output from the previous calculation step, ranging from 0% to 100%. The large language model is a pre-defined generative text model, which is pre-tuned using industrial equipment operation corpora based on a general pre-trained model. The computer vision model is a pre-defined object detection and segmentation network, which is pre-trained on a dataset of hand and equipment images collected in mixed reality.

[0047] The large language model is used to generate text prompts and voice prompts of different levels of detail based on the adaptive guidance intensity. The computer vision model is used to generate three-dimensional graphic annotation parameters of different complexities based on the adaptive guidance intensity. The text prompts, voice prompts, and three-dimensional graphic annotation parameters are combined to generate virtual prompt information.

[0048] Specifically, the large language model uses adaptive guidance strength as a control parameter. High guidance strength outputs step-by-step text and synthesized speech, while low guidance strength outputs keyword prompts. The computer vision model, also based on adaptive guidance strength, outputs dynamic 3D arrows and highlighted borders when the guidance strength is high, and static contour annotations when the guidance strength is low. By aligning the text, speech, and 3D annotations according to the operation sequence and then merging them, the virtual prompt information is obtained.

[0049] Virtual prompts are displayed in the mixed reality device, and the operator's hand movements and eye gaze points are collected in real time to generate the operation type, operation parameters and operation completion mark performed by the operator, thus obtaining the original operation record.

[0050] Specifically, mixed reality devices overlay virtual prompts onto the surface of a real device. Hand movements refer to the pose sequence of the hand's skeletal joints, and eye fixation points refer to the coordinates of the intersection of the line of sight and the virtual plane. Based on the captured data, the operation type is identified, such as click or rotation, and operation parameters such as angle or displacement are extracted. A completion marker is generated to determine whether the operation was successful, and the data is summarized into a raw operation record.

[0051] The state change sequence is obtained from the digital twin model. The state change sequence includes the device parameter change curve and alarm event log. The original operation record is associated with the state change sequence and stored to obtain operation data.

[0052] Specifically, the equipment parameter change curve refers to the time-series array of continuous quantities such as temperature and pressure output by the digital twin model. The alarm event log refers to the discrete marked records of when parameters exceed the threshold. Associated storage refers to merging the original operation records and state change sequences with the same timestamp into a complete operation data entry for subsequent on-chain evidence storage.

[0053] Data upload module: Used to package the operation data performed by the operator and the multimodal behavior data to generate a data package to be uploaded to the blockchain; In this embodiment, according to a preset time interval or a preset specific search condition, all operation data collected in the current period and all multimodal behavior data are aggregated to generate an aggregated data record.

[0054] Specifically, the preset time interval is a fixed duration set by the administrator in the configuration interface, such as performing aggregation every five minutes; preset specific search conditions are data filtering rules defined by the administrator according to business needs, such as triggering aggregation after accumulating ten operation records or detecting a fault event. The current period refers to the time period from the end of the last aggregation to the current moment. Operation data refers to the operation type, operation parameters, and completion markers generated after employees perform tasks according to virtual prompts. Multimodal behavioral data refers to hand trajectories, eye focus, and voice commands collected by mixed reality devices. Aggregation refers to merging the above two types of data into a data unit in chronological order to generate aggregated data records, thereby organizing the scattered original interaction information into compact and orderly data blocks.

[0055] Add operation timestamps and operator identification to aggregated data records to generate identifiable aggregated data records.

[0056] Specifically, the operation timestamp refers to the precise time obtained from the time synchronization server, used to mark the completion time of this aggregated data record. The operator identification refers to the unique digital certificate of the employee logging into the main control system, used to distinguish the contributions of different operators. Adding refers to appending a timestamp and identification to the header or footer of the aggregated data record, forming an identified aggregated data record, thereby providing an immutable credential for subsequent traceability and accountability.

[0057] The tagged aggregated data records are serialized into a structured data format to obtain the data packet to be uploaded to the blockchain.

[0058] Specifically, serialization refers to converting nested data structures in memory into a continuous stream of bytes for easy storage or network transmission. Structured data formats refer to representations with clearly defined fields and encoding rules, such as JSON or Protocol Buffers. Data packets to be uploaded to the blockchain refer to binary data units that, after serialization, can be directly submitted to the blockchain network for evidence storage, ensuring correct parsing and verification across different hardware and software environments.

[0059] The authentication record module is used to perform zero-knowledge proof calculations on the data packets to be uploaded to the blockchain, generate zero-knowledge proof packets, upload the zero-knowledge proof packets to the blockchain network for storage, and generate on-chain authentication records.

[0060] In this embodiment, the step of performing zero-knowledge proof computation on the data packet to be uploaded to the blockchain to generate a zero-knowledge proof packet, uploading the zero-knowledge proof packet to the blockchain network for storage, and generating an on-chain authentication record includes: Obtain the data packet to be uploaded to the blockchain, and perform proof calculation on the data packet using a zero-knowledge concise non-interactive knowledge proof algorithm to generate a zero-knowledge proof packet. The proof size, verification time, and gas consumption of the zero-knowledge proof packet satisfy the following formula: ; ; ; In the formula, For the size of the zero-knowledge proof packet, , It is an elliptic curve group. To verify the time, For bilinear pairing operations, For scalar multiplication operations, For the input quantity, This refers to the amount of gas consumed on the Ethereum network. Specifically, the data packet to be uploaded to the blockchain is a serialized binary data unit. The zero-knowledge concise non-interactive knowledge proof algorithm, namely the zk-SNARK algorithm, can generate verifiable proofs without revealing the original data. Proof size. Approximately 192 bytes, corresponding to three group element and one The system requires a group of elements; the verification time is constant, requiring three bilinear pairing operations and n scalar multiplications; and the fuel consumption is approximately 200,000. The overall effect is to compress the data packets to be uploaded to the blockchain into a very small zero-knowledge proof package, meeting the low-overhead and high-efficiency verification requirements of blockchain notarization.

[0061] The information-theoretic security and privacy leakage probability of this zero-knowledge proof package satisfies the following formula: ; In the formula, To reduce the probability of privacy breaches for zero-knowledge proof schemes, The probability of privacy breaches in encryption schemes. This represents the probability that the key will be cracked. Specifically, information-theoretical security refers to the fact that even with unlimited computing power, an attacker cannot obtain any original data from the proof. The computational security of an encryption scheme depends on key length and algorithm strength; its probability of privacy leakage equals the probability of the key being cracked, which is greater than zero. Zero-knowledge proofs mathematically prevent data leakage, providing a higher level of privacy protection compared to traditional encryption schemes.

[0062] The zero-knowledge proof package is uploaded to the blockchain network, triggering a smart contract in the blockchain network to perform a verification operation on the zero-knowledge proof package using a preset verification key. After successful verification, the transaction hash value and block height returned by the blockchain network are obtained. The transaction hash value and block height are combined as key-value pairs to generate an on-chain authentication record.

[0063] Specifically, the preset verification key is generated by the administrator through trusted settings when deploying the smart contract. It is paired with the proof key used for proof computation to verify the correctness of the proof without exposing the original data. The verification operation refers to the process where, after the smart contract performs bilinear pairing and other operations to confirm the validity of the proof, the blockchain network returns the transaction hash value and block height as an immutable proof of existence. The overall effect is to anchor the zero-knowledge proof package on the blockchain, forming a key-value pair combination of transaction hash value and block height, i.e., an on-chain authentication record.

[0064] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0065] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment, characterized in that: include: Model training module: used to acquire the physical parameters and historical operating data of the target industrial equipment, construct a causal directed graph of the target industrial equipment based on the causal discovery algorithm, optimize the structural causal model through joint training loss function, and generate a digital twin model; Data acquisition module: used to collect multimodal behavioral data generated when operators interact with the digital twin model; The profile generation module is used to extract the skill level vector, operation speed features, error pattern features and interaction preference features of operators based on multimodal behavioral data, and to generate a user experience level profile of the operator through feature fusion via a multilayer perceptron network. Strategy guidance module: used to calculate adaptive guidance strength based on user experience level profiles and preset experience level thresholds through logical functions; Information prompt module: It is used to output virtual prompt information that matches the operator's current experience level based on the adaptive guidance strength, and to collect the operation data performed by the operator based on the virtual prompt information; Data upload module: Used to package the operation data performed by the operator and the multimodal behavior data to generate a data package to be uploaded to the blockchain; The authentication record module is used to perform zero-knowledge proof calculations on the data packets to be uploaded to the blockchain, generate zero-knowledge proof packets, upload the zero-knowledge proof packets to the blockchain network for storage, and generate on-chain authentication records.

2. The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment according to claim 1, characterized in that, The process of acquiring the physical parameters and historical operating data of the target industrial equipment, constructing a causal directed graph of the target industrial equipment based on a causal discovery algorithm, optimizing the structural causal model through joint training of loss functions, and generating a digital twin model includes: Collect the physical parameters and historical operating data of the target industrial equipment, perform time alignment and normalization on the physical parameters and historical operating data, and construct an observation dataset; Based on the observation dataset, a scoring function and a sparse regularization term are used to learn the causal structure. The optimal causal structure is found by minimizing the weighted sum of the data fitting score and the number of edges in the causal directed graph. The optimal causal directed graph is generated using the following formula: ； In the formula, To form an optimal causal directed graph, For the observation dataset, Score the data fit. The regularization coefficient is . Let be the number of edges in the causal directed graph; A joint training loss function is constructed, which includes a pre-defined spatiotemporal graph neural network prediction loss, causal discovery loss, structural causal model fitting loss, and sparse regularization loss. The joint training loss function is defined by the following formula: ； In the formula, For the spatiotemporal graph neural network prediction loss, To discover loss in causal discovery, For the fitting loss of the structural causal model, For sparse regularization loss, to Weights for each loss item; Using the optimal causal directed graph as the initial graph structure, and aiming to minimize the joint training loss function, the parameters of the spatiotemporal graph neural network and the structural causal model are iteratively updated until the joint training loss function converges. The converged structural causal model is then used as a digital twin model.

3. The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment according to claim 1, characterized in that, The multimodal behavioral data generated when operators interact with the digital twin model includes: The device uses mixed reality to collect data on the operator's hand skeletal movement trajectory, eye focus, and voice control commands. Record the response time, operation type, operation parameters, and operation completion flag for each interaction between the operator and the virtual control panel in the digital twin model. Arrange the response time, operation type, operation parameters, and operation completion flag in the order of operation execution to generate an operation sequence. The hand skeletal movement trajectory, eye gaze focus data, voice control commands, and operation sequences are combined into multimodal behavioral data.

4. The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment according to claim 1, characterized in that, The process involves extracting skill level vectors, operation speed features, error pattern features, and interaction preference features from multimodal behavioral data, fusing these features using a multilayer perceptron network, and generating a user experience level profile of the operator, including: The trajectory smoothness of hand skeletal motion trajectories and the proportion of successful operation completion in the operation sequence are extracted from multimodal behavioral data. The trajectory smoothness and the success ratio are concatenated into a two-dimensional vector as a skill level vector. Extract the arithmetic mean of the reciprocals of the response times of each interaction operation in the operation sequence from the multimodal behavioral data, and use it as the operation speed feature; The number of times an operation completion marker is marked as a failure is counted in the operation sequence from multimodal behavioral data, and the operation type distribution corresponding to the number of failures is extracted. The number of failures and the operation type distribution are encoded as error pattern features. The frequency of voice control commands and the percentage of time spent on each area of ​​the virtual control panel are extracted from multimodal behavioral data. The frequency of use and the percentage of time spent on each area are then combined to form interaction preference features. The skill level vector, operation speed feature, error pattern feature, and interaction preference feature are concatenated to generate a fused feature vector. This fused feature vector is then input into a multilayer perceptron network, and the user experience level profile is calculated using the following formula: ； In the formula, To profile user experience levels, MLP stands for Multilayer Perceptron Network. For skill level vectors, For operating speed characteristics, For error mode characteristics, For interaction preference features, This is a vector concatenation operation.

5. The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment according to claim 1, characterized in that, The adaptive guidance strength is calculated using a logical function based on the user's experience level profile and a preset experience level threshold, including: Read the user's experience level profile, obtain the preset experience level threshold, calculate the difference between the user's experience level profile and the experience level threshold, and obtain the experience deviation. Based on the preset adjustment sensitivity parameters and the maximum guidance intensity value, the adaptive guidance intensity is calculated using the following formula: ； In the formula, For adaptive guidance strength, For logical functions, For maximum guiding strength, To adjust sensitivity, The experience level threshold is the threshold when the user's experience level is... When below the threshold, the guiding strength is close to When the value is above the threshold, the guiding strength approaches 0.

6. The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment according to claim 1, characterized in that, The process involves outputting virtual prompts that match the operator's current experience level based on adaptive guidance strength, and collecting operational data performed by the operator based on the virtual prompts, including: Obtain adaptive guidance strength, a pre-defined large language model, and a computer vision model; The large language model is used to generate text prompts and voice prompts of different levels of detail based on the adaptive guidance intensity. The computer vision model is used to generate three-dimensional graphic annotation parameters of different complexities based on the adaptive guidance intensity. The text prompts, voice prompts, and three-dimensional graphic annotation parameters are combined to generate virtual prompt information. Display virtual prompts in the mixed reality device, collect the operator's hand movements and eye gaze points in the mixed reality device in real time, generate the operation type, operation parameters and operation completion mark performed by the operator, and obtain the original operation record; The state change sequence is obtained from the digital twin model. The state change sequence includes the device parameter change curve and alarm event log. The original operation record is associated with the state change sequence and stored to obtain operation data.

7. The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment according to claim 1, characterized in that, The step of packaging the operation data performed by the operator with multimodal behavioral data to generate a data package to be uploaded to the blockchain includes: According to preset time intervals or preset specific search conditions, all operation data and all multimodal behavior data collected in the current period are aggregated to generate aggregated data records; Add operation timestamps and operator identification to aggregated data records to generate aggregated data records with identification; The tagged aggregated data records are serialized into a structured data format to obtain the data packet to be uploaded to the blockchain.

8. The intelligent training, evaluation, and on-chain certification and traceability system for industrial equipment according to claim 1, characterized in that, The process involves performing zero-knowledge proof computation on the data packet to be uploaded to the blockchain to generate a zero-knowledge proof packet. This packet is then uploaded to the blockchain network for storage, generating an on-chain authentication record, including: Obtain the data packet to be uploaded to the blockchain, and perform proof calculation on the data packet using a zero-knowledge concise non-interactive knowledge proof algorithm to generate a zero-knowledge proof packet. The proof size, verification time, and gas consumption of the zero-knowledge proof packet satisfy the following formula: ； ； ； In the formula, For the size of the zero-knowledge proof packet, , It is an elliptic curve group. To verify the time, For bilinear pairing operations, For scalar multiplication operations, For the input quantity, The gas consumption on the Ethereum network; the information-theoretic security and privacy leakage probability of this zero-knowledge proof package satisfies the formula: ； In the formula, To reduce the probability of privacy breaches for zero-knowledge proof schemes, The probability of privacy breaches in encryption schemes. This represents the probability that the key will be cracked. The zero-knowledge proof package is uploaded to the blockchain network, triggering a smart contract in the blockchain network to perform a verification operation on the zero-knowledge proof package using a preset verification key. After successful verification, the transaction hash value and block height returned by the blockchain network are obtained. The transaction hash value and block height are combined as key-value pairs to generate an on-chain authentication record.

Images