Intelligent recommendation method for Internet information service of medical apparatus and instruments

Abstract

The invention discloses an intelligent recommendation method for an internet information service of a medical instrument, relates to the technical field of intelligent medical treatment and data security, solves the problems of poor data isomerism, model dynamic adjustment and decision credibility of an existing recommendation method, and comprises equipment state sensing based on multi-source sensing data acquisition and edge intelligent processing; performing encryption gradient aggregation and hot update based on federated learning and a dynamic attention algorithm; generating a medical knowledge graph and extracting depth features based on the graph neural network; generating a Pareto optimal recommendation scheme based on multi-objective optimization and digital twinborn simulation; the whole-process data based on block chain multi-chain evidence storage is credible. High-precision and low-delay medical instrument service recommendation is realized, and data privacy security and model credibility are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and more specifically to an intelligent recommendation method for medical device internet information services. Background Technology

[0002] With the development of smart healthcare, intelligent recommendation systems for medical device internet information services are an important guarantee for improving the efficiency of medical services. However, current technical solutions still have technical shortcomings in achieving high-precision, safe and reliable intelligent recommendations.

[0003] In terms of data acquisition, traditional methods often employ single-modal data acquisition, lacking the ability to simultaneously perceive multiple physical quantities. Existing sensor networks typically suffer from low signal conditioning accuracy and low analog-to-digital conversion resolution, resulting in raw data signal-to-noise ratios that fail to meet the demands of high-precision diagnosis. Regarding edge computing architecture, existing systems primarily rely on data filtering and forwarding, failing to fully leverage the local computing capabilities of edge nodes. Traditional encryption methods cannot balance gradient protection and computational efficiency, leading to privacy leaks and performance degradation. Significant deficiencies exist in knowledge modeling; traditional relational database-based storage systems cannot reflect the relationships between medical devices, medical institutions, and medical records, lacking the ability to extract deep features from multidimensional medical knowledge, thus limiting the generalization ability of recommendation models. Federated learning faces unique application challenges in the medical field. Traditional federated averaging algorithms ignore the non-independent and identically distributed nature of medical data and lack dynamic evaluation of edge node credibility. Insufficient version management during model updates affects the convergence efficiency of distributed learning. In the intelligent decision-making stage, there is a lack of multi-objective optimization; existing recommendation algorithms tend to focus on single clinical indicators, failing to comprehensively consider constraints such as equipment reliability, lifecycle costs, and environmental impact. Furthermore, the decision-making scheme lacks a verification process based on physical mechanisms and real-time data; the credibility of the evidence storage is questionable as centralized audit logs are easily tampered with, and traditional blockchain evidence storage cannot meet the storage requirements of high-frequency and multi-type data in medical scenarios; the evidence storage mechanism lacks the ability to dynamically control the evidence storage process and is not applicable to medical compliance requirements.

[0004] Therefore, it is necessary to construct a comprehensive intelligent recommendation method that integrates high-precision perception, edge intelligence, secure federated learning, knowledge graph-driven approach, multi-objective optimization, digital twin verification, and blockchain notarization to overcome the aforementioned technical deficiencies in all aspects. Summary of the Invention

[0005] To address the shortcomings of medical device internet information service recommendation systems, such as insufficient data awareness, edge intelligence, privacy protection, and decision credibility, this invention discloses an intelligent recommendation method for medical device internet information services. This method first employs an edge-cloud collaborative architecture to collect device status data, using an embedded AI acceleration chip and a cryptographic coprocessor for local inference and gradient encryption at the edge. Then, it extracts medical knowledge graphs based on graph neural networks to provide semantic support for downstream decision-making, introducing a dynamic attention-weighted federated learning mechanism to continuously optimize the model while maintaining data privacy. Next, it establishes a multi-objective optimization model applicable to clinical utility, cost, and device reliability, and verifies the generation of Pareto-optimal recommendation schemes using digital twin simulation. Finally, it uses a multi-chain architecture-based blockchain storage unit to perform differentiated hashing and dynamic consensus storage of key data.

[0006] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution: A method for intelligent recommendation of medical device internet information services includes the following steps: S(1), Multi-source data acquisition and preprocessing: The sensor array collects medical device operation data in real time and receives user operation data synchronously through the interactive interface module; after amplifying, filtering and converting the collected raw analog signals into digital signals, they are adaptively transmitted to the edge computing gateway via near-field communication, Bluetooth or Wi-Fi. S(2), Edge intelligent processing and secure transmission: The edge computing gateway cleans and standardizes the received data, and uses an embedded AI acceleration chip to perform local lightweight model inference to generate device status identification results; at the same time, it uses a cryptographic coprocessor to encrypt the model gradients, and transmits the encrypted data and identification results to the central data processing platform through a TLS secure channel. S(3), Knowledge Integration and Feature Learning: The central data processing platform integrates multi-source data and constructs a knowledge graph linking medical devices, institutions, and cases; it then uses a graph neural network computing unit to perform graph convolution calculations on the knowledge graph to generate deep feature representations for downstream optimization modules to read. S(4), Federated Learning and Model Optimization: By coordinating each edge node through the parameter server, aggregating their encrypted model gradients, and after verifying version consistency, distributing the updated global model parameters to each edge node, the model can be dynamically and hotly updated. S(5), Intelligent Recommendation and Digital Verification: Based on the deep feature representation, a mixed integer programming problem with clinical utility, cost-effectiveness, and equipment reliability as objectives is constructed, and the Pareto optimal solution set is solved; the feasibility of candidate recommendation schemes is verified by simulation and data-driven models using a digital twin. S(6), Result Output and Trusted Evidence Storage: The verified recommendation scheme is output to the user terminal, and the blockchain storage unit is used to record the key storage information of device status, model parameter updates and recommendation decisions through multiple parallel private chains to ensure the immutability and traceability of the data.

[0007] In S(1), the sensor group includes a signal conditioning circuit, a first microprocessor, a vibration sensor, a temperature sensor, and an optical encoder; the signal conditioning circuit includes an operational amplifier and a low-pass filter; the first microprocessor adopts an ARM Cortex-M4 architecture and includes an analog-to-digital converter with 16-bit resolution; the vibration sensor is a MEMS piezoelectric accelerometer, the temperature sensor is a PT100 platinum resistance thermometer, and the optical encoder is an incremental photoelectric encoder; the output terminal of the physical sensor is connected to the input terminal of the signal conditioning circuit through a shielded cable, and the output terminal of the signal conditioning circuit is connected to the analog-to-digital converter pin of the first microprocessor through PCB traces.

[0008] In S(2), the lightweight model inference module includes an embedded AI acceleration chip, a model memory, and an encryption coprocessor; The lightweight model inference module operates through the following steps: S(21), Model loading and initialization: The embedded AI acceleration chip loads the pruned and quantized neural network model from the model memory into the chip's internal cache via the SPI interface; the encryption coprocessor generates an asymmetric key pair and uploads it to the federated learning scheduling module for registration via the first secure communication module; S(22), Local Data Inference: The data preprocessing module transmits the cleaned and standardized sensor data to the input buffer of the embedded AI acceleration chip via the PCIe bus; the embedded AI acceleration chip calls the loaded neural network model to perform forward inference calculations, generating device status recognition results and local decision suggestions; S(23), Gradient Calculation and Encryption The embedded AI acceleration chip calculates the model gradient based on the difference between the local inference result and the real label using the backpropagation algorithm; the encryption coprocessor uses the SM4 national cryptographic algorithm to encrypt the calculated model gradient and generate ciphertext gradient data. S(24), Secure transmission and parameter update: The first secure communication module uploads the encrypted gradient data to the gradient aggregator via a TLS secure channel; the lightweight model inference module periodically polls the parameter server to obtain the aggregated and updated global model parameters; after verifying the model version consistency, the model version manager distributes the new model parameters to each edge node via an encrypted channel. S(25), Model hot update: After reading the model parameters, the embedded AI acceleration chip pauses inference and reads the new model parameters into the spare area of ​​the model memory. After reading is complete, the embedded AI acceleration chip returns to the new model to continue providing services and updating the model.

[0009] In S(3), the process of feature learning performed by the graph neural network computing unit includes the following steps: S(31), Graph data preparation and parallel loading: The medical knowledge graph is obtained from the knowledge graph construction module, the topology is converted into an adjacency matrix, and the attribute data of each node in the graph is converted into a node feature matrix. Then, the adjacency matrix and the node feature matrix are partitioned and loaded into a multi-GPU-based parallel computing memory. S(32), Multi-layer graph convolution calculation and feature refinement: In a multi-layer graph convolutional network, the feature representations of all nodes are iteratively updated: for each target node in the current layer, the features of directly connected neighbor nodes are collected through an aggregation function to generate the target node's neighbor context vector; the neighbor context vector is fused with the target node's own current layer feature vector, and the fused result is input into a fully connected layer with a non-linear activation function for linear transformation and non-linear mapping, and the higher-order feature representation of the node in the next layer is output. S(33), Deep Feature Output and High-Speed ​​Transmission: After completing the graph convolution calculation of the preset number of layers, the depth feature representation of all nodes in the final output is directly written into the shared memory region through NVLink high-speed interconnect technology, so that the multi-objective optimization solver can perform zero-latency reading and subsequent optimization solving.

[0010] In S(4), the federated learning scheduling module includes a parameter server, a gradient aggregator, and a model version manager; The federated learning model optimization process introduces a dynamic attention weighting mechanism, which includes two steps: dynamic attention weight calculation and global model parameter update. The dynamic attention weights are calculated as follows: In formula (1), This represents the dynamic attention weight of the k-th edge node at time t; Use the Sigmoid activation function; This is the attention moderating factor; Let L2 norm be the gradient of the k-th edge node; Represents the global gradient norm; This represents the credibility score of the k-th node at time t; The global model parameters are updated as follows: In formula (2), This represents the global model parameters at iteration t; K is the total number of edge nodes. is the number of local training samples owned by the k-th edge node, and n is the total number of local training samples of all participating nodes; It is the adaptive learning rate of the k-th node; Represents the L2 regularization and adaptive L1 regularization coefficients; Represents the L2 regularization and adaptive L1 regularization coefficients; Let L represent the parameter tensor of the l-th layer of the neural network model; L is the total number of layers in the neural network model. This indicates that the private key SK is used to decrypt the encrypted content; This indicates that the input content is encrypted using the public key PK. In S(5), the optimization problem constructed by the multi-objective optimization solver introduces spatiotemporal dynamic constraints, including knowledge graph relation reasoning and multi-objective optimization: Knowledge graph dynamic relationship reasoning uses a graph attention network, and node features are updated as follows: In formula (3), It is the original feature vector of node i; Let be the updated feature vector of node i; W is the weight matrix. It is the attention weight vector; it is the attention weight vector; LeakyReLU is the activation function; For activation functions; Represents the set of neighboring nodes of node i; j and k are the indices of the neighboring nodes; Indicates feature concatenation operation; Normalize attention weights; The multi-objective optimization problem is defined by the following formula: In formula (4), All are weighting coefficients; SNR is the signal-to-noise ratio; M is the total number of tasks, devices, or subsystems participating in the optimization. It is the dynamic weight of the j-th object at time t; It is the precision index of the j-th object; It is the time decay coefficient of the j-th object; The initial investment cost is T; T is the optimization time period. It is the discount rate; At any moment The maintenance cost function; Let be the conditional fault probability density function; It is the mean time between failures (MTBF) of the i-th device; It is the health status impact coefficient; This is the health status indicator of the i-th device; This represents the total carbon emissions within the time period. In S(5), the simulation model based on physical mechanisms uses a multibody dynamics solver, and the real-time data-driven model uses a long short-term memory neural network; the two are interconnected through a data fusion processor. The real-time data-driven model introduces a multi-head spatiotemporal attention mechanism, and its working process includes two parts: physical mechanism simulation and real-time state update. The dynamic process of the simulation model based on physical mechanisms is defined by the following set of differential equations: In formula (5), J is the moment of inertia. It is angular displacement; Angular acceleration; It is the torque constant; B is the driving current; B is the viscous damping coefficient. Angular velocity; It is a nonlinear frictional torque; It is the load torque; It is the observation output; It is the noise figure; Noise term; The state update of the real-time data-driven model is defined by the following formula: In formula (6), This represents the core state of the model at time t; H is the number of heads in the multi-head attention mechanism. Let h be the attention weight for the h-th head; It is element-wise multiplication; These are the forget gate parameters for the h-th head at time t; This represents the input gate parameters of the h-th head at time t; This represents the candidate state of the h-th head at time t; represents the weighting coefficients for the frequency domain features; S is the number of frequency bands. It is the weight of the s-th frequency band; It is the linear transformation matrix of the s-th frequency band; It is a Fast Fourier Transform; It is a vibration signal; It is a temperature signal; Other coding features; It is the frequency range of the s-th frequency band; These are the weighting coefficients for the residual connections; It is a residual join function. In S(6), the trusted evidence storage using the blockchain evidence storage unit includes the following steps: S(61), Multi-source data classification and preprocessing: The key data from data collection, model optimization, and recommendation decision-making stages are classified; real-time equipment operating status data are classified as high-frequency lightweight data, model parameter update records generated by federated learning are classified as mid-frequency core asset data, and the decision-making logic and final results of the recommendation scheme are classified as low-frequency key decision data. S(62), Execution of Differentiated Hash Evidence Storage Strategy: For the high-frequency lightweight data, a timed polling hash strategy is executed to calculate the hash value of the state snapshot at preset time intervals; for the mid-frequency core asset data, an event-driven hash strategy is executed to calculate the hash value of the complete parameters only after the model version manager confirms that the global model update is successful; for the low-frequency key decision data, an instant trigger hash strategy is executed to immediately calculate the hash value of its decision path and result after the digital twin completes the scheme verification and approval output. S(63), Parallel Multi-Chain Classification, Evidence Storage and Association: The hash values, timestamps, and metadata of the above three types of data are submitted to three parallel private chains (41) for storage. At the same time, a cross-chain association index is established for the evidence records generated on different chains for the same business event to ensure the integrity of the global audit trail.

[0011] The consensus and notarization process of the parallel private blockchain includes the following steps: S(631), Dynamic consensus group formation and node admission verification: Based on the type of evidence stored, nodes are dynamically selected from the pre-authorized node pool to form the verification committee for this consensus. Members of the verification committee must verify their identity and permissions through a dual authentication mechanism based on digital certificates and attribute-based cryptography. S(632), Improved Byzantine Fault-Tolerant Consensus Implementation: Authorized nodes broadcast the evidence storage data transaction package to the dynamic consensus group; the master node verifies the validity and format compliance of the transaction package, and after the verification is passed, it assigns a sequence number and broadcasts a pre-preparation message; the verification committee member nodes conduct multiple rounds of communication voting, and when more than two-thirds of the nodes have collected valid signatures to agree, a consensus is reached; after the consensus is reached, the evidence storage data block is synchronized to all member nodes of this dynamic consensus group; S(633), Evidence Lifecycle Management: Add an updatable status marker to the newly generated evidence storage block. The initial status marker is valid. Set up an arbitration node to accept objection appeals. After the arbitration chain makes a ruling, the original evidence storage status marker can be modified to disputed or invalid. The evidence storage status is dynamically managed without modifying the original data.

[0012] The positive and beneficial technical effects of this invention are as follows: By combining a multi-physical quantity sensor group and a high-precision signal conditioning circuit, the signal-to-noise ratio and integrity of raw data acquisition are improved. Through the collaboration of an embedded AI acceleration chip and a cryptographic coprocessor, edge-side inference is achieved while ensuring gradient security. Deep feature extraction from a knowledge graph based on graph neural networks solves the problem of complex medical relationship modeling using traditional relational databases. Dynamic attention-weighted federated learning improves the model aggregation efficiency and robustness under non-independent and identically distributed medical data. Through a multi-objective optimization model and a digital twin verification system, the collaborative optimization of indicators such as clinical utility and equipment reliability is systematically addressed. A multi-chain architecture blockchain evidence storage unit and a dynamic consensus mechanism ensure the immutability and traceability of data throughout the entire medical recommendation process. These improvements resolve systemic technical deficiencies in the background technologies, such as low data accuracy, insufficient privacy protection, shallow knowledge modeling, low federated learning efficiency, single decision-making dimension, and low evidence credibility. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart of the intelligent recommendation method for medical device internet information services of the present invention; Figure 2 This is a schematic diagram of the sensor group for the intelligent recommendation method of medical device Internet information service of the present invention; Figure 3 This is a schematic diagram of the lightweight model inference module of the intelligent recommendation method for medical device Internet information services of the present invention; Figure 4 This is a schematic diagram of the federated learning scheduling module of the intelligent recommendation method for medical device Internet information services of the present invention; Figure 5 This is a flowchart of the lightweight model inference module of the intelligent recommendation method for medical device Internet information services of the present invention; Figure 6 This is a flowchart of the neural network calculation unit of the intelligent recommendation method for medical device internet information services of the present invention; Figure 7 This is a flowchart of the blockchain evidence storage unit of the intelligent recommendation method for medical device internet information services of the present invention; Figure 8 This invention provides a flowchart of the consensus and evidence storage process for a parallel private blockchain in the intelligent recommendation method for medical device internet information services. Figure 9 This is a diagram illustrating an application scenario of the intelligent recommendation method for medical device internet information services according to the present invention.

[0014] In the figure: signal conditioning circuit 11, first microprocessor 12, vibration sensor 13, temperature sensor 14, optical encoder 15, embedded AI acceleration chip 21, model memory 22, encryption coprocessor 23, parameter server 41, gradient aggregator 42, model version manager 43. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Embodiment 1: A method for intelligent recommendation of medical device internet information services includes the following steps: S(1), Multi-source data acquisition and preprocessing: The sensor array collects medical device operation data in real time and receives user operation data synchronously through the interactive interface module; after amplifying, filtering and converting the collected raw analog signals into digital signals, they are adaptively transmitted to the edge computing gateway via near-field communication, Bluetooth or Wi-Fi. S(2), Edge intelligent processing and secure transmission: The edge computing gateway cleans and standardizes the received data, and uses an embedded AI acceleration chip to perform local lightweight model inference to generate device status identification results; at the same time, it uses a cryptographic coprocessor to encrypt the model gradients, and transmits the encrypted data and identification results to the central data processing platform through a TLS secure channel. S(3), Knowledge Integration and Feature Learning: The central data processing platform integrates multi-source data and constructs a knowledge graph linking medical devices, institutions, and cases; it then uses a graph neural network computing unit to perform graph convolution calculations on the knowledge graph to generate deep feature representations for downstream optimization modules to read. S(4), Federated Learning and Model Optimization: By coordinating each edge node through the parameter server, aggregating their encrypted model gradients, and after verifying version consistency, distributing the updated global model parameters to each edge node, the model can be dynamically and hotly updated. S(5), Intelligent Recommendation and Digital Verification: Based on the deep feature representation, a mixed integer programming problem with clinical utility, cost-effectiveness, and equipment reliability as objectives is constructed, and the Pareto optimal solution set is solved; the feasibility of candidate recommendation schemes is verified by simulation and data-driven models using a digital twin. S(6), Result Output and Trusted Evidence Storage: The validated recommendation scheme is output to the user terminal, and a blockchain notarization unit is used to record key notarization information such as device status, model parameter updates, and recommendation decisions through multiple parallel private chains, ensuring the immutability and traceability of the data. Example 2: In S(1), the sensor group includes a signal conditioning circuit 11, a first microprocessor 12, a vibration sensor 13, a temperature sensor 14, and an optical encoder 15; The signal conditioning circuit 11 includes an operational amplifier and a low-pass filter; the first microprocessor 12 adopts an ARM Cortex-M4 architecture and includes an analog-to-digital converter with 16-bit resolution; the vibration sensor 13 adopts a MEMS piezoelectric accelerometer, the temperature sensor 14 adopts a PT100 platinum resistance thermometer, and the optical encoder 15 adopts an incremental photoelectric encoder. The output terminal of the physical sensor is connected to the input terminal of the signal conditioning circuit (11) via a shielded cable, and the output terminal of the signal conditioning circuit 11 is connected to the analog-to-digital conversion pin of the first microprocessor 12 via PCB traces.

[0016] In this embodiment, taking the condition monitoring of industrial rotating equipment as an example, the sensor group of this solution is used to collect vibration, temperature, and rotational speed of key components of the equipment. The collected vibration, temperature, and rotational speed signals are transmitted to the signal conditioning circuit through shielded cables. The operational amplifier of the signal conditioning circuit amplifies the signal, the low-pass filter filters out interference, and the analog signals of all signals are transmitted to the analog-to-digital converter pin of the first microprocessor through PCB traces. The first microprocessor adopts the ARM Colog-M4 architecture, and its 16-bit resolution analog-to-digital converter converts the analog signal into a high-precision digital signal. During the experiment, the system automatically compensates for the cross-interference of ambient temperature on vibration and identifies the characteristic frequency drift caused by mechanical wear through built-in algorithms, making the experimental results more reflective of the equipment condition. The experimental results are shown in Table 1.

[0017] Table 1 Comparison of Equipment Condition Monitoring Performance

[0018] This experiment achieved synchronous automatic acquisition of multiple physical quantities in rotating equipment without human intervention. The acquired state characteristic values ​​are more stable, making it easier to detect potential faults and improving the efficiency and reliability of predictive maintenance. Example 3: In S(2), the lightweight model inference module includes an embedded AI acceleration chip 21, a model memory 22, and an encryption coprocessor 23; The lightweight model inference module operates through the following steps: S(21), Model loading and initialization: The embedded AI acceleration chip 21 loads the pruned and quantized neural network model from the model memory 22 into the chip's internal cache via the SPI interface; the encryption coprocessor 23 generates an asymmetric key pair and uploads it to the federated learning scheduling module for registration via the first secure communication module; S(22), Local Data Inference: The data preprocessing module transmits the cleaned and standardized sensor data to the input buffer of the embedded AI acceleration chip 21 via the PCIe bus; the embedded AI acceleration chip 21 calls the loaded neural network model to perform forward inference calculations, generating device status recognition results and local decision suggestions; S(23), Gradient Calculation and Encryption The embedded AI acceleration chip 21 calculates the model gradient based on the difference between the local inference result and the real label using the backpropagation algorithm; the encryption coprocessor 23 uses the SM4 national cryptographic algorithm to encrypt the calculated model gradient and generate ciphertext gradient data. S(24), Secure transmission and parameter update: The first secure communication module uploads the encrypted gradient data to the gradient aggregator via a TLS secure channel; the lightweight model inference module periodically polls the parameter server to obtain the aggregated and updated global model parameters; after verifying the model version consistency, the model version manager distributes the new model parameters to each edge node via an encrypted channel. S(25), Model hot update: After reading the model parameters, the embedded AI acceleration chip 21 pauses inference and reads the new model parameters into the spare area of ​​the model memory 22. After reading, the embedded AI acceleration chip 21 returns to the new model to continue providing services and updating the model.

[0019] In the experimental example, the status of industrial edge devices is taken as the research object. The lightweight model inference module of this solution is used to analyze the device operation data in real time. At startup, the embedded AI acceleration chip downloads the encrypted neural network model from the model storage to memory. The encryption coprocessor generates an asymmetric key pair and registers it with the federated learning scheduling module through the secure communication module. During local inference, the sensor timing data is sent to the input buffer of the acceleration chip. The acceleration chip performs forward inference through the model, provides device status and maintenance suggestions, and performs backward gradient calculation according to different actual operating conditions of the model. The encryption coprocessor uses the SM4 national cryptographic algorithm to encrypt the gradient and uploads it to the cloud gradient aggregator through a TLS secure channel. The module obtains the global model parameters updated by federated aggregation by polling the parameter server. The model version manager sends the new parameters to the model storage through an encrypted channel through version consistency verification. During the update, the embedded AI acceleration chip stops inference, writes the new model parameters to the spare area of ​​the model storage, and seamlessly switches to the new model to continue service, achieving seamless hot update. The performance comparison of edge AI inference is shown in Table 2.

[0020] Table 2 Comparison of Edge AI Inference Performance

[0021] The experimental results are shown in Table 2. This device, composed of sensor data acquisition and lightweight secure federated inference, improves the real-time performance, accuracy, and security of industrial equipment condition monitoring, providing a reliable automated solution for predictive maintenance. Example 4: In S(3), the process of feature learning performed by the graph neural network computing unit includes the following steps: S(31), Graph data preparation and parallel loading: The medical knowledge graph is obtained from the knowledge graph construction module, the topology is converted into an adjacency matrix, and the attribute data of each node in the graph is converted into a node feature matrix. Then, the adjacency matrix and the node feature matrix are partitioned and loaded into a multi-GPU-based parallel computing memory. S(32), Multi-layer graph convolution calculation and feature refinement: In a multi-layer graph convolutional network, the feature representations of all nodes are iteratively updated: for each target node in the current layer, the features of directly connected neighbor nodes are collected through an aggregation function to generate the target node's neighbor context vector; the neighbor context vector is fused with the target node's own current layer feature vector, and the fused result is input into a fully connected layer with a non-linear activation function for linear transformation and non-linear mapping, and the higher-order feature representation of the node in the next layer is output. S(33), Deep Feature Output and High-Speed ​​Transmission: After completing the graph convolution calculation of the preset number of layers, the depth feature representation of all nodes in the final output is directly written into the shared memory region through NVLink high-speed interconnect technology, so that the multi-objective optimization solver can perform zero-latency reading and subsequent optimization solving.

[0022] In this embodiment, taking intelligent medical clinical decision support as the scenario, the graph neural network computing unit in the solution is used to mine the relationships and features of the medical knowledge graph. In the graph data preparation stage, the system extracts a large-scale heterogeneous graph containing entities such as diseases, symptoms, drugs, and genes from the knowledge graph modeling module, transforms its topology into a sparse adjacency matrix, and transforms the multimodal attributes of entities into node feature matrices, loading them in parallel into the memory of a multi-GPU cluster. In the feature learning stage, the computing unit performs multi-layer graph convolution calculations on the graph. For each target node, each layer uses an aggregation function to collect the features of first-order neighbor nodes, thus forming multiple neighbor context vectors. All node context vectors are transformed using a fully connected layer with an activation function according to the features of the target node, extracting new node features containing complex high-order relationships. After multiple iterations, the deep semantic embedding representation of all entities is output. Finally, the last-dimensional feature vector is directly written to a memory region shared with the multi-objective optimization solver via NVLink high-speed interconnection, enabling zero-copy reading and millisecond-level reading, achieving real-time preprocessing for subsequent real-time diagnostic reasoning and optimization schemes. Experimental results are shown in Table 3.

[0023] Table 3 Comparison of computational performance of graph neural networks

[0024] Graph neural network computing units based on multi-GPU parallelism and high-speed connectivity can achieve efficient and deep feature learning on large-scale medical knowledge graphs. The extracted high-order relation features can help further improve the accuracy and inference speed of clinical decision-making models, providing technical support for real-time intelligent assistance in complex medical scenarios. Example 5: In S(4), the federated learning scheduling module includes a parameter server 41, a gradient aggregator 42, and a model version manager 43; The federated learning model optimization process introduces a dynamic attention weighting mechanism, which includes two steps: dynamic attention weight calculation and global model parameter update. The dynamic attention weights are calculated as follows: In formula (1), This represents the dynamic attention weight of the k-th edge node at time t; Use the Sigmoid activation function; This is the attention moderating factor; Let L2 norm be the gradient of the k-th edge node; Represents the global gradient norm; This represents the credibility score of the k-th node at time t; The global model parameters are updated as follows: In formula (2), This represents the global model parameters at iteration t; K is the total number of edge nodes. is the number of local training samples owned by the k-th edge node, and n is the total number of local training samples of all participating nodes; It is the adaptive learning rate of the k-th node; Represents the L2 regularization and adaptive L1 regularization coefficients; Represents the L2 regularization and adaptive L1 regularization coefficients; Let L represent the parameter tensor of the l-th layer of the neural network model; L is the total number of layers in the neural network model. This indicates that the private key SK is used to decrypt the encrypted content; This indicates that the input content is encrypted using a public key (PK). The experiment was run in a simulated industrial edge computing environment, with a hardware environment consisting of a central server and five edge computing nodes, including a parameter server, gradient aggregator, model version manager, and client. The software was based on the PyTorch and Flower frameworks, using a multidimensional sensor signal industrial equipment fault dataset of 8000 entries, distributed to the edge nodes in a non-independent and identically distributed manner.

[0025] The purpose of the experiment is to verify the mathematical form of the dynamic attention weighting system added by the federated learning scheduling module. Formula (1) is the calculation of dynamic attention weights, and the weight of the k-th node at time t is... By the Sigmoid function The calculation shows that the input has two linear combinations: one is the local gradient L2 norm of the node. With global gradient norm The ratio; secondly, the credibility score of the node at the current moment. The attention adjustment coefficient is dynamically calculated by the model version manager based on historical behavior. and By controlling the relative intensity of the above effects, and after repeated experiments and grid searches, the values ​​were set to... =0.8, =0.5. Formula (2) is the parameter update calculation for the global model. When the parameter server calculates the information of all its nodes in round t+1, in addition to the proportion of its local information, it also calculates the information of all its nodes. In addition to weighting, dynamic attention weights were also introduced. Additionally, an L2 regularization term was set. and adaptive L1 regularization, the latter restricting the l-th layer Determine the sparsity intensity and initially set In the encrypted space, the local update is encrypted using the public key PK and the aggregation result is decrypted using the private key SK to ensure data security. The two formulas realize the federated learning optimization process of protecting node contributions, defending against unreliable updates and protecting communication security. The experimental process compares the performance difference between the proposed method and the classic federated averaging algorithm. First, a five-layer fully connected neural network is initialized. Three nodes are randomly selected to participate in 100 rounds of federated training. The control group uses the standard FedAvg algorithm. The experimental group executes the complete process. In each round, the participating nodes upload the encrypted model after local training. After receiving the model, the gradient aggregator calculates the dynamic attention weight of each node according to formula (1). According to formula (2), the data weight, dynamic attention weight and regularization term are combined to perform weighted aggregation and decryption to generate a new global model. The model version manager updates the credibility score of each node at the same time. To verify the robustness, a node is randomly selected to simulate a malicious attack and upload a gradient with negative perturbation. During the training process, the accuracy and convergence rate of the global model on the independent validation set are continuously monitored. Finally, the final test set is tested. The performance comparison results of the federated learning strategy are shown in Table 4. The experimental process compared the performance differences between the proposed method and the classic federated averaging algorithm. First, a five-layer fully connected neural network was initialized, and three nodes were randomly selected to participate in 100 rounds of federated training. The control group used the standard FedAvg algorithm. The experimental group executed the complete process. In each round, after the participating nodes trained locally, they uploaded the encrypted model. After receiving the model, the gradient aggregator calculated the dynamic attention weights of each node according to formula (1). According to formula (2), it combined the data weights, dynamic attention weights, and regularization terms for weighted aggregation and decryption to generate a new global model. The model version manager simultaneously updated the credibility score of each node. To verify robustness, a node was randomly selected to simulate a malicious attack and uploaded a gradient with negative perturbation. During training, the accuracy and convergence rate of the global model on the independent validation set were continuously monitored, and finally, the final test set was tested. The performance comparison results of the federated learning strategy are shown in Table 4.

[0026] Table 4 Performance Comparison Results of Federated Learning Strategies

[0027] Table 4 shows that dynamic attention weighting can effectively improve the final accuracy and convergence speed of the model, and shorten the impact period of malicious attacks by nearly half.

[0028] Table 5 Examples of dynamic attention weight changes during training

[0029] Table 5 shows the weight comparison before and after the introduction of malicious attacks. After a malicious attack occurs, the weight of malicious nodes can be quickly reduced and the weight of reliable nodes can be increased by formula (1) to reduce abnormal updates.

[0030] The experiment establishes and verifies a federated learning scheduling module based on formulas (1) and (2), with dynamic attention weighting and weighted contributions from each edge node. This enhances robustness against malicious attacks while ensuring convergence efficiency and final results. Security and generalizability for industrial distributed collaborative diagnostics are achieved through encrypted aggregation and regularization design. Example 6: In S(5), the optimization problem constructed by the multi-objective optimization solver introduces spatiotemporal dynamic constraints, including knowledge graph relation reasoning and multi-objective optimization: Knowledge graph dynamic relationship reasoning uses a graph attention network, and node features are updated as follows: In formula (3), It is the original feature vector of node i; Let be the updated feature vector of node i; W is the weight matrix. It is the attention weight vector; it is the attention weight vector; LeakyReLU is the activation function; For activation functions; Represents the set of neighboring nodes of node i; j and k are the indices of the neighboring nodes; Indicates feature concatenation operation; Normalize attention weights; The multi-objective optimization problem is defined by the following formula: In formula (4), All are weighting coefficients; SNR is the signal-to-noise ratio; M is the total number of tasks, devices, or subsystems participating in the optimization. It is the dynamic weight of the j-th object at time t; It is the precision index of the j-th object; It is the time decay coefficient of the j-th object; The initial investment cost is T; T is the optimization time period. It is the discount rate; At any moment The maintenance cost function; Let be the conditional fault probability density function; It is the mean time between failures (MTBF) of the i-th device; It is the health status impact coefficient; This is the health status indicator of the i-th device; This represents the total carbon emissions over the specified time period. Experimental verification was conducted using a high-performance server based on an NVIDIA RTX A6000, running Ubuntu 20.04 with Python 3.9 and the PyTorchGEic and Pymoo libraries. Data came from a simulated industrial park digital twin system, where 120 devices created an industrial knowledge graph comprising 200 entities and 500 relationships.

[0031] Experimental verification includes knowledge graph dynamic relationship reasoning and spatiotemporal constraint multi-objective optimization solver. Knowledge graph dynamic relationship reasoning uses graph attention network; Formula (3) defines the node feature update mechanism, the normalized attention weight of the node and its neighbors, the differentiated aggregation context, the attention weight is linearly transformed to the node feature concatenation according to the attention vector a and the weight matrix W, LeakyReLU activation and Softmax normalization, the node update feature neighbor node transformation feature weighted sum, and the activation function σ outputs the spatiotemporal dependency of the device state.

[0032] Formula (4) defines a multi-objective optimization problem. The maximization of the comprehensive utility function has four weighted items: signal-to-noise ratio, time-varying task weight, and accuracy index time decay; negative life cycle cost item: initial investment and fault probability discounted maintenance cost integral; system reliability benefit item: mean time between failures and equipment health status change rate; negative environmental cost item: total carbon emissions, with weighting coefficients. The targets were balanced and set to 0.4, 0.3, 0.2, and 0.1 respectively.

[0033] In the experiment, the proposed solver was compared with the traditional multi-objective particle swarm optimization algorithm. First, a graph attention network was trained to generate spatiotemporally dependent device node embedding features. A one-year optimization cycle was set, and 20 monitoring tasks were assigned to appropriate maintenance plans within the budget and carbon limits. The control group used a simple static multi-objective function. In each optimization iteration, the experimental group updated the device node features of the trained graph network and substituted the updated features into Formula 4 to calculate each index. Finally, a decomposition-based multi-objective evolutionary algorithm was used to solve the problem. Dynamic adaptability was tested by injecting device performance degradation and carbon emission price fluctuation events on days 100 and 200, respectively, and the results of the two methods were obtained. The performance comparison results of different optimization methods are shown in Table 6.

[0034] Table 6 Performance Comparison Results of Different Optimization Methods

[0035] The data in Table 6 show that the solver of this scheme is superior to traditional methods in terms of overall utility, frontier convergence quality, dynamic response speed, and environmental benefits.

[0036] Table 7 Changes in key dynamic indicators during the optimization process

[0037] Table 7 shows that the change in node attention entropy is the graph network's ability to aggregate differentiated information between stable and sudden states. The health state change rate and the peak of the failure probability correspond to dynamic events, and the optimizer restores and improves the overall utility value.

[0038] Experiments demonstrate the effectiveness of the multi-objective optimization solver. The graph attention network proposed in formula (3) can dynamically encode the relationships between devices, providing context for optimization. The optimization problem proposed in formula (4) can model the dynamic trade-offs between multiple objectives, and the solver arrives at a better, more robust, and more sustainable choice. Experiments verify the advancement and effectiveness of this method in dealing with complex dynamic optimization problems in industrial systems. Example 7: In S(5), the simulation model based on physical mechanism adopts a multibody dynamics solver, and the real-time data-driven model adopts a long short-term memory neural network; the two are interconnected through a data fusion processor. The real-time data-driven model introduces a multi-head spatiotemporal attention mechanism, and its working process includes two parts: physical mechanism simulation and real-time state update. The dynamic process of the simulation model based on physical mechanisms is defined by the following set of differential equations: In formula (5), J is the moment of inertia. It is angular displacement; Angular acceleration; It is the torque constant; B is the driving current; B is the viscous damping coefficient. Angular velocity; It is a nonlinear frictional torque; It is the load torque; It is the observation output; It is the noise figure; Noise term; The state update of the real-time data-driven model is defined by the following formula: In formula (6), This represents the core state of the model at time t; H is the number of heads in the multi-head attention mechanism. Let h be the attention weight for the h-th head; It is element-wise multiplication; These are the forget gate parameters for the h-th head at time t; This represents the input gate parameters of the h-th head at time t; This represents the candidate state of the h-th head at time t; represents the weighting coefficients for the frequency domain features; S is the number of frequency bands. It is the weight of the s-th frequency band; It is the linear transformation matrix of the s-th frequency band; It is a Fast Fourier Transform; It is a vibration signal; It is a temperature signal; Other coding features; It is the frequency range of the s-th frequency band; These are the weighting coefficients for the residual connections; This is a residual connection function. The experiment was conducted on a workstation based on a high-performance GPU, using the MATLAB Simscape and pyTorch frameworks. The experimental object was a servo turntable system, and its real-time current, position, vibration, and temperature were obtained under various operating conditions at a sampling frequency of 1 kHz.

[0039] The purpose of the experiment is to connect the physical simulation model and the real-time data-driven model through the data fusion processor. The physical simulation model consists of a set of differential equations as shown in formula (5), which represent the dynamic equilibrium relationship of the turntable system. The nonlinear frictional torque is caused by the actual frictional torque, and the observed output is caused by sensor noise. This model serves as a priori model.

[0040] Formula (6) defines the real-time data-driven model state update, which continues the three characteristics of the long short-term memory network: multi-head attention head computation weighted fusion of multiple modes; frequency domain physical feature fusion term strengthens frequency domain physical feature fusion, extracts and weightedly fuses multiple frequency band signals from Fourier transforms of real-time vibration, temperature, etc., and connects the cross-time step residuals of the frequency domain information corresponding to the mechanical state.

[0041] The experimental process compared the state estimation and early fault warning performance of the pure physics model, the standard LSTM model, and the hybrid model. First, a pure physics simulation was run to obtain the theoretical trajectory. Then, the standard LSTM and the hybrid model were trained using real data. The data fusion processor incorporated the theoretical motion state calculated by the physics simulation as an additional feature, which, together with the real sensor data, was introduced into the data-driven model based on formula (6). Two faults, bearing wear and motor overheating, were simulated to evaluate the model's displacement estimation accuracy and fault warning capability. The performance comparison results of different models are shown in Table 8.

[0042] Table 8 Performance Comparison Results of Different Models

[0043] The results in Table 8 show that the hybrid model is significantly better than the single model in terms of state estimation accuracy and the timeliness and accuracy of fault early warning.

[0044] Table 9. Analysis of Key Mechanisms within the Hybrid Model

[0045] Table 9 shows that the model's attention is more focused in the early stage of the fault, and the contribution ratio of frequency domain features changes significantly, indicating the feasibility of frequency domain fusion and attention acquisition of physical anomalies in formula (6).

[0046] Experiments have demonstrated that the hybrid modeling framework is feasible. Equation (5) provides the physical priors, and Equation (6) uses multi-head spatiotemporal attention and frequency domain hybridization to obtain the relationship between data and physical laws. The hybrid model has high accuracy and strong early detection capabilities, making it suitable for intelligent monitoring of complex equipment. Example 8: In S(6), the trusted evidence storage using the blockchain evidence storage unit includes the following steps: S(61), Multi-source data classification and preprocessing: The key data from data collection, model optimization, and recommendation decision-making stages are classified; real-time equipment operating status data are classified as high-frequency lightweight data, model parameter update records generated by federated learning are classified as mid-frequency core asset data, and the decision-making logic and final results of the recommendation scheme are classified as low-frequency key decision data. S(62), Execution of Differentiated Hash Evidence Storage Strategy: For the high-frequency lightweight data, a timed polling hash strategy is executed to calculate the hash value of the state snapshot at preset time intervals; for the mid-frequency core asset data, an event-driven hash strategy is executed to calculate the hash value of the complete parameters only after the model version manager confirms that the global model update is successful; for the low-frequency key decision data, an instant trigger hash strategy is executed to immediately calculate the hash value of its decision path and result after the digital twin completes the scheme verification and approval output. S(63), Parallel Multi-Chain Classification, Evidence Storage and Association: The hash values, timestamps, and metadata of the above three types of data are submitted to three parallel private chains (41) for storage. At the same time, a cross-chain association index is established for the evidence records generated on different chains for the same business event to ensure the integrity of the global audit trail.

[0047] In this embodiment, taking the trusted traceability of the entire lifecycle of an intelligent production line as the application scenario, the blockchain evidence storage unit of this solution is used to perform hierarchical trusted evidence storage of multi-source heterogeneous data. The main data of each stage are divided into high-frequency lightweight data based on the real-time status data of the equipment obtained by sensors; the updated global model parameters confirmed by the federated learning model version manager are divided into mid-frequency main asset data; and the final production optimization plan after digital twin approval is divided by low-frequency key decision data. Under different hash evidence storage methods, for high-frequency equipment data, the SHA-256 hash of its status snapshot is calculated at minute intervals; for model parameters, the hash of the complete parameters of the new version model is calculated after each global aggregation by federated learning; and for decision plans, the hash of the decision logic and decision results is calculated immediately after the digital twin approval and execution. The hash values, timestamps, and metadata of the three types of data are submitted to three private chains respectively, establishing a cross-chain association index for the evidence storage records on different chains for the same production optimization event. The experimental results are shown in Table 10.

[0048] Table 10 Comparison of Blockchain Evidence Storage Performance

[0049] This demonstrates that a blockchain-based evidence storage solution, based on multi-source data hierarchy and multi-chain parallel storage, can ensure the trustworthiness and controllability of the entire operation process of an industrial intelligent system. It maximizes cost savings on high-frequency data storage while ensuring that important assets and decisions are not tampered with, providing reliable support for quality traceability and compliance auditing in complex industrial scenarios. Example 9: The consensus and notarization process of the parallel private blockchain includes the following steps: S(631), Dynamic consensus group formation and node admission verification: Based on the type of evidence stored, nodes are dynamically selected from the pre-authorized node pool to form the verification committee for this consensus. Members of the verification committee must verify their identity and permissions through a dual authentication mechanism based on digital certificates and attribute-based cryptography. S(632), Improved Byzantine Fault-Tolerant Consensus Implementation: Authorized nodes broadcast the evidence storage data transaction package to the dynamic consensus group; the master node verifies the validity and format compliance of the transaction package, and after the verification is passed, it assigns a sequence number and broadcasts a pre-preparation message; the verification committee member nodes conduct multiple rounds of communication voting, and when more than two-thirds of the nodes have collected valid signatures to agree, a consensus is reached; after the consensus is reached, the evidence storage data block is synchronized to all member nodes of this dynamic consensus group; S(633), Evidence Lifecycle Management: Add an updatable status marker to the newly generated evidence storage block. The initial status marker is valid. Set up an arbitration node to accept objection appeals. After the arbitration chain makes a ruling, the original evidence storage status marker can be modified to disputed or invalid. The evidence storage status is dynamically managed without modifying the original data.

[0050] In this embodiment, taking the collaborative data storage of the manufacturing supply chain as an example, the parallel private blockchain of this solution is used to perform consensus storage of three main business data: orders, logistics, and quality inspection. When new order and logistics data needs to be stored, nodes are randomly selected from a pre-set pool of authorized nodes, such as suppliers, logistics providers, and core enterprises, to form a verification committee for this consensus, based on the data type. Committee members need to verify their identity and data operation permissions through two authentication methods: X.509 digital certificates and attribute-based encryption. At the start of consensus, authorized nodes broadcast transaction packets with logistics order hash values ​​to the dynamic consensus group. The selected master nodes verify the format and signature validity of the transaction packets. After successful verification, a sequence number is assigned to each node and a pre-preparation message is broadcast. Then, each verification node enters an improved Byzantine fault-tolerant consensus process for multiple rounds of voting communication in the preparation and submission phases. When more than two-thirds of the nodes unanimously agree on the height and content of the same block and sign it, the consensus is successful, and the new block is synchronized to the ledgers of all members of the dynamic consensus group. The stored block generates an initially valid and updatable state marker, and an independent arbitration node handles appeals against the storage. For example, if the recipient disputes the quality inspection report's storage status, they can appeal to the arbitration node. After verification, the arbitration chain changes the status of the original storage block to "in dispute" or "expired" based on the ruling, achieving dynamic management throughout the storage lifecycle without altering the original storage content. The experimental results are shown in Table 11.

[0051] Table 11 Comparison of Consensus and Evidence Storage Performance

[0052] The results show that the parallel private chain evidence storage mechanism based on dynamic consensus groups and BFT can achieve efficient consensus and state management while ensuring data immutability, and can provide a secure and auditable trust infrastructure for multi-entity collaborative industrial scenarios.

[0053] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function in substantially the same way to achieve substantially the same result falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A method for intelligent recommendation of medical device internet information services, characterized in that, Includes the following steps: S(1), Multi-source data acquisition and preprocessing: The sensor array collects medical device operation data in real time and receives user operation data synchronously through the interactive interface module; after amplifying, filtering and converting the collected raw analog signals into digital signals, they are adaptively transmitted to the edge computing gateway via near-field communication, Bluetooth or Wi-Fi. S(2), Edge intelligent processing and secure transmission: The edge computing gateway cleans and standardizes the received data, and uses an embedded AI acceleration chip to perform local lightweight model inference to generate device status identification results; at the same time, it uses a cryptographic coprocessor to encrypt the model gradients, and transmits the encrypted data and identification results to the central data processing platform through a TLS secure channel. S(3), Knowledge Integration and Feature Learning: The central data processing platform integrates multi-source data and constructs a knowledge graph linking medical devices, institutions, and cases; it then uses a graph neural network computing unit to perform graph convolution calculations on the knowledge graph to generate deep feature representations for downstream optimization modules to read. S(4), Federated Learning and Model Optimization: By coordinating each edge node through the parameter server, aggregating their encrypted model gradients, and after verifying version consistency, distributing the updated global model parameters to each edge node, the model can be dynamically and hotly updated. S(5), Intelligent Recommendation and Digital Verification: Based on the deep feature representation, a mixed integer programming problem with clinical utility, cost-effectiveness, and equipment reliability as objectives is constructed, and the Pareto optimal solution set is solved; the feasibility of candidate recommendation schemes is verified by simulation and data-driven models using a digital twin. S(6), Result Output and Trusted Evidence Storage: The verified recommendation scheme is output to the user terminal, and the blockchain storage unit is used to record the key storage information of device status, model parameter updates and recommendation decisions through multiple parallel private chains to ensure the immutability and traceability of the data.

2. The method according to claim 1, characterized in that, In S(1), the sensor group includes a signal conditioning circuit (11), a first microprocessor (12), a vibration sensor (13), a temperature sensor (14), and an optical encoder (15). The signal conditioning circuit (11) includes an operational amplifier and a low-pass filter; the first microprocessor (12) adopts an ARM Cortex-M4 architecture and includes an analog-to-digital converter with 16-bit resolution; the vibration sensor (13) adopts a MEMS piezoelectric accelerometer; the temperature sensor (14) adopts a PT100 platinum resistance thermometer; and the optical encoder (15) adopts an incremental photoelectric encoder. The output terminal of the physical sensor is connected to the input terminal of the signal conditioning circuit (11) via a shielded cable, and the output terminal of the signal conditioning circuit (11) is connected to the analog-to-digital conversion pin of the first microprocessor (12) via PCB traces.

3. The method according to claim 1, characterized in that, In S(2), the lightweight model inference module includes an embedded AI acceleration chip (21), a model memory (22), and an encryption coprocessor (23). The lightweight model inference module operates through the following steps: S(21), Model loading and initialization: The embedded AI acceleration chip (21) loads the pruned and quantized neural network model from the model memory (22) into the chip's internal cache via the SPI interface; the encryption coprocessor (23) generates an asymmetric key pair and uploads it to the federated learning scheduling module for registration via the first secure communication module; S(22), Local Data Inference: The data preprocessing module transmits the cleaned and standardized sensor data to the input buffer of the embedded AI acceleration chip (21) via the PCIe bus; the embedded AI acceleration chip (21) calls the loaded neural network model to perform forward inference calculations to generate device status recognition results and local decision suggestions; S(23), Gradient calculation and encryption: The embedded AI acceleration chip (21) calculates the model gradient based on the difference between the local inference result and the real label using the backpropagation algorithm; the encryption coprocessor (23) uses the SM4 national cryptographic algorithm to encrypt the calculated model gradient and generate ciphertext gradient data. S(24), Secure transmission and parameter update: The first secure communication module uploads the encrypted gradient data to the gradient aggregator via a TLS secure channel; the lightweight model inference module periodically polls the parameter server to obtain the aggregated and updated global model parameters; after verifying the model version consistency, the model version manager distributes the new model parameters to each edge node via an encrypted channel. S(25), Model hot update: After reading the model parameters, the embedded AI acceleration chip (21) pauses inference and reads the new model parameters in the spare area of ​​the model memory (22). After reading, the embedded AI acceleration chip (21) returns to the new model to continue providing services and updating the model.

4. The method according to claim 1, characterized in that, In S(3), the process of feature learning performed by the graph neural network computing unit includes the following steps: S(31), Graph data preparation and parallel loading: The medical knowledge graph is obtained from the knowledge graph construction module, the topology is converted into an adjacency matrix, and the attribute data of each node in the graph is converted into a node feature matrix. Then, the adjacency matrix and the node feature matrix are partitioned and loaded into a multi-GPU-based parallel computing memory. S(32), Multi-layer graph convolution calculation and feature refinement: In a multi-layer graph convolutional network, the feature representations of all nodes are iteratively updated: for each target node in the current layer, the features of directly connected neighbor nodes are collected through an aggregation function to generate the neighbor context vector of the target node; the neighbor context vector is fused with the current layer feature vector of the target node itself, and the fused result is input into a fully connected layer with a non-linear activation function for linear transformation and non-linear mapping, and the higher-order feature representation of the node in the next layer is output. S(33), Deep Feature Output and High-Speed ​​Transmission: After completing the graph convolution calculation of the preset number of layers, the depth feature representation of all nodes in the final output is directly written into the shared memory region through NVLink high-speed interconnect technology, so that the multi-objective optimization solver can perform zero-latency reading and subsequent optimization solving.

5. The method according to claim 1, characterized in that, In S(4), the federated learning scheduling module includes a parameter server (41), a gradient aggregator (42), and a model version manager (43). The federated learning model optimization process introduces a dynamic attention weighting mechanism, which includes two steps: dynamic attention weight calculation and global model parameter update. The dynamic attention weights are calculated as follows: ; In formula (1), This represents the dynamic attention weight of the k-th edge node at time t; Use the Sigmoid activation function; This is the attention moderating factor; Let L2 norm be the gradient of the k-th edge node; Represents the global gradient norm; This represents the credibility score of the k-th node at time t; The global model parameters are updated as follows: ; In formula (2), This represents the global model parameters at iteration t; K is the total number of edge nodes. is the number of local training samples owned by the k-th edge node, and n is the total number of local training samples of all participating nodes; It is the adaptive learning rate of the k-th node; Represents the L2 regularization and adaptive L1 regularization coefficients; Represents the L2 regularization and adaptive L1 regularization coefficients; Let L represent the parameter tensor of the l-th layer of the neural network model; L is the total number of layers in the neural network model. This indicates that the private key SK is used to decrypt the encrypted content; This indicates that the input content is encrypted using the public key (PK).

6. The method according to claim 1, characterized in that, In S(5), the optimization problem constructed by the multi-objective optimization solver introduces spatiotemporal dynamic constraints, including knowledge graph relation reasoning and multi-objective optimization: Knowledge graph dynamic relationship reasoning uses a graph attention network, and node features are updated as follows: ; In formula (3), It is the original feature vector of node i; Let be the updated feature vector of node i; W is the weight matrix; It is the attention weight vector; it is the attention weight vector; LeakyReLU is the activation function; For activation functions; Represents the set of neighboring nodes of node i; j and k are the indices of the neighboring nodes; Indicates feature concatenation operation; Normalize attention weights; The multi-objective optimization problem is defined by the following formula: ; In formula (4), All are weighting coefficients; SNR is the signal-to-noise ratio; M is the total number of tasks, devices, or subsystems participating in the optimization. It is the dynamic weight of the j-th object at time t; It is the precision index of the j-th object; It is the time decay coefficient of the j-th object; The initial investment cost is T; T is the optimization time period. It is the discount rate; At any moment The maintenance cost function; Let be the conditional fault probability density function; It is the mean time between failures (MTBF) of the i-th device; It is the health status impact coefficient; This is the health status indicator of the i-th device; It represents the total carbon emissions over a given time period.

7. The method according to claim 1, characterized in that, In S(5), the simulation model based on physical mechanism adopts a multibody dynamics solver, and the real-time data-driven model adopts a long short-term memory neural network; the two are interconnected through a data fusion processor. The real-time data-driven model introduces a multi-head spatiotemporal attention mechanism, and its working process includes two parts: physical mechanism simulation and real-time state update. The dynamic process of the simulation model based on physical mechanisms is defined by the following set of differential equations: In formula (5), J is the moment of inertia. It is angular displacement; Angular acceleration; It is the torque constant; It is the driving current; B is the viscous damping coefficient; Angular velocity; It is a nonlinear frictional torque; It is the load torque; It is the observation output; It is the noise figure; Noise term; The state update of the real-time data-driven model is defined by the following formula: ; In formula (6), This represents the core state of the model at time t; H is the number of heads in the multi-head attention mechanism. Let h be the attention weight for the h-th head; It is element-wise multiplication; These are the forget gate parameters for the h-th head at time t; This represents the input gate parameters of the h-th head at time t; This represents the candidate state of the h-th head at time t; These are the weighting coefficients for the frequency domain features; S is the number of frequency bands; It is the weight of the s-th frequency band; It is the linear transformation matrix of the s-th frequency band; It is a Fast Fourier Transform; It is a vibration signal; It is a temperature signal; Other coding features; It is the frequency range of the s-th frequency band; These are the weighting coefficients for the residual connections; It is a residual connection function.

8. The method according to claim 1, characterized in that, In S(6), the trusted evidence storage using the blockchain evidence storage unit includes the following steps: S(61), Multi-source data classification and preprocessing: The key data from data collection, model optimization, and recommendation decision-making stages are classified; real-time equipment operating status data are classified as high-frequency lightweight data, model parameter update records generated by federated learning are classified as mid-frequency core asset data, and the decision-making logic and final results of the recommendation scheme are classified as low-frequency key decision data. S(62), Execution of Differentiated Hash Evidence Storage Strategy: For the high-frequency lightweight data, a timed polling hash strategy is executed to calculate the hash value of the state snapshot at preset time intervals; for the mid-frequency core asset data, an event-driven hash strategy is executed to calculate the hash value of the complete parameters only after the model version manager confirms that the global model update is successful; for the low-frequency key decision data, an instant trigger hash strategy is executed to immediately calculate the hash value of its decision path and result after the digital twin completes the scheme verification and approval output. S(63), Parallel Multi-Chain Classification, Evidence Storage and Association: The hash values, timestamps, and metadata of the above three types of data are submitted to three parallel private chains (41) for storage. At the same time, a cross-chain association index is established for the evidence records generated on different chains for the same business event to ensure the integrity of the global audit trail.

9. The method according to claim 8, characterized in that, The consensus and notarization process of the parallel private blockchain includes the following steps: S(631), Dynamic consensus group formation and node admission verification: Based on the type of evidence stored, nodes are dynamically selected from the pre-authorized node pool to form the verification committee for this consensus. Members of the verification committee must verify their identity and permissions through a dual authentication mechanism based on digital certificates and attribute-based cryptography. S(632), Improved Implementation of Byzantine Fault-Tolerant Consensus: Authorized nodes broadcast the evidence storage data transaction package to the dynamic consensus group; the master node verifies the validity and format compliance of the transaction package, and after the verification is passed, it assigns a sequence number and broadcasts a pre-preparation message; the verification committee member nodes conduct multiple rounds of communication voting, and when more than two-thirds of the nodes have collected valid signatures to agree, a consensus is reached; after the consensus is reached, the evidence storage data block is synchronized to all member nodes of this dynamic consensus group; S(633), Evidence Lifecycle Management: Add an updatable status marker to the newly generated evidence storage block. The initial status marker is valid. Set up an arbitration node to accept objection appeals. After the arbitration chain makes a ruling, the original evidence storage status marker can be modified to disputed or invalid. The evidence storage status is dynamically managed without modifying the original data.

Images