Cloud-based Medical Device Production Management System and Method

Abstract

This invention relates to the field of medical device manufacturing and management technology, specifically to a cloud-based medical device production management system and method. The method includes: collecting multi-source data and employing hash binding and the discrete logarithm problem to ensure data authenticity and accuracy; using temporal convolution and multi-head self-attention to fuse spatiotemporal features and generate production line state embedding vectors; constructing a flexible task strategy graph based on twin states, and dynamically adjusting the resource consumption, compliance risk, and stability deviation weights of task paths through multi-objective optimization; achieving visual traceability through flowchart modeling and multi-scale compliance annotation, locating the causal chain of violations, and providing feedback to adjust strategy weights; and finally generating adaptive control instructions using a graph attention network. This invention integrates multimodal fusion, dynamic optimization of the strategy graph, and causal inversion technology to form a closed loop of perception-optimization-traceability-regulation, collaboratively optimizing resource scheduling, compliance, and production stability, meeting practical application needs.

Description

Technical Field

[0001] This invention relates to the field of medical device manufacturing and management technology, specifically to a cloud-based medical device production management system and method. Background Technology

[0002] Against the backdrop of the accelerated digital and intelligent transformation of the medical device industry, achieving refined, flexible, and compliant management of the production process has become crucial for improving the quality control level and production efficiency of medical devices. In particular, when faced with practical problems such as parallel manufacturing of multiple product models, dynamic switching of complex process paths, and coupling of multiple factors such as personnel, equipment, and environment, traditional static management and single-point monitoring methods are no longer sufficient to meet the requirements of precise control.

[0003] Chinese invention patent application CN115730795A discloses a cloud-based medical device production management system, including a workshop control terminal, fiber optic lines, and a cloud platform. The workshop control terminal is connected to an operation display terminal, which is connected to a central processing unit. In this cloud-based medical device production management system, information input by staff onto the touch display terminal, along with data collected from monitors, CNC equipment, and quality inspection platforms regarding the scene, progress, and quality of medical device processing, are stored in the workshop control terminal. Simultaneously, staff can wirelessly access the cloud platform's data on the scene, progress, and quality of medical device production, as well as the operation display terminal's data on the 3D model, dimensions, and materials of the medical devices. This allows for real-time, synchronized tracking of the entire factory's medical device production process, avoiding problems caused by time zone differences.

[0004] Chinese invention patent CN118798494B discloses a data-based intelligent production control method and system for the garment industry. It establishes an intelligent data acquisition platform to collect heterogeneous data, extract common features, generate a comprehensive dataset, and construct a digital twin model of garment production. Based on the comprehensive dataset, it extracts data, generates triples, and constructs a garment production knowledge graph. It then builds a garment production data analysis model, extracts key features, identifies implicit correlations, generates a cross-modal knowledge graph, identifies garment production control rules, and solves for garment production control parameters. Finally, it constructs a decision evaluation model to parse the garment production control parameters into production control decisions, evaluates them, dynamically corrects them based on the evaluation results, obtains optimized production control decisions, distributes them to equipment terminals, and adaptively optimizes garment production. It collects real-time production data and updates the garment production digital twin model and garment production knowledge graph, periodically optimizing until the production task is completed.

[0005] However, the production process exhibits complex characteristics such as multi-constraint coupling, multi-stage collaboration, and multi-state evolution. Especially in scenarios with multiple models on mixed production lines, it is necessary to balance resource utilization, quality compliance, and production stability. At the same time, strict regulatory requirements pose challenges to the traceability of the entire process and the efficiency of anomaly location. Furthermore, with the development of emerging technologies such as cloud computing, digital twins, graph intelligence, and reinforcement learning, their integration into medical device production management is expected to achieve real-time perception of multi-source states, intelligent optimization of task scheduling, and full-process traceability and feedback of compliant behaviors. This will provide technical support for building an efficient, safe, and traceable medical device production system. There is an urgent need to build a cloud-based multi-source data fusion, virtual-real linkage, and intelligent decision-making system to achieve dynamic collaborative management of the entire production chain. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the background technology by proposing a cloud-based medical device production management system and method.

[0007] The technical solution of this invention: a medical device production management method based on a cloud platform, comprising the following specific implementation steps:

[0008] S1. Collect production status data from the medical device production line, and after format conversion and time synchronization normalization, form a multi-source status input tensor. After converting to binary, generate data fidelity evaluation parameters by hash function and discrete logarithm calculation.

[0009] S2. Detect the fidelity of multi-source data, and use temporal convolution and multi-head self-attention architecture to generate spatiotemporal fusion state embedding vectors. Then, map them to the virtual production line twin state output through hierarchical kernel functions, and introduce LSTM to predict the twin mapping production line state.

[0010] S3. Construct a task dependency graph based on twin states, define a flexible strategy graph to dynamically adjust edge weights considering resource consumption, compliance risks and state deviations, and use a strategy gradient optimization algorithm to optimize task paths to balance resource utilization, quality compliance and risks. During execution, trigger feedback reconstruction strategies based on boundary trigger functions.

[0011] S4. Construct a compliance behavior graph based on the executed task path, automatically match the node status of the three compliance dimensions, initiate causal chain reconstruction to locate the upstream link for abnormal nodes, visualize the violation path and responsible person through Sankey graph hot zone highlighting, and provide feedback to adjust the prediction weight of the twin model and the preference of the strategy graph.

[0012] S5 collects real-time data on equipment status, environmental parameters, personnel behavior, and boundary signals, models them as state flow vectors, dynamically adjusts the edge weights of the strategy graph through compliance anomaly feedback, generates the optimal scheduling path and multi-dimensional control instruction set based on the control strategy generation function, and feeds back to the twin model and strategy graph to form a reinforcement learning closed-loop control.

[0013] Preferably, the data fidelity assessment parameter generation process is as follows:

[0014] Input the multi-source situation into tensor X (0) Convert to binary data BX, randomly select a one-time integer s∈[1,q-1], and calculate the controllable evaluation parameter r=g. s (mod p);

[0015] Where p is a predefined large prime number that satisfies the discrete logarithm problem difficulty; q is a predefined large prime number that satisfies q|(p-1), i.e., q divides p-1; g is the multiplicative cyclic group Z. p * Generators;

[0016] Calculate the first-order data fidelity evaluation parameter Of = H(BX||r)(mod q);

[0017] Where H is a predefined hash function that outputs a length matching the number of bits q; || represents a concatenation operation;

[0018] The calculation order data fidelity evaluation parameter Os = (s - Of × x) (mod q);

[0019] Where x is a predefined auxiliary production generation code, i.e. a randomly selected integer, x∈[1,q-1];

[0020] Output data fidelity evaluation parameters OP = {Of, Os}.

[0021] The preferred method for detecting the fidelity of multi-source data is as follows:

[0022] Obtain the data fidelity assessment parameters OP = {Of, Os} and the multi-source situational input tensor X. (0) Extract the first-order data fidelity evaluation parameter Of and the second-order data fidelity evaluation parameter Os from it, and input the multi-source situation tensor X. (0) Convert to binary data B2X;

[0023] Calculate the auxiliary detection parameter Ap = g Os ×y Of (mod p);

[0024] Where y is a predefined auxiliary production inspection code, y = g x (mod p);

[0025] Calculate the data fidelity Dc = H(B²X||AP)(mod q);

[0026] If Dc = Of, it means that the multi-source situational input tensor X has been obtained. (0)The authenticity and accuracy of the information must be verified; otherwise, an immediate warning will be issued.

[0027] Preferably, the spatiotemporal fusion state embedding vector generation process is as follows:

[0028] Input a multi-source situation tensor;

[0029] A multi-layer one-dimensional convolutional network is used to slide convolution along the time dimension by convolutional kernel weights, and the input features are passed layer by layer by combining bias terms and applying the ReLU activation function to output the convolutional layer output.

[0030] The output of the convolutional layer is input into the multi-head self-attention module, the multi-head is split and independently computed to calculate the long-range dependencies across time and space, and then connected to the projection to generate fused features;

[0031] Global average pooling is used to compress and fuse features along the time axis to obtain pooled output features. Then, a production status vector is output by combining weights and biases through a fully connected layer and nonlinear transformation, which is the spatiotemporal fusion state embedding vector.

[0032] Preferably, the virtual production line twin state output process is as follows: the input situation vector is mapped to the virtual state output in layers through three types of kernel functions: the physical perception kernel uses RBF to capture the static dependence of the equipment; the process flow kernel uses MLP modeling to dynamically transfer; the semantic structure kernel uses GCN to reflect soft constraint control; and Gaussian perturbation is superimposed to simulate external interference.

[0033] Preferably, the implementation process of the feedback reconstruction strategy triggered by the boundary triggering function during execution is as follows:

[0034] Based on the virtual production line twin state, the task structure and process dependency relationship are extracted, a task set is constructed and resource requirements, compliance requirements and quality indicators are defined, and mapped as directed graph nodes and dependency edges;

[0035] Based on the production status vector and twin mapping of production line status, a flexible strategy graph is defined to map the current production line status to strategy graph nodes, and the dynamic weight of the flexible strategy graph is defined to comprehensively consider resource consumption, compliance risks and status deviation.

[0036] The task path is optimized on the policy graph using a policy gradient optimization algorithm combined with a multi-path reward reconstruction algorithm. The objective function is:

[0037] Where J(π) represents the expected total reward of path π; L represents the task path length; U k Q represents the improvement in resource utilization after executing the k-th task; k R represents the positive contribution factor of the k-th task to overall quality compliance; k β1, β2, and β3 represent the risk factors resulting from the execution of the k-th task; β1, β2, and β3 represent the weighting coefficients of the reward function; E π(·) represents the expected reward function;

[0038] Path π * (t) is issued, parallel execution is initiated, and a boundary trigger function θ is defined. k (t):

[0039]

[0040] Where, θ k (t) represents task T k Does the out-of-bounds control function trigger at time point t? Indicates the actual execution time of task k; ε represents the expected execution time of task k in the current policy path; k This represents the maximum allowable time error range for the k-th task;

[0041] onceθ k If (t) = 1, the task will be interrupted and a policy refactoring will be triggered.

[0042] Preferably, the process of locating the upstream link by initiating causal chain reconstruction for abnormal nodes is as follows:

[0043] A compliance behavior graph is constructed based on the executed task paths. The node set N represents production operations, the relationship set R represents the relationships between nodes, and the attribute set A represents node attributes. This graph is used to represent the structure of task and compliance information.

[0044] For each node in the graph, the system automatically matches compliance clauses from the regulatory library and marks compliance dimensions including time window, operating environment, and personnel qualifications.

[0045] For nodes in the completed behavior sequence that have been marked as abnormal in compliance, initiate multi-source causal chain reconstruction, locate the upstream link that caused the abnormality, and define the abnormal causal chain as follows:

[0046] C abnormal ={n i →n j →…→n k |θ j (t)=1};

[0047] Among them, C abnormal This represents an abnormal causal chain, that is, starting from a certain historical task n. i The path from the start of the process until the exception is triggered; n i →n j →…→n k This represents a sequence of tasks on an exception chain, arranged in the order of task execution, indicating a potential causal relationship; θ j (t) represents the task. j Does the out-of-bounds behavior trigger at time t?

[0048] Preferably, the control policy generation function takes the current state vector and the updated policy graph as input, generates task path scores through a graph neural network, and selects the path with the highest score to execute control.

[0049] Preferably, multi-layer one-dimensional convolutional networks include:

[0050] First layer: kernel size is 5×5, number of input channels is C, number of output channels is 64, stride is 1, padding is 2, activation function is ReLU;

[0051] Second layer: Convolution kernel size is 3×3, number of input channels is 64, number of output channels is 128, stride is 1, padding is 1, activation function is ReLU;

[0052] The third layer has a kernel size of 3×3, 128 input channels, 256 output channels, a stride of 1, padding of 1, and an activation function of ReLU.

[0053] The technical solution of this invention: A cloud-based medical device production management system, used to execute the aforementioned cloud-based medical device production management method, comprising:

[0054] The data acquisition module is used to uniformly collect multi-source heterogeneous information involved in the medical device production process, trigger events, and perform structured modeling, and output multi-source situational input tensors.

[0055] The twin state generation module is deployed on the cloud platform. It ensures the reliability of input through data fidelity verification, uses temporal convolution and multi-head attention to fuse spatiotemporal features, generates virtual states through hierarchical mapping of physical perception kernel, process flow kernel, and semantic structure kernel, and introduces LSTM prediction mechanism to realize dynamic updating of the twin, thus constructing a digital twin modeling framework that synchronizes virtual and real.

[0056] The flexible strategy orchestration module, deployed on the cloud platform, constructs a task dependency graph structure based on the aforementioned state vectors. By introducing historical execution deviations, violation feedback factors, and dynamic resource weights, it generates a flexible task strategy graph and dynamically adjusts the task path structure in real time, supporting multi-state driven production scheduling and task switching control.

[0057] The compliance traceability analysis module, deployed on a cloud platform, is used to map various process data, task execution trajectories, and equipment and personnel operation sequences into compliance traceability maps during production execution, enabling structured representation of potential violations, identification of trigger events, and causal chain analysis.

[0058] The closed-loop collaborative control module is used to integrate the adjusted flexible task strategy diagram with the current system state, generate a multi-dimensional control instruction set through the control strategy function, and use a multi-objective optimization function to dynamically balance and optimize the control effect, and finally issue the strategy to the field control system.

[0059] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0060] This invention designs a cloud-based medical device production management system and method. It achieves dynamic synchronization between virtual and real environments through a multi-layered modeling framework, integrating multimodal perception, convolutional attention collaborative modeling, and kernel function mapping to accurately capture the spatiotemporal evolution of the production line, providing a high-confidence decision-making basis for task scheduling.

[0061] Based on a dynamic orchestration mechanism using a flexible strategy graph, and combined with multi-objective optimization and compliance feedback to adjust task paths in real time, the resource utilization and compliance of mixed-line production of multiple product models are significantly improved, and the risk of violations is reduced.

[0062] Construct a visualized compliance traceability chain and closed-loop control framework. Locate the source of violations through flowchart modeling and causal inversion, realize responsibility traceability and dynamic optimization of the entire production process, shorten the response time to anomalies, and drive the adaptive adjustment of equipment parameters and process scheduling to ensure the stability and quality consistency of medical device production.

[0063] Controllable evaluation parameters and hash bindings are constructed based on large prime numbers to ensure the authenticity and readiness of the original multimodal data (such as temperature and image streams); data non-repudiation and source traceability are achieved through the strong correlation between the order parameters and predefined parameters; during the verification phase, auxiliary detection parameters are used to quickly verify and intercept suspicious data (such as sensor signal forgery and operation log tampering), avoiding erroneous data input into the digital twin model that could lead to scheduling decision deviations. Attached Figure Description

[0064] Figure 1 This is a flowchart of a cloud-based medical device production management method proposed in this invention. Detailed Implementation

[0065] Example 1, as Figure 1 As shown, the present invention proposes a cloud platform-based medical device production management method, which includes the following specific implementation steps:

[0066] S1. Access multimodal sensing data (production status data) from various nodes of the production line, including but not limited to:

[0067] Structured data: Sensor signals (including but not limited to temperature T, pressure P, vibration A, current I, humidity H);

[0068] Unstructured data includes, but is not limited to, image streams (machine vision), voice commands (operator voice interaction), and log text;

[0069] Next, the multimodal sensing data undergoes format conversion and time synchronization to unify structured and unstructured data into a standard format suitable for computer processing. Then, normalization preprocessing is employed to obtain normalized multimodal production line situational data (i.e., multi-source situational input tensor) X. (0) ;

[0070] And provide the multi-source situation input tensor X (0) The data fidelity evaluation parameters are generated, and the generation process is as follows:

[0071] (1) Input the multi-source situation tensor X (0) Convert to binary data BX, randomly select a one-time integer s∈[1,q-1], and calculate the controllable evaluation parameter r=g. s (mod p);

[0072] Where p is a predefined large prime number (1024 bits in this embodiment) that satisfies the difficulty of the discrete logarithm problem; q is a predefined large prime number (256 bits in this embodiment) that satisfies q|(p-1), i.e., q divides p-1; g is a multiplicative cyclic group. Generators;

[0073] (2) Calculate the first-order data fidelity evaluation parameter Of = H(BX||r)(mod q);

[0074] Where H is a predefined hash function (SHA-256 is used in this embodiment), and the output length matches the number of bits q (i.e., 256 bits); || represents the concatenation operation;

[0075] (3) Calculate the order data fidelity evaluation parameter Os = (s - Of × x) (mod q);

[0076] Where x is a predefined auxiliary production generation code, i.e. a randomly selected integer, x∈[1,q-1];

[0077] (4) Output data fidelity evaluation parameters OP = {Of, Os};

[0078] S2. Construct a multi-level modeling framework that integrates multimodal perception, convolutional attention collaborative modeling, kernel function mapping, and twin prediction update. Based on a multimodal situation embedding mechanism, it unifies multi-source data and introduces a kernel mapping-type digital twin modeling structure. Furthermore, it constructs a twin state prediction-update mechanism to dynamically synchronize the virtual and real models with production evolution. The specific implementation process is as follows:

[0079] S21. Detect the multi-source situational input tensor X(0) The data fidelity is tested as follows:

[0080] (1) Obtain the data fidelity evaluation parameters OP = {Of, Os} and the multi-source situation input tensor X. (0) Extract the first-order data fidelity evaluation parameter Of and the second-order data fidelity evaluation parameter Os from it, and input the multi-source situation tensor X. (0) Convert to binary data B2X;

[0081] (2) Calculate the auxiliary detection parameter Ap = g Os ×y Of (mod p);

[0082] Where y is a predefined auxiliary production inspection code, y = g x (mod p);

[0083] (3) Calculate the data fidelity Dc = H(B2X||AP)(mod q);

[0084] (4) If Dc = Of, it means that the multi-source situation input tensor X has been obtained. (0) The authenticity and accuracy of the information must be verified; otherwise, an immediate warning will be issued.

[0085] S22. The normalized multi-source situation input tensor X output from step S1. (0) A spatiotemporal fusion modeling architecture combining temporal convolution and multi-head self-attention is constructed to deeply capture the temporal evolution features and spatial modal collaboration in the production environment, thereby generating production line state embedding vectors containing rich spatiotemporal information, specifically:

[0086] A1. Input Tensor M represents the number of modalities (including but not limited to temperature, pressure, and image features); L represents the time series length (sampling window length); and C represents the feature dimension of each modality.

[0087] A2. A multi-layer one-dimensional convolutional network (1D-CNN) is used to perform convolution operations on the time dimension to extract local temporal features, defined as: H (l) =ReLU(W (l) *H (l-1) +b (l) );

[0088] Among them, W (l) The * denotes the weights of the l-th layer 1D convolution kernel; * indicates the convolution operation, sliding along the time dimension; b (l) H represents the level l bias term; (l-1) H represents the output feature of the (l-1)th layer; (l)Represents the output features of the l-th layer; ReLU(·) represents the ReLU activation function;

[0089] It should be noted that the network structure of a multi-layer one-dimensional convolutional network (1D-CNN) is as follows: A multi-layer one-dimensional convolutional network (1D-CNN) consists of three layers of convolutional neural networks:

[0090] The first layer (Conv1) has a kernel size of 5×5, C input channels, 64 output channels, a stride of 1, padding of 2, and ReLU activation function.

[0091] The second layer (Conv2) has a kernel size of 3×3, 64 input channels, 128 output channels, a stride of 1, padding of 1, and an activation function of ReLU.

[0092] The third layer (Conv3) has a kernel size of 3×3, 128 input channels, 256 output channels, a stride of 1, padding of 1, and an activation function of ReLU.

[0093] A3, Convolutional layer output H (L) As input to the Multi-Head Self-Attention (MHSA) module, it captures long-range dependencies across time and modality. MHSA is defined as follows:

[0094]

[0095] Where Q, K, and V represent the query, key, and value matrix in the attention mechanism, and are derived from H. (L) Multiplying by the projection matrix respectively yields the attention weights and aggregated features; Attention(Q,K,V) represents standard scaled dot product attention; d k W represents the feature dimension within a single attention head. Q W K W V These represent the functions used in multi-head self-attention to select H. (L) A trainable matrix projected onto the query, key, and value space;

[0096] Specifically, input H (L) The sequence is expanded according to modality and time; it is split into h heads through a multi-head mechanism, and each head independently calculates attention; the attention outputs of each head are connected and projected back to the original dimension to obtain the fused feature representation Z;

[0097] A4. To obtain the overall state representation of the time window, a pooling operation (global average pooling in this embodiment) is used to converge along the time axis: S t =Pooling(Z); then pool the output feature S t Mapping to the final state embedding space via a fully connected layer:

[0098]

[0099] Where Pooling(·) represents the pooling function, which pools Z into a fixed length along the time dimension; W represents the generated spatiotemporal fusion state embedding vector, i.e., the production situation vector; s b represents the state mapping weights of the fully connected network; s The state mapping is indicated by the fully connected bias; Flatten(·) indicates the fully connected operation;

[0100] S23, Set the situation vector The input is fed into the digital twin state estimation model to form a virtual working condition scenario, specifically:

[0101] The situation vector As input, a set of learnable kernel function combinations are used to map to various virtual production line state outputs:

[0102] Among them, D t φ represents the virtual state output vector of the production line. k (·) represents the k-th kernel mapping function, i.e., the nonlinear characteristic transformation from the state to the twin index; ω k ε represents the kernel function weights, reflecting their importance in the mapping (through adaptive learning during training); K represents the number of kernel functions; ε t This represents a Gaussian perturbation, used to simulate unobserved external disturbances and system errors;

[0103] It should be noted that, to enhance the flexibility and interpretability of the mapping, three types of kernel functions are introduced and applied hierarchically to state variables of different dimensions:

[0104] The physical sensing kernel captures the static physical dependencies between device states (such as the relationship between temperature and pressure) using radial basis functions (RBF).

[0105] The process flow core models the dynamic transmission between production processes (such as time lag between multiple workstations), and adopts a multilayer perceptron (MLP) with a nested relationship between nonlinear activation and parameter learning.

[0106] The semantic structure kernel reflects the regulation of state transitions by soft constraints such as management / human factors (e.g., changes in management strategies). It adopts a graph convolution kernel (GCN) and uses the graph structure to connect device nodes to model the impact of human-machine collaborative states.

[0107] S24. To maintain the reliable evolutionary capability of the digital twin over time, a state update and prediction mechanism can be introduced:

[0108] in, Indicates the predicted twin state at the next moment; u t This represents the current control input / external disturbance information (including but not limited to sudden changes in temperature and humidity, and changes in order structure); LSTM(·) represents a Long Short-Term Memory (LSTM) network;

[0109] S3. Construct a flexible task strategy graph-driven production orchestration framework, introducing flexible strategy graphs, dynamic situation mapping, multi-objective strategy optimization, and path execution monitoring and feedback to generate and execute dynamic task flows under different product models, process paths, and resource states. Specifically:

[0110] S31. Combining the current digital twin state D in step S2 t Extract the task structure and process dependencies of the target medical device product, and construct the directed graph structure G = (V, E) between the task set Task and its relationship.

[0111] Specifically, Task = {Task i =(s i ,r i ,q i |i=1,2,...,n};

[0112] Among them, Task i This represents the i-th production task; n represents the total number of production tasks; s i This represents the initial resources required for the i-th task (including but not limited to machine model and manpower level); r i This represents the compliance requirements corresponding to the i-th task (including but not limited to disinfection duration and humidity range); q i V represents the quality level index of the i-th task; V represents the set of nodes, V = Task; E represents the set of relational attributes, i.e. the process dependencies between tasks, such as "packaging must be performed after sterilization";

[0113] S32. Based on the production status vector obtained in step S2 With digital twin prediction The current production line state is mapped to strategy graph nodes to achieve dynamic adjustability of the task flow graph. A flexible strategy graph G is defined. f for:

[0114]

[0115] Where W represents the set of edge weights in the graph, reflecting the overall cost under different paths; w ij Represents the Task node i →Task j State dynamic weights; c ij Represents node i(Task) i ) to node j(Task j The required unit resource consumption (including but not limited to machine occupancy rate and personnel ratio); r ij Represents node i(Task) i ) to node j(Task j The degree of risk of the switch affecting the compliance chain (including but not limited to the risk of the aseptic chain breaking due to the switch); δ ij α1, α2, and α3 represent the deviation between the current state vector and the desired state of the target task (reflecting operational stability); α1, α2, and α3 represent the weighting coefficients obtained through learning and optimization, satisfying α1 + α2 + α3 = 1.

[0116] S33. Employ an improved policy gradient optimization algorithm (PPO + multi-path reward reconstruction) in the policy graph G. f Solving for the current optimal task path π * (t), define the objective function J(π) for optimizing the task path:

[0117]

[0118] Where J(π) represents the expected total reward of path π, used to measure scheduling quality; L represents the task path length; U k Q represents the improvement in resource utilization (e.g., change in machine utilization) after executing the k-th task; k R represents the positive contribution factor of the k-th task to overall quality compliance; k β1 represents the risk factors resulting from the execution of the k-th task (including but not limited to cross-contamination and packaging before cooling); β2, β3 represent the weighting coefficients of the reward function, controlling the resource, quality, and risk trade-offs respectively; E π (·) represents the expected reward function;

[0119] S34, Path π * (t) is issued, parallel execution is initiated, and a boundary trigger function θ is defined. k (t):

[0120]

[0121] Where, θ k (t) represents task Tk Does the out-of-bounds control function trigger at time point t? Indicates the actual execution time of task k; ε represents the expected execution time of task k in the current policy path; k This represents the maximum allowable time error range for the k-th task;

[0122] onceθ k When (t) = 1, the feedback will be sent to the twin prediction model in step S2 and the compliance traceability in step S4, while interrupting the task and triggering policy reconstruction.

[0123] S4. By using process state graph modeling, multi-scale behavior annotation, and compliance causal chain inversion, a visual, interactive, and traceable compliance traceability chain is constructed. The specific implementation process is as follows:

[0124] S41. Based on the task path π that has been completed in step S3 * (t), construct the compliance behavior graph G corresponding to this path. c ,

[0125] Among them, G c The Compliance Graph represents the structure of tasks and compliance information in the medical device manufacturing process; N represents the set of nodes in the graph, where each node is n. ij R represents a specific production operation or task, including but not limited to "aseptic labeling" and "environmental handover"; R represents the set of relationships between nodes, r ij A represents the relationship between nodes i and j; A represents the set of node attributes, a ij Represents the attributes of any node;

[0126] S42. For each node n in the graph... i The system matches compliance clauses with national standards, industry norms, and other regulatory knowledge bases to automatically generate a compliance label sequence L. i And establish three types of compliance dimension mappings:

[0127] Dimension 1, Time Compliance, i.e., whether the task is executed within the specified window;

[0128] Dimension two, spatial compliance, that is, whether the operation is carried out in the prescribed environment;

[0129] Dimension three, personnel compliance, i.e., whether the operator's qualifications meet the standards;

[0130] S43. For nodes in the completed behavior sequence that are marked as non-compliant (i.e., any state is "non-compliant"), initiate the multi-source causal chain reconstruction module to locate the upstream link that caused the anomaly and define the anomaly causal chain as follows:

[0131] C abnormal ={n i →n j →…→n k |θ j (t)=1};

[0132] Among them, C abnormal This represents an abnormal causal chain, that is, starting from a certain historical task n. i The path from the start of the process until the exception is triggered; n i →n j →…→n k This represents a sequence of tasks on an exception chain, arranged in the order of task execution, indicating a potential causal relationship; θ j (t) represents the task. j Does the event trigger an out-of-bounds behavior (including but not limited to delays or unauthorized environments) at time t?

[0133] Specifically, the node chain represents the sequence of operations, and an out-of-bounds signal θ occurs in an intermediate node. k (t) = 1, triggering retrospection;

[0134] S44. Based on the traceability results, construct a visual compliance chain feedback module for management personnel to view:

[0135] Compliance Chain View: Displays the task execution path and its compliance status evolution in the form of a Sankey diagram;

[0136] Abnormal hotspot highlighting: Abnormal nodes are marked in red in the compliance chain to show the type of non-compliance;

[0137] Responsibility path location: Automatically locate upstream behaviors and responsible persons / equipment that are strongly correlated with anomalies.

[0138] Simultaneously, the feedback is sent to the digital twin model in step S1 to adjust its compliance factor prediction weights; it is also sent to the flexible strategy graph in step S3 to adjust the future scheduling path weight preferences.

[0139] S5. Construct a digital twin system + multi-source data fusion controller + compliance feedback coupling framework to perform adaptive optimization and dynamic closed-loop control of production behavior. The specific implementation process is as follows:

[0140] S51. Relying on the cloud platform, collect the following multi-source status data in real time:

[0141] Device status data (Including but not limited to temperature, vibration, and current);

[0142] Environmental parameters (Including but not limited to cleanliness, humidity, and pressure difference);

[0143] Operator behavior trajectory and identity authentication information;

[0144] Task execution log and boundary out-of-bounds signal θ k (t);

[0145] Modeling multidimensional data as state flow vectors:

[0146] S52. Based on the compliance anomaly feedback results in step S4 (such as the anomaly node set C), abnormal (Illegal paths) are used as input signals to the policy graph G. f The execution weight of the corresponding node is dynamically adjusted, and the specific update method is as follows:

[0147] Among them, w' ij This represents the updated edge weights of the strategy graph, which are adjusted to guide the next round of task scheduling; w ij This represents the process from node n in the original flexible strategy graph. i to n j The right to the side; λ represents the intensity of compliance deviations between nodes in historical scheduling (including but not limited to violation frequency and violation severity); λ represents the learning step size factor (adjustment magnitude) of edge weight adjustment;

[0148] S53, Based on the updated strategy graph G' f and the current state flow S t Functions generated using control strategies: C t =f(G' f ,S t Output multi-dimensional control instruction set C t The generated instructions will then be distributed to the local machine via the cloud platform;

[0149] Among them, C t This represents the set of control commands that need to be issued at the current moment; f(·) represents the control policy generation function, which merges the state and policy graph to deduce the control behavior.

[0150] Specifically, the control strategy generation function f(·) can be divided into the following four processing units:

[0151] (1) State graph fusion layer: The real-time state vector S t With the adjusted policy graph G' f Fusion encoding is performed, and a graph attention network (GAT) mechanism is used to inject state features into the node representation:

[0152] Where N(i) represents the set of adjacent nodes of node i; This represents the attention weight of node i to its neighboring node j; W represents a local state slice representing the state of node i; (l) This represents the trainable parameters of the l-th layer; Let represent the hidden state vector of node i in the l-th layer; This represents the output hidden state vector of node i in the (l+1)th layer; [·||·] represents the vector concatenation operation, which connects the node's own features with the state features to form an enhanced input vector;

[0153] (2) Based on the graph fusion representation from the previous step, the task paths are scored, and the optimal scheduling path π that meets the current state is selected. (*) (t):

[0154]

[0155] Select the path with the highest score:

[0156] Where Score(π) represents the overall score of candidate task path π; γ represents the adjusted edge weights between nodes; ij (S t ) represents the adaptability factor for task transition from i to j in the current state; (i,j)∈π represents the edge from node i to j in the path;

[0157] (3) For strategy path π (*) For each node and edge in (t), a corresponding control signal is generated.

[0158] Generate control signals Including but not limited to:

[0159] Equipment control commands (such as temperature control values, voltage, and speed);

[0160] Environmental control commands (such as damper adjustment, differential pressure increase);

[0161] Operator instructions (such as unlocking the workstation, prompts on the human-machine interface);

[0162] Scheduling strategies (such as skipping nodes, postponing processes, etc.)

[0163] S54. After all execution behaviors (device status changes, task completion status, boundary triggers) are perceived, they will be uniformly fed back to the digital twin compliance prediction model and strategy graph adjustment model to form a continuously evolving control closed loop. At the same time, the following reinforcement learning update mechanism will be established:

[0164] Where, π (*)(t+1) represents the policy path after the next iteration optimization; π (*) (t) represents the optimal task execution path strategy at the current time t; η represents the strategy path learning rate, which controls the rate of path change. This represents the gradient indicating the direction of policy performance improvement.

[0165] Example 2: The present invention proposes a cloud-based medical device production management system, which is used to execute the cloud-based medical device production management method proposed in Example 1, including: a data acquisition module, a twin state generation module, a flexible strategy orchestration module, a compliance traceability analysis module, and a closed-loop collaborative control module.

[0166] The data acquisition module is used to uniformly collect, event-driven trigger, and structured model multi-source heterogeneous information involved in the medical device production process, such as equipment operation data, operator behavior information, production environment parameters, and process task status, and output multi-source situation input tensors.

[0167] The twin state generation module is deployed on the cloud platform. It ensures the reliability of input through data fidelity verification, uses temporal convolution and multi-head attention to fuse spatiotemporal features, generates virtual states through hierarchical mapping of physical perception kernel, process flow kernel, and semantic structure kernel, and introduces LSTM prediction mechanism to realize dynamic updating of the twin, thus constructing a digital twin modeling framework that synchronizes virtual and real.

[0168] The flexible strategy orchestration module, deployed on the cloud platform, constructs a task dependency graph structure based on the aforementioned state vectors. By introducing historical execution deviations, violation feedback factors, and dynamic resource weights, it generates a flexible task strategy graph and dynamically adjusts the task path structure in real time, supporting multi-state driven production scheduling and task switching control.

[0169] The compliance traceability analysis module, deployed on the cloud platform, is used to map various process data, task execution trajectories, and equipment and personnel operation sequences into compliance traceability maps during production execution. This enables the structured representation of potential violations, the identification of triggering events, and the analysis of causal chains, forming a visualized reverse investigation and dynamic compliance assessment capability for the entire process.

[0170] The closed-loop collaborative control module is used to integrate the adjusted flexible task strategy diagram with the current system state. It generates a multi-dimensional control instruction set through the control strategy function, covering equipment control parameters, environmental adjustment strategies and personnel guidance tasks. It also uses a multi-objective optimization function to dynamically balance and optimize the control effect. Finally, it sends the strategy to the field control system through the cloud platform to complete the intelligent control, risk avoidance and efficiency improvement / performance optimization of the entire medical device production process.

[0171] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A cloud platform-based medical instrument production management method, characterized by, The specific implementation steps include the following: S1. Collect production state data from the medical device production line, perform time synchronization and normalization processing through format conversion to form a multi-source situation input tensor, and convert it into binary form through a hash function and discrete logarithm calculation to generate data fidelity evaluation parameters; S2. Detect the fidelity of the multi-source data, generate a spatio-temporal fusion state embedding vector using time series convolution and multi-head self-attention architecture, and then map it to a virtual production line twin state output through a hierarchical kernel function, and introduce LSTM for prediction of the twin mapping production line state; Wherein, the virtual production line twin state output process is as follows: the input situation vector is mapped to the virtual state output through three types of kernel functions: the physical perception kernel uses RBF to capture device static dependencies; the process flow kernel uses MLP to model dynamic transmission; the semantic structure kernel uses GCN to reflect soft constraint regulation; superimpose Gaussian disturbance to simulate external interference; S3. Based on the twin state, construct a task dependency graph, define a flexible strategy graph, and dynamically adjust its edge weights to consider resource consumption, compliance risk and state deviation, use a policy gradient optimization algorithm to optimize the task path to balance resource utilization, quality compliance and risk, and trigger feedback reconstruction strategy based on boundary trigger function during execution, the implementation process is as follows: Based on the virtual production line twin state, extract the task structure and process dependency relationship, construct a task set and define resource demand, compliance requirement and quality index, and map them to directed graph nodes and dependent edges; Based on the production situation vector and the twin mapping production line state, define a flexible strategy graph, map the current production line state to the strategy graph nodes, and dynamically adjust the edge weights of the flexible strategy graph to consider resource consumption, compliance risk and state deviation; The strategy gradient optimization algorithm is combined with a multi-path reward reconstruction algorithm to optimize a task path on a strategy graph, and a target function is: ; wherein, represents the expected total reward of a path π; L represents the task path length; U k represents the improvement of resource utilization after the execution of the kth task; Q k represents the positive contribution factor of the kth task to the overall quality compliance; R k represents the risk factor caused by the execution of the kth task; , , represents the reward function weighting coefficient; represents the expected reward function; path issuing, starting parallel execution, and defining boundary trigger functions : ； wherein, represents a task T k whether to trigger the out-of-bound control at time point t; represents the actual execution time of task k; represents the expected execution time of task k in the current policy path; represents the maximum time error range allowed for the kth task; Once the task is interrupted and a policy reconfiguration is triggered; S4. Based on the executed task path, build a compliance behavior graph, automatically match three types of compliance dimension label node states, start the causal chain reconstruction positioning of the upstream link for abnormal nodes, highlight the violation path and responsible person through Sankey diagram hot area visualization, and feedback to adjust the prediction weight of the twin model and the preference of the strategy graph; S5. Real-time collection of device state, environmental parameters, personnel behavior and boundary signals, modeling as a state flow vector, dynamically adjusting the edge weight of the strategy graph through compliance anomaly feedback, generating optimal scheduling path and multi-dimensional control instruction set based on control strategy generation function, and feeding back to the twin model and strategy graph after execution to form a reinforcement learning closed-loop control. 2.The cloud platform-based medical instrument production management method of claim 1, wherein, The data fidelity evaluation parameter generation process is as follows: Multiple-source situation input tensor Transformed into binary data BX, randomly select a one-time integer s ∈ [1, q-1], calculate the controllable evaluation parameter r = g s (mod p); wherein p is a predefined large prime number, satisfying a discrete logarithm problem is difficult; q is a predefined large prime number, and satisfies q | (p-1), that is, q divides p−1; g is a generator of a multiplicative cyclic group of prime order Calculate the first-order data fidelity evaluation parameter Of=H(BX||r) (mod q); Wherein, H is a pre-defined hash function, the output length matches the number of bits of q; || represents the concatenation operation; Calculate the secondary data fidelity evaluation parameter Os=(s-Of×x) (mod q); Wherein, x is a pre-defined auxiliary production generation code, that is, a randomly selected integer, x∈[1,q-1]; Output the data fidelity evaluation parameter OP={Of,Os}. 3.The cloud platform-based medical instrument production management method of claim 2, wherein, The detection process of the multi-source data fidelity is as follows: Obtaining data fidelity assessment parameters OP = {Of, Os} and multi-source situation input tensor From which the primary data fidelity assessment parameter Of and the secondary data fidelity assessment parameter Os are extracted, and the multi-source situation input tensor is converted into binary data B2X; The computer-aided detection parameter Ap = g Os x y Of (mod p); where y is a predefined auxiliary production verification code, y = g x (mod p); Calculate the data fidelity Dc=H(B2X||AP)(mod q); If Dc = Of, it means that the authenticity and accuracy of the acquired multi-source situation input tensor ; otherwise, an alarm is immediately given.

4. The cloud platform-based medical instrument production management method of claim 3, wherein, The spatio-temporal fusion state embedding vector generation process is as follows: Input the multi-source situation input tensor; The multi-layer one-dimensional convolutional network uses convolution kernel weight sliding convolution along the time dimension to pass and apply the ReLU activation function to the input features layer by layer, and outputs the convolutional layer output; The convolutional layer output is input into the multi-head self-attention module, which independently calculates the cross-temporal and spatial long-range dependencies through multi-head splitting and connection projection to generate fused features; The global average pooling is used to compress the fused features along the time axis to obtain the pooling output features, and then the non-linear transformation output is generated through the fully connected layer combined with the weight and bias to output the production situation vector, that is, the spatio-temporal fusion state embedding vector.

5. The cloud platform-based medical instrument production management method according to claim 4, characterized in that, The positioning process of the abnormal node starting the causal chain reconstruction to locate the upstream link is as follows: Based on the executed task path, a compliance behavior graph is constructed, the node set N represents the production operation, the relationship set R represents the relationship between nodes, and the attribute set A represents the node attribute, which is used to represent the structure of task and compliance information; Each node of the graph is automatically matched with the compliance clauses in the regulation library, and three compliance dimensions of time window, operation environment and personnel qualification are labeled; For the nodes with abnormal compliance labels in the completed behavior sequence, multi-source causal chain reconstruction is started to locate the upstream link that causes the abnormality, and the abnormal causal chain is defined as: ； wherein, represents an abnormal causal chain, i.e. a path from a certain historical task n i until the path that triggered the abnormality; represents a sequence of tasks on the abnormal chain, arranged in the order of task execution, i.e. there is a potential causal relationship; represents whether the task Task j triggers the out-of-bound behavior at time t.

6. The cloud platform-based medical instrument production management method according to claim 5, characterized in that, The control strategy generation function takes the current state vector and the updated policy graph as input, generates a task path score through a graph neural network, and selects the highest scoring path to execute control.

7. The cloud platform-based medical instrument production management method according to claim 4, characterized in that, The multi-layer one-dimensional convolutional network includes: The first layer: the convolution kernel size is 5*5, the input channel number is C, the output channel number is 64, the step is 1, the padding is 2, and the activation function is ReLU; The second layer: the convolution kernel size is 3*3, the input channel number is 64, the output channel number is 128, the step is 1, the padding is 1, and the activation function is ReLU; The third layer: the convolution kernel size is 3*3, the input channel number is 128, the output channel number is 256, the step is 1, the padding is 1, and the activation function is ReLU.

8. A cloud platform-based medical instrument production management system for performing the cloud platform-based medical instrument production management method of any one of claims 1 to 7. It includes: A data acquisition module for unified acquisition, event-driven triggering and structured modeling of multi-source heterogeneous information involved in the production process of medical devices, and output of multi-source situation input tensor; The twin state generation module is deployed on the cloud platform, which ensures the reliability of the input through data fidelity verification, uses temporal convolution and multi-head attention to fuse spatio-temporal features, generates virtual states through physical perception kernel, process flow kernel and semantic structure kernel, and realizes dynamic update of twin body through LSTM prediction mechanism, to build a digital twin modeling framework that synchronizes virtual and real; The flexible strategy arrangement module is deployed on the cloud platform, which constructs a task dependency graph structure based on the spatio-temporal fusion state embedding vector, generates a flexible task strategy graph by introducing historical execution bias, violation feedback factor and dynamic weight of resources, and dynamically adjusts the task path structure in real time to support multi-state driven production scheduling and task switching control; The compliance traceability analysis module is deployed on the cloud platform, which is used to map various process data, task execution trajectory and equipment personnel operation sequence into a compliance traceability graph during production execution, to realize structured representation, trigger event identification and causal chain analysis of potential violation behaviors. The closed-loop cooperative control module is used for fusing the adjusted flexible task strategy graph and the current system state, generating a multi-dimensional control instruction set through a control strategy function, dynamically balancing and optimizing the control effect through a multi-objective optimization function, and finally issuing the strategy to the field control system.

Images