Medical instrument production management system and method based on cloud platform

Abstract

The invention relates to the technical field of medical instrument manufacturing and management, in particular to a medical instrument production management system and method based on a cloud platform. The method comprises the following steps: collecting multi-source data, and ensuring the authenticity and accuracy of the data by adopting Hash binding and discrete logarithm problems; generating a production line state embedding vector by using time sequence convolution and multi-head self-attention fusion spatial-temporal characteristics; constructing a flexible task strategy graph based on a twinning state, and dynamically adjusting task path resource consumption, compliance risk and stability deviation weight in combination with multi-objective optimization; visual tracing is realized through flow map modeling and multi-scale compliance labeling, an illegal causal chain is positioned, and the weight of an adjustment strategy is fed back; and finally, generating a self-adaptive control instruction by the graph attention network. According to the method, multi-modal fusion, strategy graph dynamic optimization and causal inversion technologies are integrated, a perception-optimization-tracing-regulation closed loop is formed, resource scheduling, compliance and production stability are collaboratively optimized, and actual application requirements are met.

Description

Technical Field

[0001] The present invention relates to the technical field of medical device manufacturing and management, and in particular to a cloud platform-based medical device production management system and method. Background Art

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

[0003] A Chinese invention patent application, publication number CN115730795A, discloses a cloud-based medical device production management system, comprising a workshop control terminal, optical fiber lines, and a cloud platform. The workshop control terminal is connected to an operation display terminal, which is in turn connected to a central processing unit. In this cloud-based medical device production management system, information input by staff into the touch display terminal, while data on the medical device processing scene, process, and quality collected by monitors, numerical control equipment, and quality inspection platforms, are all stored internally in the workshop control terminal. Staff handheld devices can wirelessly access the cloud platform's internal medical device production and processing scene, process, and quality, as well as the medical device's 3D model, 3D dimensions, and materials, and other medical device data in the operation display terminal. This allows the entire factory to synchronously track the medical device production process at all times, avoiding issues caused by time differences.

[0004] The Chinese invention patent with announcement number CN118798494B discloses a production control method and system for the intelligent clothing industry based on data analysis. It sets up an intelligent data collection platform, collects heterogeneous data for common feature extraction, generates a comprehensive data set, and constructs a digital twin model of clothing production; extracts data based on the comprehensive data set, generates triples and constructs a clothing production knowledge graph, constructs a clothing production data analysis model, extracts key features and determines implicit association rules, generates a cross-modal knowledge graph, identifies clothing production control rules and solves clothing production control parameters; constructs a decision evaluation model to parse clothing production control parameters into production control decisions, evaluates them, and dynamically corrects them based on the evaluation results to obtain optimized production control decisions, which are sent to the equipment terminal and adaptively optimize clothing production. Real-time production data is collected and the digital twin model of clothing production and the clothing production knowledge graph are updated, and periodic optimization is performed until the production task is completed.

[0005] However, the production process presents complex characteristics such as multi-constraint coupling, multi-link coordination, and multi-state evolution. Especially in the scenario of multi-model mixed lines, it is necessary to take into account resource utilization, quality compliance and production stability; at the same time, strict regulatory supervision requirements pose challenges to the traceability of the entire process and the efficiency of abnormal positioning; and with the development of emerging technologies such as cloud computing, digital twins, graph intelligence, and reinforcement learning, their integration and application in 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, providing technical support for building an efficient, safe, and accountable medical device production system. It is urgent to build a multi-source data fusion, virtual-reality linkage and intelligent decision-making system based on a cloud platform to achieve dynamic collaborative management of the entire production chain. Summary of the Invention

[0006] The purpose of the present invention is to address the problems existing in the background technology and propose a medical device production management system and method based on a cloud platform.

[0007] The technical solution of the present invention is a cloud-based medical device production management method, which includes the following specific implementation steps:

[0008] S1. Collect production status data from the medical device production line, convert it into a multi-source state input tensor through format conversion and time synchronization normalization, convert it into binary form, and generate data fidelity assessment parameters through 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 a spatiotemporal fusion state embedding vector. This is then mapped to the virtual production line twin state output using a hierarchical kernel function, and an LSTM is introduced to predict the twin mapping production line state.

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

[0011] S4. Build a compliance behavior graph based on executed task paths, automatically match node status with three compliance dimensions, initiate causal chain reconstruction for abnormal nodes, locate upstream links, visualize violation paths and responsible individuals through Sankey diagram hotspots, and provide feedback to adjust the Twin model prediction weights and strategy map preferences.

[0012] S5. Real-time collection of equipment status, environmental parameters, personnel behavior, and boundary signals are modeled as state flow vectors. The edge weights of the strategy graph are dynamically adjusted through compliance anomaly feedback. The optimal scheduling path and multi-dimensional control instruction set are generated based on the control strategy generation function. After execution, the feedback is fed back to the twin model and the strategy graph to form a reinforcement learning closed-loop control.

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

[0014] Input 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), that is, q divides p-1; g is the multiplication cyclic group Z p * Generators of

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

[0017] Where H is a predefined hash function whose output length matches the number of bits in q; || represents the concatenation operation;

[0018] Calculate the order data fidelity assessment parameter Os = (s-Of×x)(mod q);

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

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

[0021] Preferably, the multi-source data fidelity detection process is as follows:

[0022] Get data fidelity evaluation parameters OP = {Of, Os} and 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, and input the multi-source situation into the 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 the predefined auxiliary production inspection code, y = g x (mod p);

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

[0026] If Dc=Of, it means that the multi-source situation input tensor X is obtained (0)Otherwise, an alarm will be issued immediately.

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

[0028] Input multi-source situation input tensor;

[0029] Use a multi-layer one-dimensional convolutional network to slide convolution along the time dimension through the convolution kernel weights, combine the bias term to pass the input features layer by layer and apply the ReLU activation function to output the convolution layer output;

[0030] The output of the convolutional layer is input into the multi-head self-attention module, and the multi-head split independently calculates the long-range dependencies across time and space, and the concatenated projections generate fused features;

[0031] Global average pooling is used to compress the fusion features along the time axis to obtain the pooled output features, and then the production situation vector is output through the fully connected layer combined with the weight and bias nonlinear transformation, that 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 hierarchically mapped to the virtual state output through three types of kernel functions: the physical perception kernel uses RBF to capture the static dependency of the equipment; the process flow kernel uses MLP to model dynamic transmission; the semantic structure kernel uses GCN to reflect soft constraint regulation; and Gaussian perturbations are superimposed to simulate external interference.

[0033] Preferably, the implementation process of the feedback reconstruction strategy based on the boundary trigger function is as follows:

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

[0035] Based on the production situation vector and twin mapping production line status, a flexible strategy graph is defined, the current production line status is mapped to the strategy graph nodes, and the dynamic weight of the flexible strategy graph is defined to integrate resource consumption, compliance risk and status deviation;

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

[0037] Where J(π) represents the expected total reward of path π; L represents the length of the task path; U k represents the improvement in resource utilization after executing 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; β1, β2, β3 represent the weighted coefficients of the reward function; E π(·) represents the expected reward function;

[0038] The path π * (t) is issued, parallel execution is started, and the boundary trigger function θ is defined k (t):

[0039]

[0040] Among them, θ k (t) represents task T k Whether out-of-bounds control is triggered 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; ε k Indicates the maximum time error range allowed for the k-th task;

[0041] Once θ k (t)=1, the task will be interrupted and the strategy reconstruction will be triggered.

[0042] Preferably, the process of starting the causal chain reconstruction and locating the upstream link of the abnormal node is as follows:

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

[0044] Automatically match compliance clauses from the regulatory database to each node in the graph, marking three compliance dimensions: time window, operating environment, and personnel qualifications;

[0045] For nodes with compliance anomalies in the completed behavior sequence, we start multi-source causal chain reconstruction, locate the upstream link that caused the anomaly, and define the anomaly causal chain as:

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

[0047] Among them, C abnormal Represents an abnormal causal chain, that is, from a historical task n i The path from the beginning until the exception is triggered; n i →n j →…→n k Represents the task sequence on the exception chain, which is arranged in the order of task execution, that is, there is a potential causal relationship; θ j (t) represents the task j Whether out-of-bounds behavior is triggered at time t.

[0048] Preferably, the control strategy generation function takes the current state vector and the updated strategy graph as input, generates a task path score through a graph neural network, and selects the highest-scoring path to perform control.

[0049] Preferably, the multi-layer one-dimensional convolutional network includes:

[0050] First layer: convolution 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: The convolution kernel size is 3×3, the number of input channels is 128, the number of output channels is 256, the stride is 1, the padding is 1, and the activation function is ReLU.

[0053] The technical solution of the present invention is a cloud-based medical device production management system, which is used to implement the above-mentioned cloud-based medical device production management method, including:

[0054] The data acquisition module is used to uniformly collect, event-driven trigger, and structure model the multi-source heterogeneous information involved in the medical device production process, and output multi-source situation input tensors;

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

[0056] The flexible strategy orchestration module, deployed on the cloud platform, builds a task dependency graph 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 the cloud platform, is used to map various process data, task execution trajectories, and equipment and personnel operation sequences into compliance traceability diagrams during production execution, enabling structured representation of potential violations, trigger event identification, 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 status, 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 technical solution of the present invention has the following beneficial technical effects:

[0060] This paper designs a cloud-based medical device production management system and method, which achieves dynamic synchronization of virtual and real through a multi-level modeling framework, integrates multimodal perception, convolutional attention collaborative modeling, and kernel function mapping, accurately captures the spatiotemporal state evolution of the production line, and provides a high-confidence decision-making basis for task scheduling.

[0061] A dynamic orchestration mechanism based on a flexible strategy graph, combined with multi-objective optimization and compliance feedback, adjusts task paths in real time, significantly improving resource utilization and compliance in mixed-line production of multiple product models and reducing the risk of non-compliance.

[0062] Build a visual compliance traceability chain and closed-loop control framework, locate the source of violations through process map modeling and causal inversion, achieve full production process responsibility traceability and dynamic optimization, shorten abnormal response time, and drive 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 binding are constructed based on large prime numbers to ensure the authenticity and readiness of original multimodal data (such as temperature and image stream); data anti-repudiation and source tracing are achieved through the strong correlation between sequence parameters and predefined parameters; auxiliary detection parameters are used for rapid inspection in the verification phase to intercept suspicious data (such as sensor signal forgery and operation log tampering) to avoid erroneous data input into the digital twin model, which may lead to scheduling decision deviations. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of a method for cloud platform-based medical device production management proposed by the present invention. DETAILED DESCRIPTION

[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 sensor data (production status data) from each node 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: including but not limited to image streams (machine vision), voice commands (operator voice interaction), and log text;

[0069] After that, the multimodal perception data is first format converted and time synchronized, and the structured and unstructured data are uniformly converted into a standard format suitable for computer processing. Then, normalization preprocessing is used to obtain the normalized multimodal production line situation data (i.e., multi-source situation input tensor) X (0) ;

[0070] And for the multi-source situation input tensor X (0) Generate data fidelity evaluation parameters. The generation process is as follows:

[0071] (1) 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);

[0072] Where p is a predefined large prime number (1024 bits in this embodiment) that satisfies the discrete logarithm problem difficulty; 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 multiplication cyclic group Generators of

[0073] (2) Calculate the first-order data fidelity assessment 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 of q (i.e., 256 bits); || represents a concatenation operation;

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

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

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

[0078] S2. Build a multi-level modeling framework that integrates multimodal perception, convolutional attention collaborative modeling, kernel function mapping, and twin prediction and updating. Based on the multimodal situation embedding mechanism, unify multi-source data, introduce a kernel mapping digital twin modeling structure, and build a twin state prediction-update mechanism to enable dynamic synchronization of virtual and real models with production evolution. The specific implementation process is as follows:

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

[0080] (1) Obtain data fidelity evaluation parameters OP = {Of, Os} and 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, and input the multi-source situation into the 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 the 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, then the multi-source situation input tensor X is obtained (0) Otherwise, an alert will be given immediately;

[0085] S22, the multi-source situation input tensor X after normalization processing output from step S1 (0) , constructing a spatiotemporal fusion modeling architecture that combines temporal convolution with multi-head self-attention to deeply capture the temporal evolution characteristics and spatial modal collaboration in the production environment, thereby generating a production line state embedding vector containing rich spatiotemporal information. Specifically:

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

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

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

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

[0090] First layer (Conv1): convolution 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;

[0091] The second layer (Conv2): convolution kernel size is 3×3, the number of input channels is 64, the number of output channels is 128, the stride is 1, the padding is 1, and the activation function is ReLU;

[0092] The third layer (Conv3): The convolution kernel size is 3×3, the number of input channels is 128, the number of output channels is 256, the stride is 1, the padding is 1, and the activation function is ReLU;

[0093] A3, convolutional layer output H (L) As the input of the Multi-Head Self-Attention (MHSA) module, it captures long-range dependencies across time and modality. The multi-head self-attention is defined as:

[0094]

[0095] Among them, Q, K, and V represent the query, key, and value matrices in the attention mechanism, which are represented by H (L) Multiply by the projection matrix to calculate the attention weight and aggregate features; Attention(Q,K,V) represents the standard scaled dot product attention; d k Represents the feature dimension within a single attention head; W Q 、W K 、W V They represent the multi-head self-attention used to (L) Projection into a trainable matrix of query, key, and value spaces;

[0096] Specifically, enter H (L) Expand into a sequence by modality and time; split into h heads through a multi-head mechanism, each head independently calculates attention; after the attention outputs of each head are connected, they are projected back to the original dimension to obtain the fused feature representation Z;

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

[0098]

[0099] Where Pooling(·) represents the pooling function, which aggregates Z into a fixed length along the time dimension; represents the generated spatiotemporal fusion state embedding vector, i.e., the production situation vector; W s represents the state mapping fully connected weight; b s Represents the state mapping full connection bias; Flatten(·) represents the full connection operation;

[0100] S23, the situation vector Input 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 status outputs:

[0102] Among them, D t Represents the virtual state output vector of the production line; φ k (·) represents the kth kernel mapping function, i.e., the nonlinear feature transformation from state to twin index; ω k Represents the kernel function weight, reflecting its importance in mapping (adaptively learned through training); K represents the number of kernel functions; ε t represents Gaussian disturbance, which is used to simulate unobserved external interference and system error;

[0103] It should be noted that in order to enhance mapping flexibility and interpretability, three types of kernel functions are introduced and act on state variables of different dimensions in a hierarchical manner:

[0104] The physical perception core captures the static physical dependencies between device states (e.g., the relationship between temperature and pressure) using radial basis functions (RBF);

[0105] The process core models the dynamic transfer between production processes (such as time lags between multiple workstations) using a multi-layer perceptron (MLP) with a nested combination of nonlinear activation and parameter learning.

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

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

[0108] in, Indicates the twin state predicted at the next moment; u t 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 the long short-term memory network (LSTM);

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

[0110] S31, combined with the current digital twin state D in step S2 t , extract the task structure and process dependency of the target medical device product, and construct the task set Task and the directed graph structure G = (V, E);

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

[0112] Among them, Task i represents the i-th production task; n represents the total number of production tasks; s i represents the starting resources required for the i-th task (including but not limited to machine model and manpower level); r i represents the compliance requirements corresponding to the i-th task (including but not limited to disinfection time and humidity range); q i represents the quality level indicator of the i-th task; V represents the node set, V = Task; E represents the relationship attribute set, that is, the process dependency between tasks, such as "encapsulation must be carried out after sterilization";

[0113] S32, the production situation vector obtained according to step S2 Prediction with digital twins Map the current production line status to the strategy graph node to achieve dynamic adjustability of the task flow graph and define the flexible strategy graph G f for:

[0114]

[0115] Among them, W represents the edge weight set in the graph, reflecting the comprehensive cost under different paths; w ij Represents the task node Task i →Task j The state dynamic weight of c ij Represents node i(Task i ) to node j(Task j ) required unit resource consumption (including but not limited to machine occupancy rate and personnel ratio); ij Represents node i(Task i ) to node j(Task j ) The risk level of the impact of the switch on the compliance chain (including but not limited to the risk of the sterile chain being broken due to the switch); ij represents the deviation between the current state vector and the expected state of the target task (reflecting the stability of the operation); α1, α2, and α3 represent the weighted coefficients obtained by learning optimization, satisfying α1+α2+α3=1;

[0116] S33, using the improved policy gradient optimization algorithm (PPO + multi-path reward reconstruction), in the strategy graph G f Solve the current optimal task path π * (t), define the task path optimization objective function J(π):

[0117]

[0118] Among them, J(π) represents the expected total reward of path π, which is used to measure the quality of scheduling; L represents the length of the task path; U k represents the improvement in resource utilization after executing the kth task (such as the change in machine utilization); Q k represents the positive contribution factor of the kth task to the overall quality compliance; R k represents the risk factors caused by the execution of the kth task (including but not limited to cross contamination and packaging before cooling); β1, β2, and β3 represent the weighted coefficients of the reward function, which control the resource, quality, and risk trade-offs respectively; E π (·) represents the expected reward function;

[0119] S34, the path π * (t) is issued, parallel execution is started, and the boundary trigger function θ is defined k (t):

[0120]

[0121] Among them, θ k (t) represents task Tk Whether out-of-bounds control is triggered 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; ε k Indicates the maximum time error range allowed for the k-th task;

[0122] Once θ k (t) = 1, which will be fed back to the twin prediction model in step S2 and compliance tracing in step S4, while interrupting the task and triggering strategy reconstruction;

[0123] S4. Through process state graph modeling, multi-scale behavior annotation, and compliance causal chain inversion, a visual, interactive, and accountable 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 the path c ,

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

[0126] S42, each node n in the graph i Match the compliance clauses in the regulatory knowledge base such as national standards and industry specifications to automatically generate the compliance annotation sequence L i , and establish three types of compliance dimension mapping:

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

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

[0129] Dimension three, personnel compliance, that is, whether the operator qualifications meet the standards;

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

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

[0132] Among them, C abnormal Represents an abnormal causal chain, that is, from a historical task n i The path from the beginning until the exception is triggered; n i →n j →…→n k Represents the task sequence on the exception chain, which is arranged in the order of task execution, that is, there is a potential causal relationship; θ j (t) represents the task j Whether out-of-bounds behavior (including but not limited to delays and illegal environments) is triggered at time t;

[0133] Specifically, the node chain represents the sequence of operation behaviors, and an out-of-bounds signal θ appears at the intermediate node. k (t) = 1, triggering abduction;

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

[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 in the compliance chain are marked in red to show the type of non-compliance;

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

[0138] At the same time, the data is fed back to the digital twin model in step S1 to adjust its compliance factor prediction weight; it is also fed back to the flexible strategy map in step S3 to adjust the weight preference of future scheduling paths;

[0139] S5. Build 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, collects the following multi-source status data in real time:

[0141] Equipment 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 crossing signal θ k (t);

[0145] Model multidimensional data as a state flow vector:

[0146] S52, based on the compliance anomaly feedback result in step S4 (such as abnormal node set C abnormal , illegal path) as input signal, the strategy graph G f The execution weight of the corresponding node is dynamically adjusted. The specific update method is:

[0147] Among them, w' ij represents the updated strategy graph edge weight, which is used to guide the next round of task scheduling after adjustment; w ij Represents the original flexible strategy graph from node n i to n j The edge right; represents the intensity of compliance deviation between nodes in historical scheduling (including but not limited to violation frequency and violation severity); λ represents the learning step factor (adjustment amplitude) of edge weight adjustment;

[0148] S53, according to the updated strategy graph G' f and the current state flow S t , using the control strategy to generate the function: C t =f(G' f ,S t ), output multi-dimensional control instruction set C t ,Then the generated instructions will be distributed to the local through the cloud platform;

[0149] Among them, C t represents the set of control instructions that need to be issued at the current moment; f(·) represents the control strategy generation function, which fuses the state and strategy graph to derive 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 and the adjusted strategy graph G' f Perform fusion encoding and use the Graph Attention Network (GAT) mechanism to inject state features into node representation:

[0152] Where N(i) represents the set of adjacent nodes of node i; represents the attention weight of node i to adjacent node j; Represents the local state slice of the state of node i; W (l) Represents the trainable parameters of layer l; represents the hidden state vector of node i in layer l; represents the hidden state vector of node i in the output l+1 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 in the previous step, score the task path and select the optimal scheduling path π that meets the current state (*) (t):

[0154]

[0155] Select the path with the maximum score:

[0156] Among them, Score(π) represents the comprehensive score of the candidate task path π; represents the adjusted edge weight between nodes; γ ij (S t ) represents the adaptability factor of task transfer from i→j in the current state; (i, j)∈π represents the edge from node i to j in the path;

[0157] (3) For the strategy path π (*) Each node and edge in (t) generates a corresponding control signal

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

[0159] Equipment control instructions (such as temperature control value, voltage, speed);

[0160] Environmental adjustment instructions (such as air valve adjustment, pressure difference increase);

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

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

[0163] S54. After being sensed, all execution behaviors (device status changes, task completion, and boundary triggering) are uniformly fed back to the digital twin compliance prediction model and the strategy map adjustment model, forming a continuously evolving control closed loop. At the same time, the following reinforcement learning update mechanism is established:

[0164] Among them, π (*)(t+1) represents the strategy path after the next round of iterative 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 path change rate; The gradient indicating the direction of policy performance improvement.

[0165] In the second embodiment, the present invention proposes a cloud platform-based medical device production management system, which is used to execute the cloud platform-based medical device production management method proposed in the first embodiment, including: a data acquisition module, a twin state generation module, a flexible policy 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 structure model multi-source heterogeneous information involved in the medical device production process, including 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, deployed on the cloud platform, ensures input reliability through data fidelity verification. It uses temporal convolution and multi-head attention to fuse spatiotemporal features. It generates virtual states through hierarchical mapping of physical perception cores, process flow cores, and semantic structure cores. It also introduces an LSTM prediction mechanism to achieve dynamic updates of the twin, building a digital twin modeling framework that synchronizes virtual and real states.

[0168] The flexible strategy orchestration module, deployed on the cloud platform, builds a task dependency graph 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 diagrams during production execution. This enables structured representation of potential violations, identification of triggering events, and causal chain analysis, providing visual backtracking and dynamic compliance assessment capabilities 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 status, generate a multi-dimensional control instruction set through the control strategy function, covering equipment control parameters, environmental adjustment strategies and personnel guidance tasks, and use multi-objective optimization functions to dynamically balance and optimize the control effect. Finally, the strategy is issued to the on-site 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 are described in detail above with reference to the accompanying drawings, but 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 medical device production management method based on a cloud platform, characterized in that: The specific implementation steps include the following: S1. Collect production status data from the medical device production line, convert it into a multi-source state input tensor through format conversion and time synchronization normalization, convert it into binary form, and generate data fidelity assessment parameters through hash function and discrete logarithm calculation. S2: Detect the fidelity of multi-source data and use temporal convolution and multi-head self-attention architecture to generate a spatiotemporal fusion state embedding vector. This is then mapped to the virtual production line twin state output using a hierarchical kernel function, and an LSTM is introduced to predict the twin mapping production line state. S3. Build a task dependency graph based on the twin state, define a flexible strategy graph to dynamically adjust edge weights to consider resource consumption, compliance risks, and state deviations, use a policy gradient optimization algorithm to optimize task paths to balance resource utilization, quality compliance, and risks, and trigger feedback reconstruction strategies based on boundary trigger functions during execution. S4. Build a compliance behavior graph based on executed task paths, automatically match node status with three compliance dimensions, initiate causal chain reconstruction for abnormal nodes, locate upstream links, visualize violation paths and responsible individuals through Sankey diagram hotspots, and provide feedback to adjust the Twin model prediction weights and strategy map preferences. S5. Real-time collection of equipment status, environmental parameters, personnel behavior, and boundary signals are modeled as state flow vectors. The edge weights of the strategy graph are dynamically adjusted through compliance anomaly feedback. The optimal scheduling path and multi-dimensional control instruction set are generated based on the control strategy generation function. After execution, the feedback is fed back to the twin model and the strategy graph to form a reinforcement learning closed-loop control.

2. A cloud platform-based medical device production management method according to claim 1, characterized in that: The process of generating data fidelity evaluation parameters is as follows: Input 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); 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), that is, q divides p-1; g is a multiplication cyclic group Generators of Calculate the first-order data fidelity assessment parameter Of = H (BX || r) (mod q); Where H is a predefined hash function whose output length matches the number of bits in q; || represents the concatenation operation; Calculate the order data fidelity assessment parameter Os = (s-Of×x)(mod q); Where x is a predefined auxiliary production code, i.e., a randomly selected integer, x∈[1,q-1]; Output data fidelity evaluation parameter OP = {Of, Os}.

3. A cloud platform-based medical device production management method according to claim 2, characterized in that: The process of detecting the fidelity of multi-source data is as follows: Get data fidelity evaluation parameters OP = {Of, Os} and 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, and input the multi-source situation into the tensor X (0) Convert to binary data B2X; Calculate the auxiliary detection parameter Ap=g Os ×y Of (mod p); Where y is the predefined auxiliary production inspection code, y = g x (mod p); Calculate data fidelity Dc = H(B2X||AP)(mod q); If Dc=Of, it means that the multi-source situation input tensor X is obtained (0) Otherwise, an alarm will be issued immediately.

4. A cloud platform-based medical device production management method according to claim 3, characterized in that: The process of generating the spatiotemporal fusion state embedding vector is as follows: Input multi-source situation input tensor; Use a multi-layer one-dimensional convolutional network to slide convolution along the time dimension through the convolution kernel weights, combine the bias term to pass the input features layer by layer and apply the ReLU activation function to output the convolution layer output; The output of the convolutional layer is input into the multi-head self-attention module, and the multi-head split independently calculates the long-range dependencies across time and space, and the concatenated projections generate fused features; Global average pooling is used to compress the fusion features along the time axis to obtain the pooled output features, and then the production situation vector is output through the fully connected layer combined with the weight and bias nonlinear transformation, that is, the spatiotemporal fusion state embedding vector.

5. A cloud platform-based medical device production management method according to claim 4, characterized in that: The twin state output process of the virtual production line is as follows: the input situation vector is hierarchically mapped to the virtual state output through three types of kernel functions: the physical perception kernel uses RBF to capture the static dependencies of the equipment; the process kernel uses MLP to model dynamic transmission; the semantic structure kernel uses GCN to reflect soft constraint regulation; and superimposed Gaussian perturbations simulate external interference.

6. A cloud platform-based medical device production management method according to claim 5, characterized in that: The implementation process of the feedback reconstruction strategy based on the boundary trigger function during execution is as follows: Based on the virtual production line twin state, the task structure and process dependencies are extracted, a task set is constructed, and resource requirements, compliance requirements, and quality indicators are defined, mapped into directed graph nodes and dependency edges. Based on the production situation vector and twin mapping production line status, a flexible strategy graph is defined, the current production line status is mapped to the strategy graph nodes, and the dynamic weight of the flexible strategy graph is defined to integrate resource consumption, compliance risk and status deviation; The policy gradient optimization algorithm is combined with the multi-path reward reconstruction algorithm to optimize the task path on the strategy graph. The objective function is: Where J(π) represents the expected total reward of path π; L represents the length of the task path; U k represents the improvement in resource utilization after executing 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; β1, β2, β3 represent the weighted coefficients of the reward function; E π (·) represents the expected reward function; The path π * (t) is issued, parallel execution is started, and the boundary trigger function θ is defined k (t): Among them, θ k (t) represents task T k Whether out-of-bounds control is triggered 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; ε k Indicates the maximum time error range allowed for the k-th task; Once θ k (t)=1, the task will be interrupted and the strategy reconstruction will be triggered.

7. A cloud platform-based medical device production management method according to claim 6, characterized in that: The process of reconstructing the causal chain and locating the upstream link of the abnormal node is as follows: A compliance behavior graph is constructed based on the executed task path. The node set N represents production operations, the relationship set R represents the relationship between nodes, and the attribute set A represents the node attributes. This is used to represent the task and compliance information structure. Automatically match compliance clauses from the regulatory database to each node in the graph, marking three compliance dimensions: time window, operating environment, and personnel qualifications; For nodes with compliance anomalies in the completed behavior sequence, we start multi-source causal chain reconstruction, locate the upstream link that caused the anomaly, and define the anomaly causal chain as: C abnormal ={n i →n j →…→n k |θ j (t)=1}; Among them, C abnormal Represents an abnormal causal chain, that is, from a historical task n i The path from the beginning until the exception is triggered; n i →n j →…→n k Represents the task sequence on the exception chain, which is arranged in the order of task execution, that is, there is a potential causal relationship; θ j (t) represents the task j Whether out-of-bounds behavior is triggered at time t.

8. A cloud platform-based medical device production management method according to claim 7, characterized in that: The control strategy generation function takes the current state vector and the updated strategy graph as input, generates task path scores through the graph neural network, and selects the highest-scoring path to execute control.

9. A cloud platform-based medical device production management method according to claim 4, characterized in that: The multi-layer one-dimensional convolutional network includes: First layer: convolution 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; 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; The third layer: The convolution kernel size is 3×3, the number of input channels is 128, the number of output channels is 256, the stride is 1, the padding is 1, and the activation function is ReLU.

10. A cloud-based medical device production management system, used to implement the cloud-based medical device production management method according to any one of claims 1 to 9, characterized in that: include: The data acquisition module is used to uniformly collect, event-driven trigger, and structure model the multi-source heterogeneous information involved in the medical device production process, and output multi-source situation input tensors; The twin state generation module, deployed on the cloud platform, ensures input reliability through data fidelity verification. It uses temporal convolution and multi-head attention to fuse spatiotemporal features. It generates virtual states through hierarchical mapping of physical perception cores, process flow cores, and semantic structure cores. It also introduces an LSTM prediction mechanism to achieve dynamic updates of the twin, building a digital twin modeling framework that synchronizes virtual and real states. The flexible strategy orchestration module, deployed on the cloud platform, builds a task dependency graph 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. 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 diagrams during production execution, enabling structured representation of potential violations, trigger event identification, and causal chain analysis. The closed-loop collaborative control module is used to integrate the adjusted flexible task strategy diagram with the current system status, 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.

Images