Neuromuscular stimulator fault calibration method based on federal transfer learning

Abstract

The invention discloses a neuromuscular stimulator fault calibration method based on federal transfer learning, and the method comprises the steps: obtaining equipment data of a neuromuscular stimulator and user physiological data, and constructing the equipment data and user physiological data into a dynamic heterogeneous graph representing the state of an equipment and user interaction system; based on a federated learning framework, a heterogeneous graph neural network model used for representing a general fault mode is trained cooperatively through multiple devices; on a single neuromuscular stimulator, on the basis of the model, a personalized calibration model adapting to individual differences is generated through transfer learning fine tuning; and finally, processing the current system state diagram by using the personalized calibration model, and outputting calibration values for the control parameters of the neuromuscular stimulator. According to the method, the equipment-user system is modeled into the heterogeneous graph, so that a fault source can be deeply understood, accurate calibration is realized, and the stimulation effect and the safety are improved; and meanwhile, the federal transfer learning normal form is adopted, so that the problems of insufficient model generalization ability and low personalized calibration precision are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault calibration of medical equipment, and particularly relates to a neuromuscular stimulator fault calibration method based on federated transfer learning. BACKGROUND

[0002] Neuromuscular Electrical Stimulation (NMES) as an important physical therapy and rehabilitation technology, it can induce muscle contraction by applying precise controlled electrical pulses to target muscle groups or nerves, and is widely used in post-stroke rehabilitation, spinal cord injury rehabilitation, muscle atrophy prevention and treatment, and sports ability enhancement. With the development of microelectronic technology and sensing technology, modern neuromuscular stimulators are evolving towards intelligence, portability and personalization. Early devices mainly adopt open-loop control mode, that is, fixed stimulation parameters (such as current amplitude, pulse width, frequency) are preset. This method cannot adapt to the optimal stimulation window drift caused by dynamic factors such as muscle fatigue, electrode-skin interface impedance changes, device component aging and individual physiological differences during treatment, resulting in a decline in stimulation effect and even causing discomfort. In order to solve this problem, researchers introduced a closed-loop control strategy, which integrates physiological feedback signals such as surface electromyography (sEMG), force sensors or inertial measurement units (IMU) to try to adjust the stimulation parameters in real time. However, these traditional closed-loop control methods mostly rely on classic PID (Proportional-Integral-Derivative) controllers or threshold judgment logic based on signal processing features, and their models are relatively simple, making it difficult to effectively process multi-modal, high-dimensional and nonlinear physiological and device data. The ability to identify and calibrate complex fault modes is limited, especially when facing output deviations caused by the coupling of multiple factors (such as capacitor aging and skin sweating), it is often difficult to accurately decouple and compensate.

[0003] To overcome the limitations of traditional methods, the industry has begun to explore the application of machine learning techniques in the intelligent calibration of NMES. Some studies attempt to use support vector machines (SVM) or shallow neural networks to classify fault states, or use time series prediction models such as ARIMA to predict stimulation effects. Although these methods have improved the level of intelligence of the device to some extent, there are still problems. First, these models usually treat multi-source data as independent, flattened data streams, ignoring the complex physical connections, causal effects, and physiological interactions between internal components, control parameters, output signals, and user physiological feedback. The lack of this structured information limits the model's ability to deeply understand the root cause of system-level faults. Second, existing machine learning solutions mostly use centralized training mode, which requires uploading large amounts of sensitive physiological data from different users to the central server, which not only brings huge data privacy and security risks, but also is difficult to adapt to the trend of edge devices working collaboratively in the Internet of Medical Things (IoMT) environment. Third, due to the significant heterogeneity of different individuals, different devices, and even the same device at different usage stages, a general model trained on laboratory data often faces serious problems of insufficient generalization ability when deployed to specific users, i.e., the model cannot quickly adapt to the unique physiological characteristics of specific users and the individual aging trajectory of the device, resulting in a significant reduction in calibration accuracy. SUMMARY

[0004] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, the abstract and the title, and such simplifications or omissions are not used to limit the scope of the present application.

[0005] In view of the above existing problems, the present application is proposed. Therefore, the present application provides a neural muscle stimulator fault calibration method based on federated transfer learning to solve the problems proposed in the background art.

[0006] To solve the above technical problems, the present application provides the following technical solutions: a neural muscle stimulator fault calibration method based on federated transfer learning, comprising:

[0007] Obtaining device data and user physiological data of a neural muscle stimulator, and constructing graph structure data representing the state of the neural muscle stimulator and user interaction system;

[0008] Based on the federated learning framework, the graph structure data is taken as input, and a global graph neural network model for representing fault patterns is trained collaboratively by multiple neural muscle stimulators;

[0009] On each neuromuscular stimulator, based on the global graph neural network model, a personalized calibration model is generated through transfer learning;

[0010] Using the personalized calibration model, the graph structure data corresponding to the current system state is processed, and the calibration value of the control parameter of the neuromuscular stimulator is output.

[0011] As a preferred scheme of the neuromuscular stimulator fault calibration method based on federal transfer learning, wherein: the graph structure data representing the neuromuscular stimulator and user interaction system state is constructed, including:

[0012] The preset entities in the device data and physiological data are defined as multiple types of nodes, and the heterogeneous edges connecting the nodes are defined according to the preset relationship between the preset entities, thereby constructing a dynamic heterogeneous graph as the graph structure data.

[0013] As a preferred scheme of the neuromuscular stimulator fault calibration method based on federal transfer learning, wherein: further comprising:

[0014] Before constructing the dynamic heterogeneous graph, feature extraction is performed on the device data and physiological data to generate node embedding vectors;

[0015] The feature extraction is performed on different types of nodes using neural network models suitable for the different types of nodes.

[0016] As a preferred scheme of the neuromuscular stimulator fault calibration method based on federal transfer learning, wherein: the specific processing process of the feature extraction is:

[0017] When the data associated with the node is a sequence waveform, a time convolution network is used for processing;

[0018] When the data associated with the node is a statistical value, a multilayer perceptron is used for processing, and the features of all nodes are mapped to the same dimension embedding space.

[0019] As a preferred scheme of the neuromuscular stimulator fault calibration method based on federal transfer learning, wherein: the graph neural network model is a heterogeneous graph neural network, and its processing process includes: when performing information aggregation, different parameter matrices are used for operation for different types of edges.

[0020] As a preferred scheme of the neuromuscular stimulator fault calibration method based on federal transfer learning, wherein: the heterogeneous graph neural network uses an attention mechanism, and its processing process includes:

[0021] According to the target node, the source node and the edge type between the target node and the source node, the projection operation of the edge type is performed on the embedding vector of the node to calculate the attention weight;

[0022] The projected source node is weighted by using the attention weight.

[0023] As a preferred scheme of the neural muscle stimulator fault calibration method based on federal transfer learning, the processing process of the collaborative training global model comprises:

[0024] Each neural muscle stimulator locally performs model training and uploads update information containing local training results;

[0025] The server performs weighted aggregation on the received update information from the plurality of neural muscle stimulators, and updates the global graph neural network model.

[0026] As a preferred scheme of the neural muscle stimulator fault calibration method based on federal transfer learning, the processing process of the collaborative training global model comprises:

[0027] The update information further comprises meta-information for evaluating the contribution degree of local training;

[0028] The server assigns a weight to the training result of each neural muscle stimulator according to the meta-information, and performs weighted aggregation.

[0029] As a preferred scheme of the neural muscle stimulator fault calibration method based on federal transfer learning, the processing process of the collaborative training global model comprises:

[0030] The global graph neural network model is loaded as a pre-training model;

[0031] The bottom network parameters of the pre-training model are frozen, and the unfrozen top network parameters are fine-tuned by using a local calibration data set.

[0032] As a preferred scheme of the neural muscle stimulator fault calibration method based on federal transfer learning, the processing process of the collaborative training global model comprises:

[0033] The graph-level embedding vector obtained after processing by the individualized calibration model is processed by a regression network to output a numerical calibration increment of the control parameter.

[0034] Compared with the prior art, the application has the following advantages:

[0035] 1、 The present application constructs a dynamic heterogeneous space-time graph by combining the neuromuscular stimulator with the user's physiological system, and uses a heterogeneous graph neural network for analysis, so that the model can understand the failure mechanism from the root, thereby realizing accurate calibration of the stimulation parameters, and improving the effectiveness and safety of the stimulation;

[0036] 2、 Adopting a federated learning framework allows multiple neuromuscular stimulators to train models locally using user data, and only uploads encrypted model update results that do not contain original sensitive information to the server for aggregation. Through this distributed learning paradigm, the data privacy and security risks brought about by centralized training can be fundamentally addressed. At the same time, by aggregating all model training results, the model can learn more diverse and varied failure patterns, avoiding the limitations and biases of models trained by a single device. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings. Among them:

[0038] Figure 1 The overall flowchart of the neuromuscular stimulator fault calibration method based on federated transfer learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0040] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the spirit of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0041] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in this specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0042] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0043] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0044] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integrated connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0045] Example 1

[0046] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for fault calibration of a neuromuscular stimulator based on federated transfer learning, comprising:

[0047] S1. Acquire device data and user physiological data of the neuromuscular stimulator, and construct a graph structure data representing the state of the interaction system between the neuromuscular stimulator and the user;

[0048] It should be noted that the neuromuscular stimulator (NMS) device continuously monitors the user's physiological feedback and calculates the integrated electromyography (iEMG) value of the surface electromyography (sEMG) signal in real time. When the iEMG value is lower than 70% of the historical normal range for 1 consecutive second, it is determined that an "effect deviation" event has occurred, and a complete calibration process is immediately triggered.

[0049] Specifically, within a 200ms time window prior to the triggering event, all relevant synchronization data is extracted from the data buffer, including control and internal state data (reflecting the device itself) and output and feedback data. The control and internal state data includes control parameters and internal components. Control parameters are used to read the MCU registers to obtain the currently set stimulation parameters: current amplitude and pulse width. Internal components are used to read the ADC channels related to hardware monitoring and the real-time voltage of the energy storage capacitor. The output and feedback data (reflecting device-human interaction) includes actual output and interface and physiological states. Actual output is used to read the high-speed ADC data connected in parallel with the output terminal to obtain the actual output pulse waveform. The output pulse waveform is controlled by the current amplitude and pulse width, and is affected by the real-time voltage state of the energy storage capacitor. If the energy storage capacitor ages (i.e., the real-time voltage of the energy storage capacitor decreases), then even if the settings remain unchanged, the actual pulse waveform at the output end will be deformed (e.g., amplitude decreases, tailing). The interface and physiological state are used to read external sensor data: the impedance of the electrode-skin interface and sEMG. The sEMG response of the target muscle is an sEMG waveform with a significantly reduced amplitude. The sEMG response of the target muscle is the final physiological manifestation after the actual pulse waveform at the output end acts on the human body and is attenuated by the impedance of the electrode-skin interface. The increase in impedance will directly weaken the effective stimulation intensity, resulting in a decrease in EMG response.

[0050] Furthermore, the data collected in the above process (i.e., control and internal state data, output and feedback data) are instantiated, and the node types are defined as follows: device component node (energy storage capacitor), control parameter node (current amplitude, pulse width), output signal node (pulse waveform), and physiological feedback node (sEMG, impedance of electrode-skin interface).

[0051] Specifically, based on known physical and physiological causal chains, heterogeneous edges are established to explicitly establish conditional associations between nodes:

[0052] (1) Since the control parameters can directly determine the basic shape of the output waveform, the current amplitude set by the MCU (microcontroller) of the neuromuscular stimulator and the actual pulse waveform at the output end are established as a control relationship, which is represented as the edge established between the control parameter node and the output signal node. This edge is affected by the control parameter node.

[0053] (2) Since the capacitor voltage is the source of output energy, its fluctuation will directly affect the amplitude and stability of the actual pulse. Therefore, it is necessary to establish the edge relationship between the real-time voltage of the energy storage capacitor and the actual pulse waveform at the output end. This edge relationship is represented as the actual pulse waveform at the output end under the influence of the real-time voltage of the energy storage capacitor, and is represented as the edge established between the device component node and the output signal node.

[0054] (3) Considering that the stimulation pulse itself will cause instantaneous changes in interface impedance (electrode polarization effect), it is necessary to establish the relationship between the actual pulse waveform at the output end and the electrode-skin interface impedance. This relationship represents the impedance of the electrode-skin interface under the influence of the actual pulse waveform at the output end. The edge consists of output signal nodes and physiological feedback nodes.

[0055] (4) In addition, after the actual pulse waveform passes through the skin interface (its state is reflected by the impedance of the electrode-skin interface), it will trigger the electrophysiological response of the muscle. Therefore, the relationship is established as the relationship between the actual pulse waveform at the output end and the sEMG response of the target muscle. The edge consists of the output signal node and the physiological feedback node.

[0056] It should be noted that although the composition of the edge nodes is the same in conditions (3) and (4), the relationship between the nodes is different. In condition (3), the node relationship is represented by the relationship when the output signal comes into contact with the human body, while in condition (4), the node relationship is represented by the relationship after the output signal responds to the human body.

[0057] Furthermore, nodes that meet the conditions are input into the conditions (1) to (4) above to obtain multiple heterogeneous edges and generate a heterogeneous graph, where each edge in the heterogeneous graph represents a potential fault propagation path.

[0058] Specifically, by calculating the maximum cross-correlation value between edge nodes that meet the conditions, the corresponding edge weights can be obtained, such as the maximum cross-correlation value between the actual pulse waveform and sEMG.

[0059] Furthermore, to enable heterogeneous graphs to classify heterogeneous node features, we first use a pre-trained Temporal Convolutional Network (TCN) to encode heterogeneous nodes containing waveform sequences. This TCN contains three dilated convolutional layers with a kernel size of 3 and dilation rates of 1, 2, and 4, respectively. Finally, a 64-dimensional embedding vector is output through a global average pooling layer. In addition, for heterogeneous node features containing statistical values, they are first processed into fixed-length feature vectors, and then passed through a two-layer Multilayer Perceptron (MLP). This MLP consists of an input layer, a first hidden layer, a second hidden layer, and an output layer. The first and second hidden layers each contain 128 neurons and are followed by a ReLU activation function. Through the 64-bit neurons of the output layer, the processed fixed-length feature vector is mapped into a 64-dimensional node embedding vector.

[0060] It should be noted that the generated node embedding vectors can provide a pre-processed and semantic input for heterogeneous graphs, so that heterogeneous graphs no longer need to painstakingly search for patterns from the most primitive data, but can directly infer the state of the entire system in this highly structured and information-based semantic space by analyzing the relationship between different node embedding vectors (i.e. the edges of the heterogeneous graph).

[0061] S2. Based on the federated learning framework, graph structure data is used as input, and a global graph neural network model for representing fault modes is trained collaboratively by multiple neuromuscular stimulators.

[0062] It should be noted that a global model with universal knowledge, which has been co-trained on several neuromuscular stimulators, is used as a federated learning framework. The aim is to use data from all devices in the network to co-train a powerful fault mode recognition model while protecting privacy.

[0063] Specifically, for each neuromuscular stimulator participating in the collaborative training, it uses graph structure data from the heterogeneous graph to train the model on its local model.

[0064] Specifically, the local model uses a heterogeneous graph neural network for information propagation to update the embedding vector of the actual output pulse waveform node (denoted as the target node), and the process is as follows:

[0065] The model creates a learnable parameter matrix for each defined heterogeneous edge type (represented as a triple: source node type, edge type, target node type). This parameter matrix is ​​used to perform a relation-type linear transformation (projection) on the node embedding vectors. This linear transformation includes changes to the query, key, and value. For example, heterogeneous edge 1 represents a control relationship, heterogeneous edge 2 represents an influence relationship, and W... Q,r1 ,、W K,r1 W V,r1 For heterogeneous edge 1, W Q,r2 ,、W K,r2 W V,r2 For heterogeneous edge 2, when the model needs to calculate the influence of current amplitude node characteristics (source node) on output pulse waveform node characteristics (target node), since they are control edges (type r1), the model will automatically select and use W. Q,r1 ,、W K,r1 W V,r1The original embedding vectors of the source and target nodes are transformed. When the model calculates the influence of the energy storage capacitor node features (source node) on the same output pulse waveform node features (target node), since they are physical influence edges (type r2), the model will automatically switch to and use the second set of matrices. This process is repeated, and different parameter matrices are used for different types of edges. Essentially, this allows the model to learn a specific nonlinear transformation operation for each causal relationship, so that the model can mathematically separate the logic of the control edge from the logic of the physical influence edge. This allows the model to independently and accurately learn and distinguish the information transmission patterns under different causal relationships.

[0066] Furthermore, an attention mechanism is employed on top of the heterogeneous graph neural network. Assuming the source node is connected to the target node via edge types, the model first extracts the parameter matrix corresponding to the relationship, projects the embedding vectors of the source and target nodes to generate query and key vectors, representing the target and source nodes under the heterogeneous edge type relationship. Then, the dot product similarity of the query and key vectors is calculated to obtain the original attention score. It's important to note that this attention score is no longer an importance metric, but rather represents the importance of the source node to the target node under the constraint of the heterogeneous edge type relationship, expressed as:

[0067]

[0068] Among them, e st Represented as the raw attention score, Q t Represented as a query vector, Represented as a key vector, since the similarity between the query vector and the key vector is usually calculated using a dot product, and multiplying two matrices requires that the number of columns in the first matrix must equal the number of rows in the second matrix, a transpose operation is used here, where T represents transpose; d k The dimension of the key vector;

[0069] Specifically, the Softmax function is used to normalize all the original attention scores to obtain the final attention weights. Then, the calculated attention weights are used to weight the projected source nodes to obtain the aggregated information from all source nodes connected to the target node through heterogeneous edge types.

[0070] In addition, the amount of data used for local training and the loss value of the local model are used as metadata to evaluate the contribution of the local model training.

[0071] It should be noted that the neuromuscular stimulator only retains the aggregation information of the source nodes that are heterogeneous edge types connected to the target node, while the rest of the raw data is kept in the local model.

[0072] Specifically, the neuromuscular stimulator uploads the update information to the central server. After receiving the update information from multiple neuromuscular stimulators, the server assigns weights to each neuromuscular stimulator based on the metadata (for example, the larger the amount of data or the more significant the loss reduction of the neuromuscular stimulator, the higher the weight), and performs weighted aggregation on all the update information. The aggregated result is then used to update the heterogeneous graph neural network model.

[0073] S3. On each neuromuscular stimulator, a personalized calibration model is generated through transfer learning based on a global graph neural network model.

[0074] Furthermore, the neuromuscular stimulator downloads the latest heterogeneous graph neural network model from the server and loads it locally as a pre-trained model, fully loading all weights of the heterogeneous graph neural network model and subdividing the feature extractor in the heterogeneous graph neural network model into:

[0075] The bottom feature layer is designed to learn general, cross-device physical and physiological laws (such as increased voltage leading to increased current, and muscle fatigue leading to decreased sEMG amplitude), and these contents are marked as basic knowledge.

[0076] The top feature layer is designed to learn how to combine raw node features to adapt to more complex, individual-related patterns (such as the impedance of the electrode-skin interface and the correlation pattern of sEMG when the electrode is slightly loose).

[0077] Specifically, since the effectiveness of fine-tuning is highly dependent on high-quality local data, in addition to the aforementioned update information and metadata, the dataset should also consider user feedback information (e.g., when the "stimulus is too strong / too weak", the adjustment amount of the parameter is recorded after the user reaches a satisfactory state through user feedback).

[0078] Specifically, the bottom feature layer is frozen, meaning that the gradients of these parameters are always zero when calculating gradients during backpropagation, to ensure that all parameters of the top feature layer are trainable, i.e., fine-tuned.

[0079] Specifically, multiple rounds of training are performed on the local calibration dataset, with Adam selected as the optimizer and a learning rate of 1e-4. In each round, the data is fed into the model in batches. For each batch:

[0080] Perform forward propagation to obtain the predicted neuromuscular stimulator calibration value;

[0081] Calculate the loss for the current batch;

[0082] Perform backpropagation to compute the gradient of the loss with respect to all trainable parameters;

[0083] The Adam optimizer updates the trainable parameters of the lower-level feature layers based on the calculated gradients.

[0084] Specifically, the training process can continue for a fixed number of rounds, or until the loss on the validation set no longer decreases, to obtain a personalized neuromuscular stimulator calibration model.

[0085] It should be noted that, through the calibration model of the personalized neuromuscular stimulator, it is possible to find the physiological characteristics that match each different user, and to adaptively adjust the calibration values ​​of the neuromuscular stimulator control parameters based on these physiological characteristics.

[0086] S4. Using a personalized calibration model, process the graph structure data corresponding to the current system state and output calibration values ​​for the control parameters of the neuromuscular stimulator.

[0087] Furthermore, when the device triggers the calibration process again, it first executes the graph structure data construction process, which constructs a graph structure data by combining the currently collected device and physiological data. Then, the graph structure data is input into the trained personalized calibration model for a complete forward propagation.

[0088] Specifically, the forward propagation passes through the feature extractor part of the model (which includes a frozen bottom feature layer and a fine-tuned top feature layer) to generate an updated, context-rich embedding vector for each node in the heterogeneous graph.

[0089] In addition, in order to obtain a representation that can represent the current "health status" of the neuromuscular stimulator, average pooling is used, which is to sum the final embedding vectors of all nodes in the heterogeneous graph and then average them to output the graph-level embedding vector.

[0090] Specifically, the graph-level embedding vector is fed into a regression network for processing. For the compatibility of the model components, this regression network, like the heterogeneous graph mentioned above, uses a multilayer perceptron. It includes an input layer with 64 neurons to receive the graph-level embedding vector; a hidden layer with 32 neurons and ReLU as the activation function; and an output layer with 2 neurons, corresponding to the two control parameters that need to be calibrated, namely the current amplitude and the pulse width. The output layer does not use any activation function, ensuring that its output can be a real number in any range, which can accurately represent the increase or decrease of the control parameters.

[0091] Specifically, when a fault occurs, the MCU (microcontroller) of the neuromuscular stimulator receives the increase or decrease of the adjusted control parameters and immediately performs an update operation on the control parameters, thereby achieving fault calibration of the neuromuscular stimulator.

[0092] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0097] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A fault calibration method for a neuromuscular stimulator based on federated transfer learning, characterized in that, include: Acquire device data and user physiological data of the neuromuscular stimulator, and construct a graph structure data representing the state of the interaction system between the neuromuscular stimulator and the user; Based on the federated learning framework, the graph structure data is used as input, and a global graph neural network model for representing fault modes is trained collaboratively by multiple neuromuscular stimulators. On each neuromuscular stimulator, a personalized calibration model is generated through transfer learning based on the global graph neural network model. Using the personalized calibration model, the graph structure data corresponding to the current system state is processed, and calibration values ​​for the control parameters of the neuromuscular stimulator are output.

2. The neuromuscular stimulator fault calibration method based on federated transfer learning as described in claim 1, characterized in that, The construction of the graph structure data characterizing the state of the neuromuscular stimulator and user interaction system includes: The preset entities in the device data and physiological data are defined as nodes of various types, and heterogeneous edges connecting the nodes are defined according to the preset relationships between the preset entities, thereby constructing a dynamic heterogeneous graph as the graph structure data.

3. The method for fault calibration of neuromuscular stimulators based on federated transfer learning as described in claim 2, characterized in that, Also includes: Before constructing the dynamic heterogeneous graph, feature extraction is performed on the device data and physiological data to generate node embedding vectors; The feature extraction is performed using a neural network model adapted to different types of nodes.

4. The neuromuscular stimulator fault calibration method based on federated transfer learning as described in claim 3, characterized in that, The specific processing procedure for feature extraction is as follows: When the data associated with a node is a sequence waveform, a temporal convolutional network is used for processing; When the data associated with a node is a statistical value, a multilayer perceptron is used to process it, mapping the features of all nodes to the same-dimensional embedding space.

5. The neuromuscular stimulator fault calibration method based on federated transfer learning as described in claim 1, characterized in that, The graph neural network model is a heterogeneous graph neural network, and its processing includes: when performing information aggregation, different parameter matrices are used for operations on different types of edges.

6. The neuromuscular stimulator fault calibration method based on federated transfer learning as described in claim 5, characterized in that, The heterogeneous graph neural network employs an attention mechanism, and its processing includes: Based on the target node, source node, and the edge type between the target node and the source node, the embedding vector of the node is projected using the edge type to calculate the attention weights. The projected source nodes are weighted using the attention weights.

7. The method for fault calibration of a neuromuscular stimulator based on federated transfer learning as described in claim 1, characterized in that, The process of collaboratively training the global model includes: Each neuromuscular stimulator trains the model locally and uploads updated information containing the local training results; The server performs weighted aggregation of the received update information from multiple neuromuscular stimulators and updates the global graph neural network model accordingly.

8. The neuromuscular stimulator fault calibration method based on federated transfer learning as described in claim 7, characterized in that, The weighted aggregation process is as follows: The updated information also includes metadata used to assess the contribution of local training; The server assigns weights to the training results of each neuromuscular stimulator based on the metadata and performs weighted aggregation.

9. The method for fault calibration of a neuromuscular stimulator based on federated transfer learning as described in claim 1, characterized in that, The process of generating a personalized calibration model through transfer learning includes: The global graph neural network model is loaded as a pre-trained model; The underlying network parameters of the pre-trained model are frozen, and the unfrozen top-level network parameters are fine-tuned using a local calibration dataset.

10. The method for fault calibration of a neuromuscular stimulator based on federated transfer learning as described in claim 1, characterized in that, The processing of the output calibration value includes: The graph-level embedding vector obtained after processing by the personalized calibration model is processed by a regression network to output the numerical calibration increment of the control parameters.

Images