Cross-domain metric matrix based zero sample matching method and system for time-series physiological signals

Abstract

The application discloses a kind of zero sample matching methods and systems of time physiological signal based on cross-domain measure matrix, it is related to physiological signal decoding field, the method is mapped into high-dimensional physiological feature matrix by pre-training physiological feature coding network, to be decoded time physiological signal;Through pre-training semantic feature coding network, the dynamic candidate instruction set containing several candidate control instructions is converted into the high-dimensional instruction semantic reference matrix set corresponding to each instruction;Based on the preset cross-domain feature measurement algorithm, the feature matching degree is calculated, the inference stage cross-domain measure matrix is generated, the target semantic reference matrix is determined by extreme value optimization and the corresponding control instruction is output.This scheme does not need to update network weight, does not need to retrain sample, can realize new instruction zero sample decoding, break through the limitation of traditional classification model fixed class, solve the problem of cross-modal time sequence misalignment, greatly reduce the threshold of equipment use, significantly improve decoding accuracy, generalization ability and system robustness.

Description

Technical Field

[0001] This application relates to the field of physiological signal decoding technology, and in particular to a zero-sample matching method and system for time-series physiological signals based on cross-domain metric matrices. Background Technology

[0002] With the rapid development of brain-computer interfaces (BCI) and wearable physiological sensing technologies, human-computer interaction, digital therapy, and intelligent control systems based on physiological signals such as electroencephalograms (EEGs) have been widely applied in consumer and medical settings. The core of these technologies' practical application lies in accurately translating the user's implicit intentions into explicit control commands.

[0003] Existing EEG intention decoding systems generally employ classification models or conventional clustering algorithms based on fixed output dimensions. These systems can only predict within a pre-defined, limited set of categories and cannot recognize entirely new semantic commands not included in the original training, resulting in extremely poor system scalability. Adding new control commands requires forcing users to undergo prolonged, high-intensity directional physiological signal sample collection, followed by model retraining and weight fine-tuning. This process is not only cumbersome and has a high barrier to entry for users, but it also easily induces cognitive fatigue, leading to a decline in the quality of the collected signals and further affecting the model's decoding accuracy.

[0004] Meanwhile, existing cross-modal decoding schemes for time-series physiological signals suffer from problems such as temporal misalignment, high computational consumption, and distorted decoding results. Furthermore, existing schemes fail to form a complete closed loop between the training logic for feature space alignment and the zero-sample matching in the inference stage. The feature alignment target in the training stage is disconnected from the matching logic in the inference stage, resulting in insufficient model generalization ability and difficulty in meeting the actual application needs of medical and consumer scenarios. Summary of the Invention

[0005] The purpose of this application is to provide a zero-shot matching method and system for time-series physiological signals based on cross-domain metric matrices, which can achieve zero-shot adaptive decoding of new instructions without collecting new training samples or updating network weights.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a zero-sample matching method for time-series physiological signals based on a cross-domain metric matrix, comprising the following steps: The time-series physiological signal to be decoded is acquired, and the time-series physiological signal is input into the pre-trained physiological feature encoding network to generate a high-dimensional physiological feature matrix.

[0007] A dynamic candidate instruction set is obtained and input into a pre-trained semantic feature encoding network to obtain a high-dimensional instruction semantic benchmark matrix set. The dynamic candidate instruction set includes several candidate control instructions. Each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set corresponds one-to-one with each candidate control instruction in the dynamic candidate instruction set.

[0008] Based on a pre-defined cross-domain feature measurement algorithm, the feature matching degree between the high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set is calculated, and a cross-domain measurement matrix for the inference stage is generated.

[0009] The cross-domain metric matrix in the inference stage is subjected to extreme value optimization processing to determine the high-dimensional instruction semantic benchmark matrix corresponding to the optimal matching extreme value, which is used as the target high-dimensional instruction semantic benchmark matrix, and the candidate control instructions corresponding to the target high-dimensional instruction semantic benchmark matrix are output.

[0010] Optionally, before acquiring the time-series physiological signal to be decoded, inputting the time-series physiological signal into the pre-trained physiological feature encoding network, and generating a high-dimensional physiological feature matrix, the method further includes the following steps: Obtain the training sample set; the training sample set includes N pairs of time-series physiological signal samples and candidate control command samples.

[0011] Each time-series physiological signal sample is input into the physiological feature encoding network to be trained, generating N corresponding high-dimensional physiological feature matrices.

[0012] Each candidate control instruction sample is input into the semantic feature encoding network to be trained, generating N corresponding high-dimensional instruction semantic benchmark matrices.

[0013] Based on the preset cross-domain feature measurement algorithm, the feature matching degree between each high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix is ​​calculated, and an N×N dimension cross-domain measurement matrix for the training phase is generated.

[0014] The goal is to optimize the training phase by finding the extreme values ​​of the diagonal elements of the cross-domain metric matrix. The network weights of the physiological feature encoding network and the semantic feature encoding network are then updated until the training convergence condition is met.

[0015] Optionally, the cross-domain feature measurement algorithm is a spatial distance measurement function or a statistical similarity evaluation function; the spatial distance measurement function includes at least one of the mean squared error function and the L1 norm function, and the statistical similarity evaluation function includes at least one of the Pearson correlation coefficient function and the cosine similarity function.

[0016] Optionally, the dimensions of the high-dimensional physiological feature matrix and the high-dimensional instruction semantic benchmark matrix are both 1 to M, where M is an integer greater than or equal to 2, and the dimensions of the high-dimensional physiological feature matrix and the high-dimensional instruction semantic benchmark matrix are the same.

[0017] Optionally, both the high-dimensional physiological feature matrix and the high-dimensional instruction semantic benchmark matrix are two-dimensional matrices of T×F dimensions, where T is the time frame dimension and F is the feature channel dimension or frequency dimension.

[0018] Optionally, the dynamic candidate instruction set includes M candidate instructions, where M is an integer greater than or equal to 1; the cross-domain metric matrix during the inference phase is a 1×M dimension matrix.

[0019] Optionally, the optimal matching extreme value is the minimum value of feature difference or the maximum value of feature correlation.

[0020] Optionally, the time-series physiological signals include at least one of electroencephalogram (EEG) signals, electromyogram (EMG) signals, electrooculogram (EOG) signals, and functional near-infrared spectroscopy signals.

[0021] Optionally, the physiological feature encoding network is a neural network containing temporal convolutional layers or attention mechanisms; the semantic feature encoding network is a natural language processing model.

[0022] Secondly, this application provides a zero-sample matching system for time-series physiological signals based on a cross-domain metric matrix, comprising the following functional modules: The physiological feature matrix acquisition module is used to acquire the time-series physiological signal to be decoded, input the time-series physiological signal into the pre-trained physiological feature encoding network, and generate a high-dimensional physiological feature matrix.

[0023] The semantic benchmark matrix acquisition module is used to acquire a dynamic candidate instruction set. The dynamic candidate instruction set is input into a pre-trained semantic feature encoding network to obtain a high-dimensional instruction semantic benchmark matrix set. The dynamic candidate instruction set includes several candidate control instructions. Each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set corresponds one-to-one with each candidate control instruction in the dynamic candidate instruction set.

[0024] The cross-domain metric matrix generation module is used to calculate the feature matching degree between the high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set based on a preset cross-domain feature metric algorithm, and generate the cross-domain metric matrix for the inference stage.

[0025] The extreme value optimization decoding output module is used to perform extreme value optimization processing on the cross-domain metric matrix during the inference stage, determine the high-dimensional instruction semantic benchmark matrix corresponding to the optimal matching extreme value, use it as the target high-dimensional instruction semantic benchmark matrix, and output the candidate control instructions corresponding to the target high-dimensional instruction semantic benchmark matrix.

[0026] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a zero-shot matching method and system for time-series physiological signals based on cross-domain metric matrices. The method acquires the time-series physiological signal to be decoded and inputs it into a pre-trained physiological feature encoding network to generate a high-dimensional physiological feature matrix. This accurately extracts the latent neural patterns of the time-series physiological signal, mapping the physiological signal to a standardized high-dimensional feature space, laying a unified feature benchmark for subsequent cross-domain matching. Simultaneously, it avoids redundant feature extraction computations in the real-time inference stage, ensuring low-latency processing. By acquiring a dynamic candidate instruction set and inputting it into a pre-trained semantic feature encoding network, a high-dimensional instruction semantic benchmark matrix set corresponding one-to-one with each candidate control instruction in the instruction set is obtained. This allows for unlimited access to dynamic instructions containing entirely new, untrained categories, enabling instruction expansion without modifying the network structure and weights, completely breaking away from the fixed output class of traditional classification models. Other limitations; by calculating the feature matching degree between the physiological feature matrix and each semantic benchmark matrix through a preset cross-domain feature measurement algorithm, a cross-domain measurement matrix is ​​generated in the inference stage. The matching relationship between physiological signals and instruction semantics can be quantified in the shared latent space, transforming cross-modal intent decoding into a quantifiable matrix numerical matching problem, effectively solving the problem of temporal misalignment between temporal physiological signals and semantic instructions, and improving matching robustness; finally, by performing extreme value optimization processing on the cross-domain measurement matrix in the inference stage, the target semantic benchmark matrix corresponding to the optimal matching extreme value is determined and its corresponding candidate control instruction is output. Zero-sample intent decoding can be completed without collecting new training samples or updating network weights, completely eliminating the cumbersome targeted data collection and model retraining process of traditional solutions, greatly reducing the threshold for using brain-computer interface devices, while ensuring high accuracy of intent parsing. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram illustrating the technical route of a zero-sample matching method for time-series physiological signals based on a cross-domain metric matrix, provided in one embodiment of this application.

[0029] Figure 2 This is a flowchart of the training phase in a zero-sample matching method for time-series physiological signals based on a cross-domain metric matrix, provided as an embodiment of this application.

[0030] Figure 3 This is a flowchart of the inference stage in a zero-sample matching method for time-series physiological signals based on a cross-domain metric matrix, provided as an embodiment of this application.

[0031] Figure 4 This is a schematic diagram of the functional modules of a zero-sample matching system for time-series physiological signals based on a cross-domain metric matrix, provided as an embodiment of this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] The technical solution provided in this application consists of two core stages: the first stage is the feature encoding network training stage, which is used to complete the alignment of the shared latent space of physiological features and semantic features, and outputs the pre-trained physiological feature encoding network and semantic feature encoding network. The core is diagonal extremum optimization based on an N×N cross-domain metric matrix. The second stage is the zero-shot matching inference stage, which, based on the pre-trained dual-feature encoder, realizes adaptive decoding of new instructions without retraining. The core is global extremum optimization of a 1×M matrix. The two stages form a complete technical closed loop. The pre-trained network output from the first stage is the core execution body of the second stage, and the diagonal optimization target of the first stage provides the core guarantee for the extremum matching accuracy of the second stage.

[0034] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] This application provides a zero-sample matching method for time-series physiological signals based on a cross-domain metric matrix, such as... Figure 1 As shown, this method can be divided into a training phase and an inference phase. For the training phase, the execution entity is a computer device with computing power, including but not limited to PCs, cloud servers, and high-performance edge computing terminals. In an exemplary embodiment, as... Figure 2 As shown, the training phase of this method includes the following steps: A1. Obtain the training sample set; the training sample set includes N pairs of time-series physiological signal samples and candidate control command samples.

[0036] Specifically, in each pair of samples, the temporal physiological signal sample is the physiological signal data collected when the user performs the intentional task of executing the corresponding semantic command, including but not limited to 0-125Hz full-band EEG signals, EMG signals, EOG signals, and functional near-infrared spectroscopy (fNIRS) signals; the candidate control command sample is the control command corresponding to the physiological signal sample, and the two are paired in a positive sample relationship. The N sets of samples in the training sample set cover the preset basic command categories and are used to complete the feature space alignment learning of the dual-feature encoder.

[0037] A2. Input the physiological signal samples of each time series into the physiological feature encoding network to be trained to generate the corresponding N high-dimensional physiological feature matrices.

[0038] Specifically, the physiological feature encoding network employs a neural network containing temporal convolutional layers or attention mechanisms, such as the TCN temporal convolutional network or the Transformer network. The network input is a temporal physiological signal sample, and the output is a high-dimensional physiological feature matrix of 1-M dimensions, where M is an integer greater than or equal to 2. In this embodiment, a T×F two-dimensional matrix is ​​used as a preferred embodiment, where T is the time frame dimension and F is the feature channel dimension or frequency dimension. That is, each set of temporal physiological signal samples generates a set of T×F high-dimensional physiological feature matrices, and N sets of samples generate N sets of T×F high-dimensional physiological feature matrices, with an overall dimension of N×T×F.

[0039] A3. Input each candidate control instruction sample into the semantic feature encoding network to be trained to generate N corresponding high-dimensional instruction semantic benchmark matrices.

[0040] Specifically, the semantic feature encoding network employs a pre-trained natural language processing model, such as BERT or GPT. The network input is the text data of candidate control command samples, and the output is a high-dimensional command semantic benchmark matrix with the same dimensions as the high-dimensional physiological feature matrix. In this embodiment, each group of candidate control command samples generates a set of T×F-dimensional high-dimensional command semantic benchmark matrices, and N groups of samples generate a total of N sets of T×F-dimensional high-dimensional command semantic benchmark matrices, with an overall dimension of N×T×F.

[0041] A4. Based on the preset cross-domain feature measurement algorithm, calculate the feature matching degree between each high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix, and generate an N×N dimension cross-domain measurement matrix for the training phase.

[0042] Specifically, cross-domain feature measurement algorithms are used to quantify the degree of matching between two sets of feature matrices in a shared latent space. These include spatial distance measurement functions to characterize feature differences and statistical similarity evaluation functions to characterize feature correlations. Spatial distance measurement functions include, but are not limited to, the mean squared error (MSE) function and the L1 norm function; statistical similarity evaluation functions include, but are not limited to, the Pearson correlation coefficient (PCC) function and the cosine similarity function.

[0043] In this embodiment, the mean squared error function is used as the cross-domain feature measurement algorithm. The formula for calculating the matching degree value in the N×N dimension cross-domain measurement matrix during the training phase is as follows: .

[0044] in, For the first i The high-dimensional physiological feature matrix and the first j The matching degree value of a high-dimensional instruction semantic benchmark matrix, i.e. the mean squared error loss value, the smaller the value, the smaller the difference between the two sets of features and the higher the matching degree. For the first i A high-dimensional physiological feature matrix at time index t Feature Index f The value at that location; For the first j A high-dimensional instruction semantic benchmark matrix at time index t Feature Index f The value at that location; T The time frame dimension of the feature matrix. F The feature channel dimension of the feature matrix; t For the summation iteration index of the time axis, f This is the summation iteration index for the characteristic axis.

[0045] By iterating through and calculating the pairwise matching degree values ​​of N sets of high-dimensional physiological feature matrices and N sets of high-dimensional instruction semantic benchmark matrices, an N×N dimension cross-domain metric matrix for the training phase is generated. The diagonal elements in the matrix are the matching degree values ​​of paired positive samples, and the off-diagonal elements are the matching degree values ​​of unpaired negative samples.

[0046] Although the above embodiments use a T×F two-dimensional matrix as an example, the solution of this application can be applied to high-dimensional matrices of 3 dimensions and above. For high-dimensional feature matrices of 3 dimensions and above, the calculation logic of this embodiment can be directly extended: include all dimensional elements of the feature matrix in the summation calculation range, traverse and calculate the overall matching degree of the two sets of multi-dimensional feature matrices, and generate the cross-domain metric matrix of the corresponding dimension. Without modifying the core calculation logic, full-scene adaptation of 1-M dimensions can be achieved.

[0047] A5. Optimize the training phase by ensuring the diagonal elements of the cross-domain metric matrix reach their extreme values. Update the network weights of the physiological feature encoding network and the semantic feature encoding network until the training convergence condition is met. Specifically, when using a spatial distance metric function, the optimization objective is to minimize the diagonal elements and maximize the off-diagonal elements of the cross-domain metric matrix; when using a statistical similarity evaluation function, the optimization objective is to maximize the diagonal elements and minimize the off-diagonal elements of the cross-domain metric matrix.

[0048] In an exemplary embodiment, extreme value optimization matching is performed on the cross-domain metric matrix generated above during the training phase to find the semantic target with the smallest difference from the current physiological feature matrix. The formula for the minimum value matching condition is as follows: .

[0049] in, Indicates for the first i A high-dimensional physiological feature matrix is ​​used as the target index of the high-dimensional instruction semantic benchmark matrix with minimal loss value matched in the candidate set; N is the total number of instructions in the dynamic candidate control instruction set. Simultaneously, through built-in diagonal matching judgment logic, its verification formula is an indicator function: .

[0050] when When the physiological signal and the corresponding true intention semantic benchmark are perfectly aligned diagonally, the match is successful; if the match fails, the system reduces the accuracy. Acc During the training phase, based on the diagonal matching results described above, the overall accuracy of zero-shot intent decoding can be quickly output using the following formula: .

[0051] in, Acc The zero-sample matching accuracy of the model is the overall accuracy. This metric directly reflects the discriminative performance of the system's dual-stream feature encoding network, enabling a quantitative and intuitive evaluation of the generalization support capability for new instruction categories without the need for fine-tuning the network weights.

[0052] In this embodiment, taking the mean squared error function as an example, the optimization objective is to compress the loss value of the diagonal elements. The network weights of the two feature encoders are updated through the backpropagation algorithm until the training loss value is lower than the preset threshold, or the matching accuracy reaches the preset standard, or the number of iterations reaches the preset upper limit. The training convergence condition is then determined to be met, and the pre-trained physiological feature encoding network and semantic feature encoding network are output.

[0053] In the inference phase of this method, the execution entities are a multimodal physiological signal acquisition device and an edge computing terminal or cloud server. The multimodal physiological signal acquisition device is responsible for the acquisition and preprocessing of time-series physiological signals. The edge computing terminal or cloud server acquires the time-series physiological signals to be decoded, loads the pre-trained dual-feature encoder, and is responsible for the core feature encoding, matrix calculation, and instruction decoding. In an exemplary embodiment, such as... Figure 3 As shown, the reasoning phase of this method includes the following steps: B1. Obtain the time-series physiological signal to be decoded, input the time-series physiological signal into the pre-trained physiological feature encoding network, and generate a high-dimensional physiological feature matrix.

[0054] Specifically, the time-series physiological signals are collected in real time by a wearable multimodal physiological signal acquisition device. The microprocessor built into the acquisition device performs analog-to-digital conversion (ADC), noise reduction filtering, and data packetization on the acquired raw signals. The signals are then transmitted to the edge computing terminal via wireless communication methods such as Bluetooth Low Energy and Wi-Fi to obtain the time-series physiological signals to be decoded.

[0055] The time-series physiological signal to be decoded is input into the pre-trained physiological feature encoding network to generate a high-dimensional physiological feature matrix with the same dimensions as the training phase. In this embodiment, a two-dimensional matrix with T×F dimensions is preferred. The time-series physiological signal to be decoded corresponds to one set of high-dimensional physiological feature matrices with T×F dimensions, and the overall dimension is 1×T×F.

[0056] B2. Obtain the dynamic candidate instruction set. Input the dynamic candidate instruction set into the pre-trained semantic feature encoding network to obtain a high-dimensional instruction semantic benchmark matrix set. The dynamic candidate instruction set includes several candidate control instructions. Each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set corresponds one-to-one with each candidate control instruction in the dynamic candidate instruction set.

[0057] Specifically, the dynamic candidate instruction set is a dynamically updated set of candidate control instructions, which can contain any number of entirely new categories of instructions that did not participate in the dual encoder training process described above. The number of instructions is denoted as M, where M is an integer greater than or equal to 1. Each candidate control instruction in the dynamic candidate instruction set is input into the pre-trained semantic feature encoding network to generate M sets of high-dimensional instruction semantic benchmark matrices with the same dimensions as the high-dimensional physiological feature matrices (T×F). All high-dimensional instruction semantic benchmark matrices constitute a set of high-dimensional instruction semantic benchmark matrices, with an overall dimension of M×T×F.

[0058] B3. Based on the preset cross-domain feature measurement algorithm, calculate the feature matching degree between the high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set, and generate the cross-domain measurement matrix for the inference stage.

[0059] Specifically, the cross-domain feature measurement algorithm used in this step is consistent with that used in the training phase, ensuring the uniformity of the feature matching degree calculation logic. In this embodiment, taking the mean squared error function as an example, the matching degree value in the cross-domain measurement matrix during the inference phase is calculated using the following formula: .

[0060] in, The high-dimensional physiological feature matrix to be decoded and the first m The matching degree values ​​of the high-dimensional instruction semantic baseline matrix are calculated by traversing and calculating the high-dimensional physiological feature matrix to be decoded. M The matching degree values ​​of the high-dimensional instruction semantic benchmark matrix are used to generate a 1× M The cross-domain metric matrix for the dimensional reasoning stage, where each element corresponds to the matching degree loss value between the physiological feature matrix and the semantic benchmark matrix of a single instruction.

[0061] B4. Perform extreme value optimization on the cross-domain metric matrix during the inference stage to determine the high-dimensional instruction semantic benchmark matrix corresponding to the optimal matching extreme value, which is used as the target high-dimensional instruction semantic benchmark matrix, and output the candidate control instructions corresponding to the target high-dimensional instruction semantic benchmark matrix.

[0062] Specifically, a global extremum search is performed on the 1×M-dimensional cross-domain metric matrix. The optimal matching extremum is determined based on the cross-domain feature metric algorithm used: when using a spatial distance metric function, the optimal matching extremum is the minimum feature difference; when using a statistical similarity evaluation function, the optimal matching extremum is the maximum feature correlation. In this embodiment, when using the mean squared error function, the formula for calculating the minimum matching is: .

[0063] In the formula, The target index is the high-dimensional instruction semantic benchmark matrix that has the highest matching degree with the high-dimensional physiological feature matrix to be decoded. M Let be the total number of instructions in the dynamic candidate instruction set. Using the above formula, the target index corresponding to the minimum difference value is determined, and then the high-dimensional instruction semantic benchmark matrix corresponding to this target index is determined as the target high-dimensional instruction semantic benchmark matrix.

[0064] Subsequently, based on the target index corresponding to the target high-dimensional instruction semantic benchmark matrix, the corresponding control instruction is extracted from the dynamic candidate instruction set and output as a decoded control instruction to the controlled execution terminal. The controlled execution terminal includes, but is not limited to, smart home devices, human-computer interaction interfaces, and digital therapy devices, and the controlled execution terminal executes the corresponding action according to the decoded control instruction.

[0065] In an optional embodiment, for medical-grade silent communication and voice prosthesis scenarios, the dynamic candidate instruction set can be replaced with an audio feature library of standard pronunciations, and the semantic feature encoding network can be replaced with an audio feature encoding network. By calculating the cross-domain metric matrix between the physiological feature matrix and each standard audio feature matrix, extreme value optimization matching is completed, and the standard pronunciation with the highest matching degree is directly output. There is no need to perform high-computational speech generation, and the clarity of the output speech is guaranteed with extremely low latency.

[0066] In an optional embodiment, for application scenarios of high-dimensional feature matrices, the physiological feature encoding network can output a high-dimensional physiological feature matrix of 3 dimensions or higher, and the semantic feature encoding network can output a high-dimensional instruction semantic benchmark matrix of the same dimension. The cross-domain feature measurement algorithm traverses and calculates the matching degree of all-dimensional elements to generate a cross-domain measurement matrix of the corresponding dimension. Without modifying the core matching logic, it can achieve full-scenario adaptation of 1-M dimensions.

[0067] Based on the same inventive concept, this application also provides a system for implementing the aforementioned zero-sample matching method for time-series physiological signals based on cross-domain metric matrices. The solution provided by this system is similar to the implementation described in the above method. In one embodiment, such as... Figure 4 As shown, a zero-sample matching system for time-series physiological signals based on a cross-domain metric matrix is ​​provided, including the following functional modules: The physiological feature matrix acquisition module is used to acquire the time-series physiological signal to be decoded, input the time-series physiological signal into the pre-trained physiological feature encoding network, and generate a high-dimensional physiological feature matrix.

[0068] The semantic benchmark matrix acquisition module is used to acquire a dynamic candidate instruction set. The dynamic candidate instruction set is input into a pre-trained semantic feature encoding network to obtain a high-dimensional instruction semantic benchmark matrix set. The dynamic candidate instruction set includes several candidate control instructions. Each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set corresponds one-to-one with each candidate control instruction in the dynamic candidate instruction set.

[0069] The cross-domain metric matrix generation module is used to calculate the feature matching degree between the high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set based on a preset cross-domain feature metric algorithm, and generate the cross-domain metric matrix for the inference stage.

[0070] The extreme value optimization decoding output module is used to perform extreme value optimization processing on the cross-domain metric matrix during the inference stage, determine the high-dimensional instruction semantic benchmark matrix corresponding to the optimal matching extreme value, use it as the target high-dimensional instruction semantic benchmark matrix, and output the candidate control instructions corresponding to the target high-dimensional instruction semantic benchmark matrix.

[0071] In addition to the above, the system hardware includes physiological signal acquisition devices, edge computing terminals, and controlled execution terminals. The physiological signal acquisition devices employ flexible wearable structures, such as EEG headbands and patch sensors. The edge computing terminal, serving as the underlying physical support carrier of the system, includes, but is not limited to, smartphones, tablets, PCs, and cloud servers. It establishes a bidirectional data connection with the main control and wireless transmission modules via a wireless communication network, serving as the computational carrier for the core algorithms and executing the entire process of feature encoding network training and zero-shot matching inference. The controlled execution terminal communicates with the edge computing terminal and is used to receive decoded control commands and execute corresponding device actions.

[0072] The zero-shot matching method and system for temporal physiological signals based on cross-domain metric matrices provided in the above embodiments of this application, by constructing a dual-stream mapping architecture of physiological features and semantic features, maps temporal physiological signals and dynamic candidate instruction sets to a unified shared latent space. This breaks through the limitation of fixed output dimensions in traditional classification models, allowing support for any number of new instructions without modifying the network structure, thus achieving unlimited instruction expansion capabilities. Through the construction of cross-domain metric matrices and the extreme value optimization mechanism, adaptive matching of new instructions can be completed without collecting new training samples or updating network weights, completely eliminating the cumbersome directional data collection and model retraining process in traditional solutions, significantly reducing the usage threshold of brain-computer interface devices, and avoiding cognitive fatigue for users. Through a generalized cross-domain feature metric algorithm, it can flexibly adapt to different signal qualities and application scenarios, effectively resisting individual physiological differences and environmental noise interference, and improving decoding accuracy and system robustness. At the same time, through the diagonal feature optimization mechanism, it solves the temporal misalignment problem in cross-modal decoding of temporal physiological signals and provides a quantifiable zero-shot matching accuracy index, providing an intuitive standard for system performance evaluation.

[0073] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the data protection laws and policies of the relevant countries.

[0074] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A zero-sample matching method for time-series physiological signals based on cross-domain metric matrices, characterized in that, include: Obtain the training sample set; The training sample set includes N pairs of time-series physiological signal samples and candidate control command samples; Each time-series physiological signal sample is input into the physiological feature encoding network to be trained, generating N corresponding high-dimensional physiological feature matrices; Each candidate control instruction sample is input into the semantic feature encoding network to be trained to generate N corresponding high-dimensional instruction semantic benchmark matrices; Based on the preset cross-domain feature measurement algorithm, the feature matching degree between each high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix is ​​calculated, and an N×N dimension cross-domain measurement matrix for the training phase is generated. The network weights of the physiological feature encoding network and the semantic feature encoding network are updated with the goal of optimizing the diagonal elements of the cross-domain metric matrix during the training phase to reach their extreme values, until the training convergence condition is met. Acquire the time-series physiological signal to be decoded, input the time-series physiological signal into a pre-trained physiological feature encoding network, and generate a high-dimensional physiological feature matrix; A dynamic candidate instruction set is obtained, and the dynamic candidate instruction set is input into a pre-trained semantic feature encoding network to obtain a high-dimensional instruction semantic benchmark matrix set; the dynamic candidate instruction set includes several candidate control instructions; each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set corresponds one-to-one with each candidate control instruction in the dynamic candidate instruction set. Based on a preset cross-domain feature measurement algorithm, the feature matching degree between the high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix in the set of high-dimensional instruction semantic benchmark matrices is calculated to generate a cross-domain measurement matrix for the inference stage; the high-dimensional physiological feature matrix and the high-dimensional instruction semantic benchmark matrix are both two-dimensional matrices of T×F dimensions, where T is the time frame dimension and F is the feature channel dimension or frequency dimension. The cross-domain metric matrix in the inference stage is subjected to extreme value optimization processing to determine the high-dimensional instruction semantic benchmark matrix corresponding to the optimal matching extreme value, which is used as the target high-dimensional instruction semantic benchmark matrix, and the candidate control instruction corresponding to the target high-dimensional instruction semantic benchmark matrix is ​​output; the cross-domain feature measurement algorithm is a spatial distance metric function or a statistical similarity evaluation function; the spatial distance metric function includes at least one of the mean square error function and the L1 norm function, and the statistical similarity evaluation function includes at least one of the Pearson correlation coefficient function and the cosine similarity function; The dimensions of the high-dimensional physiological feature matrix and the high-dimensional instruction semantic reference matrix are both 1 to M, where M is an integer greater than or equal to 2, and the dimensions of the high-dimensional physiological feature matrix and the high-dimensional instruction semantic reference matrix are the same.

2. The zero-sample matching method for time-series physiological signals based on cross-domain metric matrices according to claim 1, characterized in that, The dynamic candidate instruction set includes M candidate control instructions, where M is an integer greater than or equal to 1; the cross-domain metric matrix of the inference stage is a 1×M dimension matrix.

3. The zero-sample matching method for time-series physiological signals based on cross-domain metric matrices according to claim 1, characterized in that, The optimal matching extreme value is either the minimum value of feature difference or the maximum value of feature correlation.

4. The zero-sample matching method for time-series physiological signals based on cross-domain metric matrices according to claim 1, characterized in that, The time-series physiological signals include at least one of electroencephalogram (EEG) signals, electromyogram (EMG) signals, electrooculogram (EOG) signals, and functional near-infrared spectroscopy signals.

5. The zero-sample matching method for time-series physiological signals based on cross-domain metric matrices according to claim 1, characterized in that, The physiological feature encoding network is a neural network containing temporal convolutional layers or attention mechanisms; the semantic feature encoding network is a natural language processing model.

6. A zero-sample matching system for time-series physiological signals based on a cross-domain metric matrix, characterized in that, The system for implementing the method as described in any one of claims 1-5 includes: The physiological feature matrix acquisition module is used to acquire the time-series physiological signal to be decoded, input the time-series physiological signal into the pre-trained physiological feature encoding network, and generate a high-dimensional physiological feature matrix. The semantic benchmark matrix acquisition module is used to acquire a dynamic candidate instruction set, input the dynamic candidate instruction set into a pre-trained semantic feature encoding network, and obtain a high-dimensional instruction semantic benchmark matrix set; the dynamic candidate instruction set includes several candidate control instructions; each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set corresponds one-to-one with each candidate control instruction in the dynamic candidate instruction set. The cross-domain metric matrix generation module is used to calculate the feature matching degree between the high-dimensional physiological feature matrix and each high-dimensional instruction semantic benchmark matrix in the high-dimensional instruction semantic benchmark matrix set based on a preset cross-domain feature metric algorithm, and generate the cross-domain metric matrix for the inference stage. The extreme value optimization decoding output module is used to perform extreme value optimization processing on the cross-domain metric matrix of the inference stage, determine the high-dimensional instruction semantic benchmark matrix corresponding to the optimal matching extreme value, use it as the target high-dimensional instruction semantic benchmark matrix, and output the candidate control instruction corresponding to the target high-dimensional instruction semantic benchmark matrix.

Images