Rehabilitation nursing device for patients with limb disorder based on myoelectric signal recognition

Abstract

This application discloses a rehabilitation and nursing device for patients with limb disorders based on electromyography (EMG) signal recognition. It includes an EMG acquisition module for collecting the patient's EMG signals, a main control module for signal processing and command generation, and a rehabilitation and nursing execution mechanism for outputting rehabilitation assistive movements. The signal output terminal of the EMG acquisition module is electrically connected to the signal input terminal of the main control module, and the control output terminal of the main control module is electrically connected to the controlled terminal of the rehabilitation and nursing execution mechanism. The EMG acquisition module is a single-channel surface EMG acquisition module. This application addresses the limitations of existing solutions for Brunnstrom stage 1-2 patients with muscle strength grade 1-2 who can only produce micro-muscle contractions. It overcomes the limitations of existing solutions that require multi-channel acquisition, pre-collected labeled samples, and forced resting calibration, achieving zero-sample, unconstrained adaptation under single-channel hardware, thus solving the problem of patients at this stage being unable to cooperate with calibration and sample collection.

Description

Technical Field

[0001] This application relates to the field of limb rehabilitation nursing technology, and in particular to a rehabilitation nursing device for patients with limb disorders based on electromyography signal recognition. Background Technology

[0002] Epidemiological data shows that over 70% of patients experience varying degrees of hemiplegic limb dysfunction after onset, severely impacting their quality of life. The early rehabilitation stage, characterized by Brunnstrom stages 1-2 and muscle strength grades 1-2, represents a golden window for neurofunctional remodeling. Effective intervention at this stage can significantly reduce long-term disability rates; however, patients require addressing rigid needs such as pressure ulcer prevention, postural adjustments, and assisted nursing care due to prolonged bed rest. Surface electromyography (sEMG), with its rapid response and ability to capture micro-muscle contractions, has become a core signal source for rehabilitation and nursing devices, reflecting movement intentions in advance and exhibiting greater sensitivity than force and angle signals.

[0003] Existing rehabilitation devices based on electromyography (EMG) signals have significant limitations. Their design is primarily geared towards patients with muscle strength ≥3 who can perform active joint movements. They rely on multi-channel acquisition units and supervised learning algorithms, requiring patients to cooperate in completing standard movements to collect training samples and undergo resting calibration. For Brunnstrom stage 1-2 patients, they can only produce barely perceptible micro-muscle contractions, unable to initiate active joint movements or complete the calibration process. This mismatch between technological requirements and physiological capabilities renders these devices ineffective during the crucial golden rehabilitation period, missing the optimal intervention opportunity.

[0004] Some solutions for weak electromyography (EMG) signal recognition use high-performance deep learning algorithms to optimize accuracy. While this improves performance, it significantly increases hardware costs and power consumption, making it difficult to meet the portability needs of home or grassroots scenarios. Some solutions only focus on front-end signal denoising, failing to solve the core problem of zero-sample adaptation. They also cannot distinguish between active microcontractions and artifacts such as pathological fasciculations and involuntary microspasms commonly seen in early hemiplegia. According to clinical statistics, the incidence of such artifacts exceeds 60% in early hemiplegic patients. These artifacts are easily misinterpreted as movement intentions, resulting in a high false trigger rate of the device. This not only weakens the effectiveness of intervention but may also cause patient anxiety or muscle fatigue.

[0005] In summary, existing technologies, due to their reliance on active cooperation and high-cost hardware, are limited in their applicability to various scenarios and struggle to reach the golden window of rehabilitation. Furthermore, at the signal recognition level, they are limited by artifact interference and the lack of zero-sample adaptation capability, resulting in insufficient reliability and a high false trigger rate in weak signal recognition.

[0006] Therefore, a rehabilitation and nursing device for patients with limb disorders based on electromyography signal recognition is proposed to achieve precise intervention and effective care for patients with early-stage hemiplegia. Summary of the Invention

[0007] This application aims to at least partially solve one of the technical problems in the aforementioned technologies.

[0008] To achieve the above objectives, the first aspect of this application proposes a rehabilitation and nursing device for patients with limb disorders based on electromyography (EMG) signal recognition, comprising an EMG acquisition module for acquiring the patient's EMG signals, a main control module for signal processing and instruction generation, and a rehabilitation and nursing execution mechanism for outputting rehabilitation assistive actions. The signal output terminal of the EMG acquisition module is electrically connected to the signal input terminal of the main control module, and the control output terminal of the main control module is electrically connected to the controlled terminal of the rehabilitation and nursing execution mechanism. The EMG acquisition module is a single-channel surface EMG acquisition module.

[0009] The main control module has a front-end preprocessing unit, a multi-domain feature extraction unit, an unconstrained baseline self-construction unit, an unsupervised integration discrimination unit, a pathological artifact temporal separation unit, and an identification and execution closed-loop control unit that are connected in sequence.

[0010] The front-end preprocessing unit is used to perform power frequency interference filtering, weak signal adaptive noise reduction, and sliding window processing on the raw electromyography signals acquired by the single-channel surface electromyography acquisition module, and output pure electromyography signals.

[0011] The multi-domain feature extraction unit is used to extract multi-dimensional electromyographic features from the pure electromyographic signal of each sliding window and generate corresponding feature vectors.

[0012] The unconstrained baseline self-construction unit is used to construct a resting-state baseline library from unconstrained electromyographic features acquired through an unsupervised clustering algorithm.

[0013] The unsupervised ensemble discrimination unit is used to perform multi-domain ensemble discrimination on the real-time generated feature vectors based on the resting-state baseline library and output the result of suspected active motion intent.

[0014] The pathological artifact temporal separation unit is used to perform temporal verification on electromyographic signals corresponding to suspected active movement intentions, filter out pathological electromyographic artifacts in patients, and output valid active movement intention results.

[0015] The identification and execution closed-loop control unit is used to generate drive control commands based on the results of effective active motion intentions and send them to the rehabilitation and nursing execution mechanism. At the same time, it feeds back the operating status of the rehabilitation and nursing execution mechanism to the front-end preprocessing unit and the unconstrained baseline self-construction unit.

[0016] In addition, the rehabilitation and nursing device for patients with limb disorders based on electromyography signal recognition proposed above in this application may also have the following additional technical features:

[0017] As a further description of the above technical solution:

[0018] The front-end preprocessing unit has a built-in second-order minimum mean square adaptive notch filter.

[0019] The front-end preprocessing unit performs an improved ensemble empirical mode decomposition on the notch-filtered electromyographic signal, decomposing the signal into 12 intrinsic mode function components. It removes the first two high-frequency noise components with the highest energy ratio and the last three baseline drift components with the lowest energy ratio. The remaining seven middle-order components are then locally denoised using a soft threshold function based on the standard deviation of each component, and a clean electromyographic signal is reconstructed.

[0020] As a further description of the above technical solution:

[0021] The multi-domain feature extraction unit extracts a total of 12 electromyographic features across four dimensions: higher-order statistical domain, frequency subband domain, nonlinear dynamics domain, and phase domain. Specifically, these features include skewness, kurtosis, and fourth-order cumulant in the higher-order statistical domain.

[0022] Among them, the frequency domain includes the energy proportion of the 125-250Hz frequency band, wavelet packet entropy, and spectral centroid frequency; the nonlinear dynamics domain includes the improved sample entropy, detrended fluctuation analysis scaling index, and Hurst exponent; and the phase domain includes the instantaneous phase variance, phase synchronization index, and joint higher-order statistics.

[0023] As a further description of the above technical solution:

[0024] The unconstrained baseline self-constructing unit adopts an improved density peak clustering algorithm, and its clustering distance metric is Mahalanobis distance;

[0025] The unconstrained baseline self-construction unit unconstrainedly acquires 60s of electromyography (EMG) signals, performs unsupervised clustering on all extracted feature vectors, and determines the cluster with the highest sample size and the smallest intra-cluster dispersion as the resting-state baseline cluster. Based on the resting-state baseline cluster, Gaussian distributed baselines are established for each of the 12 EMG features, and the mean of each feature is calculated. Standard deviation The resting state interval of each feature is defined as follows: Simultaneously, the maximum rate of change threshold of each feature in the adjacent window under the resting state is calculated, and a clustering calculation is performed every 10 seconds after startup to continuously optimize the parameters of the resting state baseline library.

[0026] As a further description of the above technical solution:

[0027] The unsupervised integrated discriminant unit is configured with an independent intra-domain discriminant for each of the four feature domains, and each intra-domain discriminant adopts dual-threshold discrimination logic.

[0028] The dual-threshold discrimination logic specifically states that the first threshold is the absolute value threshold of the features, and the threshold is the upper threshold of the corresponding resting state interval for two or more of the three features within a single domain. If so, it passes the first threshold verification;

[0029] The second threshold is the feature change rate threshold. If the change rate of two or more of the three features in a single domain exceeds the corresponding change rate threshold, then the second threshold is passed. The discriminator in the domain that passes both thresholds will output the initial judgment result of the active intent. Otherwise, the resting state result will be output.

[0030] The unsupervised integrated discrimination unit uses a fixed weighted majority voting rule to fuse the output results of the discriminators in each domain. The discriminator weights in the higher-order statistical domain and phase domain are 1.2, and the discriminator weights in the frequency domain subband domain and nonlinear dynamics domain are 1.0. After weighted voting, the total number of votes is ≥3, and the result of suspected active motion intention is output.

[0031] The unsupervised integrated discrimination unit is also equipped with an automatic anomaly domain shielding mechanism. If the dispersion of a feature in a single domain in the resting state exceeds a preset threshold, the discriminator in that domain is automatically shielded, and its weight is evenly distributed to the other normally functioning discriminators in the domain.

[0032] As a further description of the above technical solution:

[0033] The pathological artifact temporal separation unit adopts a three-level temporal verification logic, specifically the first-level temporal slope verification, which extracts the feature vectors of the current suspected active motion intention window and the previous three consecutive sliding windows, calculates the temporal change slope of each feature, and if more than 80% of the features show a continuous positive increasing trend, it passes the first-level verification; otherwise, it is judged as a pathological artifact and filtered out.

[0034] The secondary action potential firing synchronization rate verification involves extracting the electromyographic action potential waveform from the window signal that passes the primary verification and calculating the action potential firing synchronization rate. If the synchronization rate is ≥60%, the secondary verification is passed; otherwise, it is judged as a pathological artifact and filtered out.

[0035] A three-level artifact library matching and verification process is used to establish a dynamic pathological artifact feature library. For window feature vectors that pass the second-level verification, the Mahalanobis distance between them and features in the artifact library is calculated. If the Mahalanobis distance ≥ If the result is valid, the active motion intent will be verified and output. Otherwise, it will be judged as a pathological artifact and filtered. The pathological artifact temporal separation unit will automatically update the feature vector of the pathological artifact to the dynamic pathological artifact feature library.

[0036] As a further description of the above technical solution:

[0037] The recognition and execution closed-loop control unit has a built-in zero-sample electromyography and muscle strength linear mapping model. The zero-sample electromyography and muscle strength linear mapping model uses the feature mean value corresponding to the patient's first effective active movement intention as the benchmark muscle strength value, which corresponds to 10% of the basic assistance force of the rehabilitation nursing execution mechanism. It calculates in real time the ratio of the feature mean value of the current effective active movement intention window to the benchmark muscle strength value. The feature ratio has a negative linear relationship with the assistance force of the rehabilitation nursing execution mechanism.

[0038] The recognition and execution closed-loop control unit also has a built-in spasm prediction and safety inhibition mechanism and an electromyography-driven buttonless human-computer interaction module.

[0039] As a further description of the above technical solution:

[0040] The spasticity prediction and safety inhibition mechanism monitors the root mean square rate of change of the electromyography signal in the current window in real time. If the rate of change exceeds the preset safety threshold, regardless of the current judgment result, the auxiliary force of the rehabilitation nursing execution mechanism is immediately reduced linearly until the output is suspended.

[0041] The triggering logic of the electromyography-driven buttonless human-computer interaction module is that when the continuous and effective active movement intention exceeds 3 seconds, the working mode of the rehabilitation nursing execution mechanism is switched, without the need for the patient to operate physical buttons.

[0042] As a further description of the above technical solution:

[0043] The unconstrained baseline self-construction unit is also equipped with a sliding window-type baseline library incremental update subunit. It adopts a sliding baseline library with a capacity of 1000 sets of feature vectors. Every time a set of high-confidence resting-state feature vectors is collected, it is added to the sliding baseline library and the earliest set of data is removed. The Gaussian distribution parameters of the sliding baseline library are recalculated every 5 minutes.

[0044] Advantages of this invention:

[0045] According to the rehabilitation and nursing device for patients with limb disorders based on electromyography signal recognition in this application, for patients with Brunnstrom stage 1-2 and muscle strength grade 1-2 who can only produce micro-muscle contractions, it breaks through the limitations of existing solutions that require multi-channel acquisition, pre-collection and labeling samples, and forced resting calibration, and achieves zero-sample and unconstrained adaptation under single-channel hardware, solving the problem that patients at this stage cannot cooperate with calibration and sample collection.

[0046] The reliability and anti-interference ability of identification are significantly improved. Through the combined design of multi-domain feature fusion, unsupervised integrated discrimination and three-level pathological artifact separation, the common problems of insensitivity to weak signals at the microvolt level and inability to distinguish between active intent and pathological tremor artifacts are solved, and the false trigger rate is reduced to below 0.5%.

[0047] By deeply integrating intent recognition with the principles of proactive rehabilitation, and through adaptive assistance matching and spasticity prediction and inhibition, the problem of simply performing on / off recognition and being disconnected from clinical rehabilitation needs is avoided.

[0048] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0049] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0050] Figure 1 This is a schematic diagram of the logic flow of a rehabilitation and nursing device for patients with limb disabilities based on electromyography signal recognition according to an embodiment of this application; Figure 2 This is a schematic diagram of the principle of a rehabilitation and nursing device for patients with limb disabilities based on electromyography signal recognition according to an embodiment of this application. Detailed Implementation

[0051] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0052] The following description, in conjunction with the accompanying drawings, describes a rehabilitation and nursing device for patients with limb disabilities based on electromyography signal recognition, according to an embodiment of this application.

[0053] like Figure 1 As shown, the rehabilitation and nursing device for patients with limb disorders based on electromyography signal recognition in Embodiment 1 of this application may include an electromyography acquisition module for acquiring the patient's electromyography signals, a main control module for signal processing and instruction generation, and a rehabilitation and nursing execution mechanism for outputting rehabilitation assistive actions. The signal output terminal of the electromyography acquisition module is electrically connected to the signal input terminal of the main control module, and the control output terminal of the main control module is electrically connected to the controlled terminal of the rehabilitation and nursing execution mechanism. The electromyography acquisition module is a single-channel surface electromyography acquisition module.

[0054] The main control module has a front-end preprocessing unit, a multi-domain feature extraction unit, an unconstrained baseline self-construction unit, an unsupervised integration discrimination unit, a pathological artifact temporal separation unit, and an identification and execution closed-loop control unit connected in sequence.

[0055] The front-end preprocessing unit is used to perform power frequency interference filtering, weak signal adaptive noise reduction, and sliding window processing on the raw electromyography signals acquired by the single-channel surface electromyography acquisition module, and output pure electromyography signals.

[0056] The multi-domain feature extraction unit is used to extract multi-dimensional electromyographic features from the pure electromyographic signals of each sliding window and generate corresponding feature vectors.

[0057] Unconstrained baseline self-construction unit is used to construct a resting-state baseline library from unconstrained electromyographic features acquired through an unsupervised clustering algorithm;

[0058] The unsupervised ensemble discrimination unit is used to perform multi-domain ensemble discrimination on real-time generated feature vectors based on the resting-state baseline library and output results of suspected active motion intentions.

[0059] The pathological artifact temporal separation unit is used to perform temporal verification on electromyographic signals corresponding to suspected active movement intentions, filter out pathological electromyographic artifacts in patients, and output valid active movement intention results.

[0060] The closed-loop control unit is identified and executed to generate drive control commands based on the results of effective active motion intentions and send them to the rehabilitation and nursing execution mechanism. At the same time, the operating status of the rehabilitation and nursing execution mechanism is fed back to the front-end preprocessing unit and the unconstrained baseline self-construction unit.

[0061] The above-mentioned scheme, by limiting the acquisition of surface electromyography (EMG) to a single channel, abandons the multi-channel approach commonly used in existing technologies. It constructs a full-link serial processing architecture that includes preprocessing, feature extraction, baseline construction, integrated discrimination, artifact separation, and closed-loop control. This architecture fully covers the entire process from signal acquisition to action execution. It utilizes unconstrained baseline self-construction and zero-sample unsupervised integrated discrimination, eliminating the need for pre-acquiring labeled samples from patients and mandatory resting calibration. Through the temporal separation of pathological artifacts, it specifically filters out interference such as fasciculations and involuntary microspasms that are common in early-stage hemiplegic patients. By designing a reverse feedback closed loop for recognition and execution, it achieves adaptive updating of parameters throughout the entire process.

[0062] The overall principle and process are as follows: single-channel acquisition of patient electromyography signals → preprocessing and purification of microvolt-level weak signals → extraction of multi-domain sensitive features → automatic construction of resting-state baseline → multi-domain integration to determine motion intention → filtering of pathological artifacts → generation of adaptive drive commands to control the actuator, while simultaneously providing backfeed to optimize the baseline and preprocessing parameters.

[0063] like Figure 1 As shown:

[0064] The front-end preprocessing unit has a built-in second-order least mean square adaptive notch filter. The front-end preprocessing unit performs an improved ensemble empirical mode decomposition on the notch-filtered electromyographic signal, decomposing the signal into 12 intrinsic mode function components. It removes the first two high-frequency noise components with the highest energy ratio and the last three baseline drift components with the lowest energy ratio. The remaining seven middle components are then locally denoised using a soft threshold function based on the standard deviation of each component, and a clean electromyographic signal is reconstructed.

[0065] The above scheme uses a second-order least mean square adaptive notch filter with a fixed ±0.5Hz ultra-narrow band, filtering out only the 50Hz power frequency and its harmonics without damaging the weak electromyographic signals in adjacent frequency bands. It employs an improved ensemble empirical mode decomposition (EEMD) to split the signal into 12-order components, accurately removing high-frequency noise and low-frequency baseline drift. After reconstructing the signal by retaining the effective weak signal through a soft threshold function, it also fixes the 50ms window length and 10ms sliding step size to balance the real-time performance and computational load of the identification.

[0066] like Figure 1 As shown:

[0067] The multi-domain feature extraction unit extracts a total of 12 electromyographic features across four dimensions: higher-order statistical domain, frequency subband domain, nonlinear dynamics domain, and phase domain. Specifically, these features include skewness, kurtosis, and fourth-order cumulant in the higher-order statistical domain.

[0068] Among them, the energy proportion of the 125-250Hz frequency band in the frequency sub-band domain, wavelet packet entropy, and spectral centroid frequency; the improved sample entropy, detrended fluctuation analysis scaling index, and Hurst exponent in the nonlinear dynamics domain; and the instantaneous phase variance, phase synchronization index, and joint higher-order statistics in the phase domain.

[0069] The above scheme extracts 12 features across four dimensions: higher-order statistical domain, frequency subband domain, nonlinear dynamic domain, and phase domain. Each dimension corresponds to a different physical property of the electromyographic signal. While the signal amplitude is completely submerged by noise, the phase feature can still stably reflect the muscle contraction state. At the same time, the core parameters of the improved sample entropy are fixed to ensure sensitivity to weak signals.

[0070] like Figure 1 As shown:

[0071] The unconstrained baseline self-constructed unit adopts an improved density peak clustering algorithm, and its clustering distance metric is Mahalanobis distance;

[0072] Unconstrained baseline self-construction units were used to collect 60 seconds of electromyography (EMG) signals without constraints. Unsupervised clustering was performed on all extracted feature vectors. The cluster with the highest sample size and the smallest intra-cluster dispersion was determined as the resting-state baseline cluster. Gaussian distributed baselines were established for each of the 12 EMG features based on the resting-state baseline cluster, and the mean of each feature was calculated. Standard deviation The resting state interval of each feature is defined as follows: Simultaneously, the maximum rate of change threshold of each feature in the adjacent window under the resting state is calculated, and a clustering calculation is added every 10 seconds after startup to continuously optimize the parameters of the resting state baseline library;

[0073] The above scheme uses an improved density peak clustering algorithm with Mahalanobis distance as the clustering metric. After the device is powered on, it collects signals for 60 seconds without constraints and automatically identifies the cluster with the highest sample size and the smallest intra-class dispersion as the resting baseline cluster. This does not require the patient to relax. At the same time, Gaussian distribution baselines are established for each of the 12 features, and the discrimination threshold is determined. The baseline parameters are continuously optimized every 10 seconds after the device is powered on.

[0074] like Figure 1 As shown:

[0075] The unsupervised ensemble discriminant unit is configured with an independent intra-domain discriminant for each of the four feature domains, and each intra-domain discriminant uses dual-threshold discrimination logic.

[0076] The dual-threshold discrimination logic specifically involves the first threshold being the absolute value threshold of the features, and the second threshold being the upper threshold of the corresponding resting state interval for two or more of the three features within a single domain. If so, it passes the first threshold verification;

[0077] The second threshold is the feature change rate threshold. If the change rate of two or more of the three features in a single domain exceeds the corresponding change rate threshold, then the second threshold is passed. The discriminator in the domain that passes both thresholds will output the initial judgment result of the active intent. Otherwise, the resting state result will be output.

[0078] The unsupervised integrated discrimination unit uses a fixed weighted majority voting rule to fuse the output results of the discriminators in each domain. The discriminators in the higher-order statistical domain and phase domain have a weight of 1.2, while the discriminators in the frequency domain subband domain and nonlinear dynamics domain have a weight of 1.0. After weighted voting, if the total number of votes is ≥3, the result of suspected active motion intention is output.

[0079] The unsupervised ensemble discriminant unit is also equipped with an automatic anomaly domain shielding mechanism. If the dispersion of a feature in a single domain in the resting state exceeds a preset threshold, the discriminant in that domain is automatically shielded, and its weight is evenly distributed to the other normally functioning discriminants in the domain.

[0080] The above scheme adopts an integrated discrimination mechanism of multi-domain independent discriminators, dual threshold verification, weighted majority voting, and automatic masking of abnormal domains. Independent discriminators are configured for each of the four feature domains. Each discriminator is verified by both absolute value threshold and rate of change threshold to avoid misjudgment in one dimension. When higher-order statistical domains and phase domains that are more sensitive to weak signals are given higher weights, the final result is output through voting rules. When a single domain is disturbed, it is automatically masked and the weights are distributed to the remaining normal domains.

[0081] like Figure 1 As shown:

[0082] The pathological artifact temporal separation unit adopts a three-level temporal verification logic. Specifically, the first level is temporal slope verification, which extracts the feature vectors of the current suspected active motion intention window and the previous three consecutive sliding windows, calculates the temporal change slope of each feature, and passes the first level verification if more than 80% of the features show a continuous positive increasing trend; otherwise, it is judged as a pathological artifact and filtered out.

[0083] The secondary action potential firing synchronization rate verification involves extracting the electromyographic action potential waveform from the window signal that passes the primary verification and calculating the action potential firing synchronization rate. If the synchronization rate is ≥60%, the secondary verification is passed; otherwise, it is judged as a pathological artifact and filtered out.

[0084] A three-level artifact library matching and verification process is used to establish a dynamic pathological artifact feature library. For window feature vectors that pass the second-level verification, the Mahalanobis distance between them and features in the artifact library is calculated. If the Mahalanobis distance ≥ If the result is valid, the active motion intent will be verified and output. Otherwise, it will be judged as a pathological artifact and filtered. The pathological artifact temporal separation unit will automatically update the feature vector of the pathological artifact to the dynamic pathological artifact feature library.

[0085] The above scheme identifies the essential differences between active contraction and pathological artifacts. Active contraction is gradual, temporally continuous, and involves synchronous action potential delivery, while pathological artifacts are sudden, irregular, and involve asynchronous action potentials. Through three levels of filtering—temporal slope verification, action potential synchronization rate verification, and artifact library matching verification—the scheme accurately separates effective active intentions from pathological artifacts and automatically updates the artifact library.

[0086] like Figure 1 As shown:

[0087] The closed-loop control unit for identification and execution has a built-in zero-sample electromyography and muscle strength linear mapping model. The zero-sample electromyography and muscle strength linear mapping model uses the feature mean value corresponding to the patient's first effective active movement intention as the benchmark muscle strength value, which corresponds to 10% of the basic assist force of the rehabilitation nursing execution mechanism. It calculates in real time the ratio of the feature mean value of the current effective active movement intention window to the benchmark muscle strength value. The feature ratio has a negative linear relationship with the assist force of the rehabilitation nursing execution mechanism.

[0088] The recognition and execution closed-loop control unit also has a built-in spasm prediction and safety inhibition mechanism and an electromyography-driven buttonless human-machine interaction module;

[0089] The above scheme uses a zero-sample electromyography-muscle strength linear mapping model, taking the characteristic mean of the patient's first effective active contraction as the benchmark muscle strength value. The characteristic ratio is negatively correlated with the device's assistance force, that is, the stronger the patient's active contraction, the less assistance the device provides, strictly following the principle of active participation in neurorehabilitation.

[0090] like Figure 1 As shown:

[0091] The spasticity prediction and safety inhibition mechanism monitors the root mean square rate of change of the current window's electromyography signal in real time. If the rate of change exceeds the preset safety threshold, regardless of the current judgment result, it immediately and linearly reduces the assistance force of the rehabilitation nursing execution mechanism until the output is suspended.

[0092] The triggering logic of the electromyography-driven buttonless human-computer interaction module is that when the continuous and effective active movement intention exceeds 3 seconds, the working mode of the rehabilitation nursing execution mechanism is switched, without the need for the patient to operate physical buttons;

[0093] The above solution uses a spasm prediction and safety inhibition mechanism to monitor the root mean square rate of change of electromyographic signals in real time, reducing assistance in advance during the pre-existing stage of excessive muscle contraction to avoid inducing spasms. The electromyographic-driven buttonless human-computer interaction module triggers mode switching through effective active contraction for 3 seconds, eliminating the need for physical buttons.

[0094] like Figure 1 As shown:

[0095] The unconstrained baseline self-construction unit is also equipped with a sliding window-type baseline library incremental update sub-unit. It uses a sliding baseline library with a capacity of 1000 sets of feature vectors. Every time a set of high-confidence resting-state feature vectors is collected, it is added to the sliding baseline library and the earliest set of data is removed. The Gaussian distribution parameters of the sliding baseline library are recalculated every 5 minutes.

[0096] The above scheme adopts a sliding window-type baseline library incremental update sub-unit, with a fixed baseline library capacity of 1000 sets of feature vectors. Each time high-confidence resting-state data is collected, it is automatically added to the library. At the same time, the oldest historical data is removed, and the baseline parameters are recalculated every 5 minutes to adapt to the state changes of the electrode skin interface during long-term wear.

[0097] Example 2, further illustrated below with a specific case (e.g.) Figure 2 (as shown)

[0098] This embodiment targets upper limb hemiplegic patients with Brunnstrom stage 1-2 and muscle strength grade 1-2 after stroke. These patients can only produce micro-contractions of the biceps brachii muscle that are invisible to the naked eye, and have no active elbow flexion joint movement. They cannot cooperate with the sample collection and resting calibration of conventional rehabilitation devices, and there is a rigid demand for intervention during the golden period of rehabilitation, as well as for daily eating and postural adjustment nursing assistance.

[0099] The rehabilitation and nursing device for patients with limb disorders based on electromyography signal recognition in this embodiment includes the following hardware units:

[0100] The single-channel surface electromyography (EMG) acquisition module uses three disposable Ag / AgCl surface electrodes. One pair of differential acquisition electrodes is placed along the muscle fiber direction of the biceps brachii on the affected side of the patient, with an electrode spacing of 2 cm. One reference electrode is placed at the bony prominence of the elbow joint on the same side. The acquisition parameters are fixed at the sampling rate. 16-bit AD resolution, input range ±5mV, preamplifier gain 1000 times;

[0101] The main control module uses an STM32F030F4P6 8-bit MCU with a main frequency of 48MHz and built-in 32KB Flash and 4KB RAM, which fully meets the design requirements of low computing power and low cost. The single-window full-link processing time is ≤25ms.

[0102] Rehabilitation nursing execution mechanism: The upper limb lightweight exoskeleton driven by a small DC servo motor has a maximum assist torque of 5 N·m and an assist adjustment range of 0%-100%, which can realize rehabilitation nursing actions such as elbow flexion assistance, feeding support, and body position adjustment.

[0103] The signal output terminal of the single-channel surface electromyography acquisition module is electrically connected to the AD sampling input terminal of the main control module, and the PWM control output terminal of the main control module is electrically connected to the controlled terminal of the servo motor drive board of the rehabilitation nursing actuator.

[0104] The specific working principle of the above embodiments is as follows:

[0105] 1) The single-channel surface electromyography (EMG) acquisition module acquires the raw EMG signals of the patient's biceps brachii in real time and transmits them to the main control module;

[0106] 2) The front-end preprocessing unit sequentially performs power frequency notch filtering, adaptive noise reduction, and sliding window sizing on the raw signal to output a clean electromyographic signal;

[0107] 3) The multi-domain feature extraction unit extracts 12 fixed features in 4 dimensions from the clean signal of each sliding window to generate a feature vector;

[0108] 4) Unconstrained resting baseline self-construction unit automatically constructs a resting baseline library through unsupervised clustering, without requiring patients to undergo mandatory resting calibration throughout the process;

[0109] 5) The unsupervised ensemble discrimination unit performs multi-domain dual-threshold discrimination on real-time feature vectors based on the baseline library, and outputs the results of suspected active motion intentions through weighted voting. The entire process is zero-sample adaptation and does not require pre-collection of labeled samples.

[0110] 6) The pathological artifact timing separation unit performs three-level timing verification on suspected intention signals, filters out pathological artifacts such as fasciculations and microspasms, and outputs valid active motion intention results;

[0111] The identification-execution closed-loop control unit generates adaptive drive commands based on valid intentions, controls the actuator to output corresponding auxiliary force, and feeds back the operating status of the actuator to the front-end preprocessing unit and the baseline self-construction unit to achieve adaptive update of parameters across the entire link.

[0112] In this embodiment, all signal processing is performed based on a fixed sliding window, with a window length of [missing information]. The corresponding number of sampling points Sliding step size That is, the window data is updated every 10ms to complete one full-link processing. The discrete electromyographic signal sequence within a single window is uniformly denoted as... .

[0113] The principle of the front-end preprocessing unit is to completely preserve the effective electromyographic signal at the microvolt level while filtering out interference, which is implemented in three steps:

[0114] An ultra-narrowband second-order LMS adaptive power frequency notch filter is used to target 50Hz power frequency interference. It employs a second-order least mean square (LMS) adaptive notch filter, and its ultra-narrowband design avoids damaging weak electromyographic signals in adjacent frequency bands. Specific formulas and parameters are as follows:

[0115] Reference input sequence: ;

[0116] Notch filter output error (pure signal after removing power frequency): ,in These are the real-time weighting coefficients of the filter;

[0117] Weight Iterative Update Formula (Step Size Factor) =0.01, fixed bandwidth ±0.5Hz):

[0118] ;

[0119] The notch filter can track fluctuations of the power frequency within ±0.1Hz in real time, filtering out only the 50Hz power frequency and its harmonics, without affecting the effective electromyographic signal frequency band of 100-300Hz.

[0120] Among them, the improved EEMD adaptive noise reduction improves the signal after notch filtering. Performing improved ensemble empirical mode decomposition (EEMD) to address baseline drift and high-frequency noise issues, with specific implementation steps and formulas as follows:

[0121] right Gaussian white noise with an amplitude of 0.2 times the signal standard deviation is added, and EMD decomposition is performed. After repeating this process 100 times, the mean is taken to obtain the 12th-order intrinsic mode function (IMF) component, denoted as... ;

[0122] Calculate the normalized energy ratio for each IMF component: ;

[0123] Remove the top two high-frequency noise components with the highest energy ratio. 1. 2, and the last three baseline drift components with the lowest energy ratio. 10 , 11 , 12 Retain the middle 7th order components 3. 9;

[0124] For each retained IMF component, perform soft thresholding denoising. The soft thresholding function is:

[0125] Among them, threshold , For the first Standard deviation of each IMF component;

[0126] Reconstructing the thresholded IMF components yields a pure electromyographic signal. : ;

[0127] For the reconstructed pure signal The system is divided into windows with a fixed window length of 50ms and a sliding step of 10ms. Each window performs subsequent feature extraction and discrimination operations independently.

[0128] The multi-domain feature extraction unit extracts 12 features from the pure signal s(n) in each window, across four dimensions: higher-order statistical domain, frequency subband domain, nonlinear dynamic domain, and phase, generating a 12-dimensional standardized feature vector. All features are adapted to the computing power requirements of 8-bit MCUs. The calculation formulas and fixed parameters for each feature are as follows:

[0129] Based on the non-Gaussian enhancement of electromyographic signals during active contraction, the following was extracted:

[0130] Skewness This reflects the asymmetry in signal distribution:

[0131]

[0132] Kudo This reflects the sharpness of the signal distribution:

[0133]

[0134] Fourth-order cumulative amount Strengthening non-Gaussian characteristics:

[0135]

[0136] in, The mean of the window signal. The standard deviation of the window signal. Let be the mathematical expectation.

[0137] Frequency domain sub-band features are extracted based on the characteristic that electromyographic energy is concentrated in the 125-250Hz frequency band during active contraction:

[0138] Energy percentage in the 125-250Hz frequency band Using db4 wavelet basis pairs Perform 3-level wavelet packet decomposition, extract the 5th, 6th and 7th sub-bands of the 3rd level (corresponding to 125-250Hz), and calculate the proportion of the total energy of this frequency band to the total signal energy.

[0139] Wavelet packet entropy This reflects the complexity of signal energy distribution across different subbands.

[0140]

[0141] in, For the first The energy percentage of each sub-band;

[0142] Spectral centroid frequency Reflects the centroid shift of the signal spectrum

[0143]

[0144] in, The results of the fast Fourier transform, The frequency corresponding to each frequency point.

[0145] Nonlinear dynamics domain features are extracted to address the increased complexity of electromyographic signals during active contraction.

[0146] Improved sample entropy Fixed embedding dimension $$m=2$$, tolerance

[0147] The calculation formula is:

[0148]

[0149] in, dimensional vector matching number Number of dimensional vector matches;

[0150] Detrended volatility analysis (DFA) scaling index This reflects the long-range correlation of the signal;

[0151] Hurst Index This reflects the persistence of the signal's trend.

[0152] Phase domain features are used to extract the characteristic that phase features can still stably reflect muscle contraction even when the amplitude is submerged in noise.

[0153] Instantaneous phase variance ,right Perform Hilbert transform to obtain the analytic signal, and extract the instantaneous phase sequence. Calculate its variance;

[0154] Phase synchronization index This reflects the phase synchronization of electromyographic action potentials;

[0155] Combined higher-order statistics: the product of phase variance and fourth-order cumulant, enhances the characteristics of weak signals.

[0156] The principle of unconstrained resting-state baseline self-construction unit is to automatically identify resting-state signals and construct a standardized baseline library without requiring the patient to consciously relax. Specific implementation steps and formulas are as follows:

[0157] After the device is powered on, 60 seconds of unconstrained electromyography (EMG) signals are acquired, resulting in 5996 sets of 12-dimensional feature vectors with fixed window parameters. An improved density peak clustering (DPC) algorithm is used to perform unsupervised clustering on all feature vectors, with Mahalanobis distance as the clustering distance metric to avoid the influence of differences in feature dimensions. ;

[0158] in , There are two sets of feature vectors. Let be the covariance matrix of all eigenvectors;

[0159] After clustering, the cluster with the highest sample size and the smallest intra-cluster dispersion is identified as the resting-state baseline cluster (core logic: patients with grade 1-2 muscle strength always have the highest proportion of resting time and cannot generate sustained active contractions); based on the feature vector of the baseline cluster, Gaussian distribution baselines are established for each of the 12 features, and the resting-state baseline is calculated. Mean of the feature Standard deviation Its resting state interval is defined as Simultaneously, the threshold of the maximum rate of change of this feature in adjacent windows under resting state is calculated. ;

[0160] After powering on, clustering calculations are performed every 10 seconds to continuously optimize baseline library parameters, reaching a stable state after 60 seconds.

[0161] The sliding window incremental update uses a sliding baseline library with a capacity of 1000 feature vectors. Each time a high-confidence resting-state feature vector (Mahaviron distance) is collected... This involves adding data to the sliding baseline database while removing the oldest historical data set, and recalculating the baseline database every 5 minutes. It adapts to the changes in the state of the electrode skin interface during long-term wear.

[0162] The unsupervised ensemble discriminant unit (ALU) operates on the principle of multi-domain cross-validation, avoiding misjudgments caused by the failure of single-dimensional features. It employs zero-sample adaptation throughout the process. Specific implementation details are as follows:

[0163] Each of the four feature domains is configured with an independent intra-domain discriminator, with each discriminator corresponding to three features, and a fixed double threshold discrimination logic is used:

[0164] First threshold (absolute value threshold): Among the three features within a single domain, two or more exceed the upper threshold of the corresponding resting state interval. If so, it passes the first threshold verification;

[0165] Second threshold (rate of change threshold): Within a single domain, the rate of change of two or more of the three features in adjacent windows exceeds the corresponding threshold. If so, then it passes the second threshold verification;

[0166] Simultaneously, the intra-domain discriminator, which uses two threshold checks, outputs the initial result of the active intent judgment; otherwise, it outputs the resting state result.

[0167] The weighted majority voting fusion adopts a fixed-weight voting rule: the weight of the discriminator in the higher-order statistical domain and phase domain is 1.2, the weight of the discriminator in the frequency domain sub-band domain and nonlinear dynamic domain is 1.0, and the total weight is 4.4. After weighted voting, if the total number of votes is ≥3.0, the result of suspected active motion intention is output; otherwise, the result of resting state is output.

[0168] In the automatic anomaly domain masking process, if the dispersion of features within a single domain in the resting state exceeds a preset threshold ( The discriminator in that domain is automatically masked (by multiples of the initial standard deviation), and its weight is evenly distributed to the remaining normally functioning discriminators to prevent interference with a single domain from causing overall failure.

[0169] The core principle of the pathological artifact temporal separation unit is to grasp the essential difference between active contraction and pathological artifacts: active contraction is gradual, temporally continuous, and with synchronized action potentials; pathological artifacts are sudden, irregular, and with asynchronous action potentials. Accurate separation is achieved through three levels of verification.

[0170] Extract the feature vectors of the current suspected intent window and the previous three consecutive sliding windows, and calculate the temporal change slope of 12 features; if more than 80% of the features show a continuous positive increasing trend, it passes the first-level verification; otherwise, it is judged as a pathological artifact and filtered out.

[0171] For the window signal that passes the first-level verification, extract the electromyographic action potential (MUAP) waveform and calculate the MUAP firing synchronization rate. : ;

[0172] like If it passes the second-level verification, it is judged as a pathological artifact and filtered out.

[0173] Establish a dynamic pathological artifact feature library, and analyze the feature vectors that pass the second-level verification. Calculate the minimum Mahalanobis distance between it and all feature vectors in the artifact library. ;like If the result passes the final verification, a valid active motion intent result will be output; otherwise, it will be judged as an artifact and filtered out.

[0174] The principle of the closed-loop control unit for identification and execution is to deeply bind intention recognition with the principle of active rehabilitation to achieve adaptive assistive control and safety protection. Specific implementation method:

[0175] The baseline muscle strength value was the average of 12 characteristics corresponding to the patient's first effective active intention. This corresponds to a basic assistance level of 10% for rehabilitation nursing institutions; it also calculates the average feature value of the current effective intent window in real time. Adjust the auxiliary force linearly according to the following formula. : ;

[0176] in, Basic support level, This is the proportionality coefficient. The adjustment range is 0%-100%;

[0177] Core logic: Feature mean The larger the size, the stronger the patient's active contraction and the greater the assist force. The smaller the size, the more proactive the patient should be in neurorehabilitation, and the less assistance the device should provide, in order to avoid disuse atrophy of muscles.

[0178] During device operation, feature vectors identified as pathological artifacts are automatically updated to the artifact library, which has a maximum capacity of 500 sets and is updated using a first-in-first-out rule.

[0179] Real-time monitoring of the root mean square (RMS) rate of change of electromyography signals in the current window :

[0180] like If the preset safety threshold (200% / 100ms) is exceeded, regardless of the current judgment result, the auxiliary force of the actuator will be reduced linearly immediately until the output is stopped to avoid inducing muscle spasms in the patient.

[0181] The fixed trigger logic of the electromyography-driven buttonless human-computer interaction is as follows: when the continuous effective active movement intention exceeds 3 seconds, the working mode of the rehabilitation nursing execution mechanism is automatically switched (such as elbow flexion assistance → arm raising assistance → body position adjustment), without the need for the patient to operate physical buttons, which is fully adapted to the operation ability of patients with limb disabilities.

[0182] During closed-loop feedback updates, the operating status of the actuator and the changing trend of the patient's electromyography signal are fed back to the front-end preprocessing unit and the baseline self-construction unit, and the denoising parameters and baseline library are adjusted in real time to achieve end-to-end adaptive optimization.

[0183] In summary, the rehabilitation and nursing device for patients with limb disorders based on electromyography signal recognition according to the embodiments of this application has an intent recognition response delay of ≤60ms, meeting the needs of real-time assistance; an active micro-contraction recognition accuracy of ≥92% for patients with grade 1-2 muscle strength; a pathological artifact filtering rate of ≥98% and a steady-state false trigger rate of ≤0.5%; a power-on readiness time of ≤60s, requiring no mandatory resting calibration and having zero cooperation threshold; continuous and stable wear for ≥8 hours without manual recalibration; and a total hardware cost of ≤20 yuan, making it suitable for large-scale promotion in home and primary care settings.

[0184] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0185] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0186] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A rehabilitation nursing device for patients with limb disorders based on myoelectric signal recognition, comprising a myoelectric signal acquisition module for acquiring myoelectric signals of a patient, a main control module for signal processing and instruction generation, and a rehabilitation nursing execution mechanism for outputting rehabilitation auxiliary actions, wherein the signal output end of the myoelectric signal acquisition module is electrically connected with the signal input end of the main control module, and the control output end of the main control module is electrically connected with the controlled end of the rehabilitation nursing execution mechanism. The electromyography (EMG) acquisition module is a single-channel surface EMG acquisition module. The main control module has a front-end preprocessing unit, a multi-domain feature extraction unit, an unconstrained baseline self-construction unit, an unsupervised integration discrimination unit, a pathological artifact temporal separation unit, and an identification and execution closed-loop control unit that are connected in sequence. The front-end preprocessing unit is used to perform power frequency interference filtering, weak signal adaptive noise reduction, and sliding window processing on the raw electromyography signals acquired by the single-channel surface electromyography acquisition module, and output pure electromyography signals. The multi-domain feature extraction unit is used to extract multi-dimensional electromyographic features from the pure electromyographic signal of each sliding window and generate corresponding feature vectors. The unconstrained baseline self-construction unit is used to construct a resting-state baseline library from unconstrained electromyographic features acquired through an unsupervised clustering algorithm. The unsupervised ensemble discrimination unit is used to perform multi-domain ensemble discrimination on the real-time generated feature vectors based on the resting-state baseline library and output the result of suspected active motion intent. The pathological artifact temporal separation unit is used to perform temporal verification on electromyographic signals corresponding to suspected active movement intentions, filter out pathological electromyographic artifacts in patients, and output valid active movement intention results. The identification and execution closed-loop control unit is used to generate drive control commands based on the results of effective active motion intentions and send them to the rehabilitation and nursing execution mechanism. At the same time, it feeds back the operating status of the rehabilitation and nursing execution mechanism to the front-end preprocessing unit and the unconstrained baseline self-construction unit.

2. The rehabilitation nursing device for a patient with a limb disorder based on myoelectric signal recognition according to claim 1, characterized in that, The front-end preprocessing unit has a built-in second-order minimum mean square adaptive notch filter. The front-end preprocessing unit performs an improved ensemble empirical mode decomposition on the notch-filtered electromyographic signal, decomposing the signal into 12 intrinsic mode function components. It removes the first two high-frequency noise components with the highest energy ratio and the last three baseline drift components with the lowest energy ratio. The remaining seven middle-order components are then locally denoised using a soft threshold function based on the standard deviation of each component, and a clean electromyographic signal is reconstructed.

3. The rehabilitation nursing device for the patient with limb disorder based on the recognition of myoelectric signal according to claim 1, characterized in that, The multi-domain feature extraction unit extracts a total of 12 electromyographic features across four dimensions: higher-order statistical domain, frequency subband domain, nonlinear dynamics domain, and phase domain. Specifically, these features include skewness, kurtosis, and fourth-order cumulant in the higher-order statistical domain. Among them, the frequency domain includes the energy proportion of the 125-250Hz frequency band, wavelet packet entropy, and spectral centroid frequency; the nonlinear dynamics domain includes the improved sample entropy, detrended fluctuation analysis scaling index, and Hurst exponent; and the phase domain includes the instantaneous phase variance, phase synchronization index, and joint higher-order statistics.

4. The rehabilitation nursing device for a patient with a limb disorder based on myoelectric signal recognition according to claim 1, characterized in that, The unconstrained baseline self-constructing unit adopts an improved density peak clustering algorithm, and its clustering distance metric is Mahalanobis distance; The unconstrained baseline self-construction unit collects 60s electromyographic signals without constraint, performs unsupervised clustering on all extracted feature vectors, determines the cluster with the highest sample size ratio and the smallest intra-class dispersion as the resting state baseline cluster, establishes a Gaussian distribution baseline for 12 electromyographic features based on the resting state baseline cluster, calculates the mean , standard deviation of each feature , and defines the resting state interval of each feature Meanwhile, the maximum change rate threshold of each feature in the resting state is calculated, and the clustering calculation is supplemented every 10s after starting, continuously optimizing the parameters of the resting state baseline library.

5. The rehabilitation nursing device for the patient with limb disorder based on the recognition of myoelectric signal according to claim 4, characterized in that, The unsupervised integrated discriminant unit is configured with an independent intra-domain discriminant for each of the four feature domains, and each intra-domain discriminant adopts dual-threshold discrimination logic. The double-threshold discrimination logic is specifically that the first threshold is an absolute value threshold of the feature, and two or more of the three features in the single domain exceed the upper threshold of the corresponding resting state interval Then the first threshold is checked. The second threshold is the feature change rate threshold. If the change rate of two or more of the three features in a single domain exceeds the corresponding change rate threshold, then the second threshold is passed. The discriminator in the domain that passes both thresholds will output the initial judgment result of the active intent. Otherwise, the resting state result will be output. The unsupervised integrated discrimination unit uses a fixed weighted majority voting rule to fuse the output results of the discriminators in each domain. The discriminator weights in the higher-order statistical domain and phase domain are 1.2, and the discriminator weights in the frequency domain subband domain and nonlinear dynamics domain are 1.

0. After weighted voting, the total number of votes is ≥3, and the result of suspected active motion intention is output. The unsupervised integrated discrimination unit is also equipped with an automatic anomaly domain shielding mechanism. If the dispersion of a feature in a single domain in the resting state exceeds a preset threshold, the discriminator in that domain is automatically shielded, and its weight is evenly distributed to the other normally functioning discriminators in the domain.

6. The rehabilitation nursing device for a patient with a limb disorder based on myoelectric signal recognition according to claim 1, characterized in that, The pathological artifact temporal separation unit adopts a three-level temporal verification logic, specifically the first-level temporal slope verification, which extracts the feature vectors of the current suspected active motion intention window and the previous three consecutive sliding windows, calculates the temporal change slope of each feature, and if more than 80% of the features show a continuous positive increasing trend, it passes the first-level verification; otherwise, it is judged as a pathological artifact and filtered out. The secondary action potential firing synchronization rate verification involves extracting the electromyographic action potential waveform from the window signal that passes the primary verification and calculating the action potential firing synchronization rate. If the synchronization rate is ≥60%, the secondary verification is passed; otherwise, it is judged as a pathological artifact and filtered out. Third level artifact library matching verification, a dynamic pathological artifact feature library is established, and the Mahalanobis distance between the window feature vector passing the second level verification and the features in the artifact library is calculated. If the Mahalanobis distance is greater than or equal to then the final verification is passed and the effective voluntary motion intention result is output, otherwise it is determined as a pathological artifact and filtered, and the pathological artifact timing separation unit automatically updates the feature vector determined as a pathological artifact to the dynamic pathological artifact feature library.

7. The rehabilitation nursing device for a patient with a limb disorder based on myoelectric signal recognition according to claim 1, characterized in that, The recognition and execution closed-loop control unit has a built-in zero-sample electromyography and muscle strength linear mapping model. The zero-sample electromyography and muscle strength linear mapping model uses the feature mean value corresponding to the patient's first effective active movement intention as the benchmark muscle strength value, which corresponds to 10% of the basic assistance force of the rehabilitation nursing execution mechanism. It calculates in real time the ratio of the feature mean value of the current effective active movement intention window to the benchmark muscle strength value. The feature ratio has a negative linear relationship with the assistance force of the rehabilitation nursing execution mechanism. The recognition and execution closed-loop control unit also has a built-in spasm prediction and safety inhibition mechanism and an electromyography-driven buttonless human-computer interaction module.

8. The rehabilitation nursing device for the patient with limb disorder based on the recognition of myoelectric signal according to claim 7, characterized in that, The spasticity prediction and safety inhibition mechanism monitors the root mean square rate of change of the electromyography signal in the current window in real time. If the rate of change exceeds the preset safety threshold, regardless of the current judgment result, the auxiliary force of the rehabilitation nursing execution mechanism is immediately reduced linearly until the output is suspended. The triggering logic of the electromyography-driven buttonless human-computer interaction module is that when the continuous and effective active movement intention exceeds 3 seconds, the working mode of the rehabilitation nursing execution mechanism is switched, without the need for the patient to operate physical buttons.

9. The rehabilitation nursing device for the patient with limb disorder based on the recognition of myoelectric signal according to claim 4, characterized in that, The unconstrained baseline self-construction unit is also equipped with a sliding window-type baseline library incremental update subunit. It adopts a sliding baseline library with a capacity of 1000 sets of feature vectors. Every time a set of high-confidence resting-state feature vectors is collected, it is added to the sliding baseline library and the earliest set of data is removed. The Gaussian distribution parameters of the sliding baseline library are recalculated every 5 minutes.

Images