Rehabilitation robot force perception network method and system based on dynamic weight fusion

Abstract

The application belongs to the technical field of robots, and discloses a rehabilitation robot force perception network method and system based on dynamic weight fusion, which acquires original electromyography signals and original force signals, and respectively performs feature enhancement on the original electromyography signals and the original force signals by using a generative adversarial network; based on a muscle force grading adaptive mechanism, the original electromyography signals and the original force signals after feature enhancement are respectively fused to obtain a fused force perception value; by using an adaptive calibration mechanism and an adaptive compensation mechanism, the acquired temperature and humidity data and the calibration signals are processed to extract a baseline drift correction value and a temperature compensation coefficient, and the baseline drift correction value is input into the muscle force grading adaptive mechanism for optimization; by using a multi-scale feature fusion mechanism, the extracted fused force perception value and the temperature compensation coefficient are fused and processed to obtain fused multi-scale features. The application significantly improves the perception effect of patients with extremely weak muscle strength and spasm.

Description

Technical Field

[0001] This invention relates to the field of robotics, and in particular to a method and system for force sensing networks in rehabilitation robots based on dynamic weight fusion. Background Technology

[0002] The force sensing performance of rehabilitation robots is crucial in determining the accuracy and safety of force control, but existing technologies have significant limitations: traditional force sensing uses fixed-weight single-modality or simple multi-modal fusion, which cannot dynamically adapt to fluctuations in patient muscle strength, such as changes in muscle tension and unstable force exertion; the perception and decision-making layers are disconnected, lacking a collaborative mechanism, resulting in a force sensing lag of over 150ms; it is susceptible to drift during long-term operation, leading to rapid decay in perception accuracy; and it cannot be specifically adjusted according to different muscle strength levels, resulting in poor perception for patients with extremely weak muscles or spasticity. These problems make force sensing a key bottleneck restricting the force control performance of rehabilitation robots, urgently requiring innovative breakthroughs in network architecture and fusion mechanisms.

[0003] Therefore, how to provide a method and system for force sensing networks for rehabilitation robots based on dynamic weight fusion is an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method and system for a force sensing network for rehabilitation robots based on dynamic weight fusion, in order to solve the problems mentioned above in the prior art.

[0005] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or to describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0006] According to a first aspect of the present invention, a method for a force sensing network for a rehabilitation robot based on dynamic weight fusion is provided.

[0007] In one embodiment, a force sensing network method for rehabilitation robots based on dynamic weight fusion includes:

[0008] The original electromyography (EMG) signal and the original force signal were acquired, and the original EMG signal and the original force signal were enhanced by generative adversarial network. Based on the muscle force grading adaptive mechanism, the original EMG signal and the original force signal after feature enhancement were fused to obtain the fused force perception value.

[0009] The acquired temperature and humidity data and calibration signals are processed using adaptive calibration and adaptive compensation mechanisms to extract baseline drift correction values ​​and temperature compensation coefficients. The baseline drift correction values ​​are then input into the muscle strength grading adaptive mechanism for optimization.

[0010] The extracted fusion force perception value and temperature compensation coefficient are fused using a multi-scale feature fusion mechanism to obtain the fused multi-scale features.

[0011] A cross-layer feedback adaptation mechanism is used to predict and correct multi-scale features, obtain perceptual reliability scores and weight adjustment parameters, and input the weight adjustment parameters into the muscle strength grading adaptive mechanism for secondary optimization.

[0012] In one embodiment, the process of acquiring the original electromyography (EMG) signal and the original force signal, performing feature enhancement on the original EMG signal and the original force signal using a generative adversarial network (GAN), and fusing the feature-enhanced original EMG signal and the original force signal based on a muscle strength grading adaptive mechanism to obtain a fused force perception value includes:

[0013] An eight-channel surface electromyography (EMG) sensor and an S-shaped force sensor were used to acquire raw EMG signals and raw force signals, respectively.

[0014] Generative adversarial networks were used to enhance the features of the original electromyography and force signals, respectively.

[0015] The original electromyography signal and the original force signal after feature enhancement were processed by a second-order Butterworth filter and a fourth-order Bessel filter, respectively, to construct a first-level feature vector and a second-level feature vector.

[0016] Convolutional neural networks and long short-term memory networks are used to extract features from the first-level feature vector and the second-level feature vector, respectively, to obtain the muscle force feature vector and the contact force feature vector.

[0017] Based on the attention fusion mechanism, the muscle force feature vector and contact force feature vector are fused with the pre-configured muscle force level to obtain the fused force perception value.

[0018] In one embodiment, the feature enhancement of the original electromyography signal and the original force signal using a generative adversarial network includes:

[0019] The original electromyography signals and original force signals were constructed and compared with pre-configured clinical assessment results to obtain the training dataset;

[0020] The generator network and the judge network are trained on the training dataset simultaneously. The key features of the generator network and the judge network are extracted by knowledge distillation mechanism, and feature enhancement is performed on the input raw electromyography signal and raw force signal.

[0021] In one embodiment, the step of simultaneously training the training dataset using a generator network and a judge network, extracting key features from the generator network and the judge network using a knowledge distillation mechanism, and performing feature enhancement on the input raw electromyography signal and raw force signal includes:

[0022] Random noise is input into the generator network to generate simulated feature vectors, and a discriminator network is used to distinguish between real and generated first-order simulated features.

[0023] The loss function of the first-order simulated features is optimized by using local feature matching terms, and the first-order simulated features are compressed into second-order key features by using a knowledge distillation mechanism.

[0024] The second-order key features were input into the original electromyography signal and the original force signal using an adaptive moment estimation optimizer to enhance the features.

[0025] In one embodiment, the fusion of muscle force feature vector and contact force feature vector with pre-configured muscle force levels based on the attention fusion mechanism to obtain a fused force perception value includes:

[0026] The dynamic weight calculation function is used to calculate the weights of the muscle force feature vector and the contact force feature vector with the pre-configured muscle force level to obtain the first-level weight and the second-level weight.

[0027] The weights of the first-level and second-level weights are adjusted using a weight smoothing transition mechanism, and the fusion force perception value is obtained through the fusion force perception value calculation formula.

[0028] In one embodiment, the step of using a dynamic weight calculation function to calculate the weights of the muscle strength feature vector and the contact force feature vector with a pre-configured muscle strength level to obtain primary weights and secondary weights includes:

[0029] The dynamic weight calculation function is used to monitor muscle tension and force fluctuation coefficient in real time. Different target weight calculation paths are triggered according to whether the preset threshold is exceeded, and an initial target weight value is generated.

[0030] The initial target weight value is constrained by limiting conditions, and the initial target weight value that exceeds the limit is truncated to obtain the first-level weight and the second-level weight.

[0031] In one embodiment, the process of using an adaptive calibration mechanism and an adaptive compensation mechanism to process the acquired temperature and humidity data and calibration signal, extracting baseline drift correction values ​​and temperature compensation coefficients, and inputting the baseline drift correction values ​​into the muscle strength grading adaptive mechanism for optimization includes:

[0032] By triggering a periodic calibration process, the robot is brought into contact with a standard load and sensor readings are collected. The raw data is then input into a linear correction formula to calculate and generate proportional coefficients and offset parameters, which are then stored in a calibration parameter table, and the calibration time is recorded.

[0033] The system continuously collects environmental data through temperature and humidity sensors. When environmental changes exceed a preset threshold, a compensation mechanism is triggered. Real-time data is input into the temperature compensation formula for correction, and a temperature compensation coefficient is calculated and stored in the environmental compensation parameter table.

[0034] Combining calibration time and pre-configured muscle strength level data, the original force signal is compensated by inputting into the time-varying drift compensation formula, a baseline drift correction value is calculated, and the baseline drift correction value is input into the muscle strength grading adaptive mechanism for optimization.

[0035] A three-dimensional coupling matrix is ​​constructed, and an adaptive compensation triggering mechanism is used to dynamically correct the temperature compensation coefficient and drift correction value in tandem.

[0036] In one embodiment, the process of fusing the extracted fusion force sensing value and temperature compensation coefficient through a multi-scale feature fusion mechanism to obtain the fused multi-scale features includes:

[0037] The force sensing value and temperature compensation coefficient are fused and input into the adaptive Kalman filter and the multi-scale fusion module for processing, respectively, to generate the preliminary filtered force signal estimate and the multi-scale fused force signal.

[0038] By calculating the muscle activation index of electromyography and performing correlation analysis with multi-scale force signals, a correction mechanism is activated to generate a correction force value when abnormalities are detected. At the same time, physiological delay is calculated through dynamic time warping and the signal is corrected for time offset.

[0039] By using a cross-modal spatiotemporal weighted fusion formula, time-corrected multimodal signals are weighted and fused to finally output time-aligned multi-scale features.

[0040] In one embodiment, the step of using a cross-layer feedback adaptation mechanism to predict and correct multi-scale features, obtaining a perceptual reliability score and weight adjustment parameters, and then inputting the weight adjustment parameters into a muscle strength grading adaptive mechanism for secondary optimization includes:

[0041] Based on the pre-configured decision layer, a two-way interaction mechanism is built through the initialization of the communication interface to receive the adjustment parameters and training modes issued by the decision layer, and to establish a parameter mapping table and threshold adjustment strategy.

[0042] By reading historical fusion force data and inputting it into a long short-term memory prediction network, a future force change trend curve is generated. By comparing the deviation between the actual measured value and the predicted value, a dynamic correction coefficient is generated to calibrate the prediction trajectory.

[0043] By calculating the perceived reliability score, the score result is obtained. The real-time force signal, the calibrated predicted trajectory and the reliability score result are generated into a data frame and sent back to the decision-making level for optimization.

[0044] According to a second aspect of the present invention, a force sensing network system for rehabilitation robots based on dynamic weight fusion is provided.

[0045] In one embodiment, the force sensing network system for rehabilitation robots based on dynamic weight fusion includes:

[0046] The perception fusion module is used to acquire the original electromyography (EMG) signal and the original force signal. It uses a generative adversarial network to enhance the features of the original EMG signal and the original force signal respectively. Based on the muscle force grading adaptive mechanism, the feature-enhanced original EMG signal and the original force signal are fused to obtain the fused force perception value.

[0047] The correction and compensation module is used to process the acquired temperature and humidity data and calibration signals using the adaptive calibration mechanism and the adaptive compensation mechanism, extract the baseline drift correction value and temperature compensation coefficient, and input the baseline drift correction value into the muscle strength grading adaptive mechanism for optimization.

[0048] The feature fusion module is used to fuse the extracted fusion force perception value and temperature compensation coefficient through a multi-scale feature fusion mechanism to obtain the fused multi-scale features.

[0049] The prediction and correction module is used to predict and correct multi-scale features using a cross-layer feedback adaptation mechanism, obtain the perceptual reliability score and weight adjustment parameters, and input the weight adjustment parameters into the muscle strength grading adaptive mechanism for secondary optimization.

[0050] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0051] 1. This invention employs a dual-channel and dynamic weighting and muscle strength grading architecture, breaking through the traditional fixed fusion mode to achieve active adaptation of force sensing, maintaining high accuracy even when the patient's muscle strength fluctuates by 30%. Traditional single-modal sensing schemes exhibit significant sensing deviations (>0.5N) when muscle strength fluctuations exceed 15%, while this invention, through dynamic weighting adjustment, can control the deviation within 0.1N, without requiring manual intervention for recalibration.

[0052] 2. This invention reduces force perception error to ±0.08N through dynamic calibration, multi-scale fusion, and adversarial training, representing a 75% improvement over traditional methods. In a 5N standard load test, the average error of the traditional algorithm is 0.32N, while that of this invention is only 0.08N. In dynamic force tracking experiments simulating active force exertion by the patient, the tracking error is reduced from the traditional 0.5N to 0.09N, meeting the needs of fine rehabilitation training.

[0053] 3. This invention employs a cross-layer feedback mechanism and a predictive correction structure to shorten the perception-decision delay to 40ms and enable anomaly recognition 60ms earlier. The perception-decision closed-loop delay of traditional force control systems for rehabilitation robots is typically 150-200ms, making it difficult to cope with sudden spasms (attack time approximately 100ms). This invention can issue an early warning 60ms before the spasm occurs, providing sufficient response time for safety protection.

[0054] 4. This invention utilizes a muscle strength grading adaptive mechanism to enable the algorithm to adapt to patients with different muscle strength levels (0-5), significantly improving the perception effect, especially for patients with extremely weak muscle strength (0-1) and spasticity. For patients with muscle strength level 0 (complete paralysis), through... EMG By combining weak signal feature amplification with high-weight fusion of force sensor, the sensing sensitivity is improved by 3 times; for patients with spasticity, the abnormality recognition accuracy is increased from the traditional 78% to 97.5%, effectively avoiding secondary injury.

[0055] 5. This invention adopts a modular algorithm design to support rapid deployment. Tests on three different models of rehabilitation robots show that the integration time is reduced from 72 hours in the traditional solution to 12 hours. The patient adaptation process is simplified, and the parameter debugging time for new patients using the device for the first time is reduced from 30 minutes to 5 minutes, which greatly lowers the threshold for clinical operation.

[0056] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0057] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0058] Figure 1 This is a flowchart illustrating a force sensing network method for rehabilitation robots based on dynamic weight fusion, according to an exemplary embodiment.

[0059] Figure 2 This is a structural block diagram of a force sensing network system for a rehabilitation robot based on dynamic weight fusion, according to an exemplary embodiment.

[0060] Figure 3This is a diagram illustrating the overall architecture of a force sensing network for a rehabilitation robot based on dynamic weight fusion, according to an exemplary embodiment.

[0061] Figure 4 This is a diagram of the dynamic weighted force sensing network structure of a force sensing network method for rehabilitation robots based on dynamic weight fusion, according to an exemplary embodiment.

[0062] Figure 5 This is a flowchart illustrating the dynamic adjustment of attention weights in a force sensing network method for rehabilitation robots based on dynamic weight fusion, according to an exemplary embodiment.

[0063] Figure 6 This is a cross-layer feedback interaction timing diagram of a force sensing network method for rehabilitation robots based on dynamic weight fusion, illustrated according to an exemplary embodiment.

[0064] Figure 7 This is a flowchart illustrating the dynamic calibration and drift compensation process of a force sensing network method for rehabilitation robots based on dynamic weight fusion, according to an exemplary embodiment.

[0065] Figure 8 This is a structural diagram of an adversarial training optimization module for a force-sensing network method for rehabilitation robots based on dynamic weight fusion, according to an exemplary embodiment. Detailed Implementation

[0066] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some embodiments may include or substitute parts and features of other embodiments. The scope of the embodiments herein encompasses the entire scope of the claims and all available equivalents thereof. Throughout this document, the terms “first,” “second,” etc., are used only to distinguish one element from another without requiring or implying any actual relationship or order between the elements. Indeed, a first element can also be referred to as a second element, and vice versa. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a structure, apparatus, or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a structure, apparatus, or device. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the structure, apparatus, or device that includes said element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0067] The terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer" used in this document to indicate orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings. They are used solely for the convenience of describing the document and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description herein, unless otherwise specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two elements; they can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0068] In this document, unless otherwise stated, the term "multiple" means two or more.

[0069] In this article, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0070] In this article, the term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0071] It should be understood that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the diagram may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0072] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0073] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0074] Figure 1An embodiment of the force sensing network method for rehabilitation robots based on dynamic weight fusion of the present invention is shown.

[0075] In this optional embodiment, the force sensing network method for rehabilitation robots based on dynamic weight fusion includes:

[0076] Step S101: Obtain the raw electromyography (i.e., EMG ) signals and original force signals, using generative adversarial networks (i.e. GAN The original electromyography (EMG) signal and the original force signal were respectively enhanced in terms of features; based on the muscle force grading adaptive mechanism, the enhanced original EMG signal and the original force signal were fused to obtain the fused force sensing value.

[0077] Step S102: Process the acquired temperature and humidity data and calibration signal using the adaptive calibration mechanism and adaptive compensation mechanism, extract the baseline drift correction value and temperature compensation coefficient, and input the baseline drift correction value into the muscle strength grading adaptive mechanism for optimization.

[0078] Step S103: The extracted fusion force perception value and temperature compensation coefficient are fused through a multi-scale feature fusion mechanism to obtain the fused multi-scale features;

[0079] Step S104: Use the cross-layer feedback adaptation mechanism to predict and correct multi-scale features, obtain the perceptual reliability score and weight adjustment parameters, and input the weight adjustment parameters into the muscle strength grading adaptive mechanism for secondary optimization.

[0080] Specifically, such as Figure 3 As shown, the system initialization and hardware connection are as follows: An 8-channel surface electromyography (EMG) sensor is attached to the patient's gastrocnemius muscle, in the middle of the muscle belly, at a spacing of 2 cm. An S-shaped force sensor is installed at the contact point between the ankle exoskeleton and the foot brace. A temperature and humidity sensor is fixed within 1 meter of the robot body from the patient. Parameter configuration: Patient information (age 65, post-stroke sequelae, left ankle muscle strength grade L=2) is entered through the human-computer interaction interface. The system automatically loads the initial weight parameters corresponding to the muscle strength grade. W 1 = 0.44 W 2 = 0.56. Network loading: Call the pre-trained... CNN, LSTM and GAN Model parameters, storage path: / model / rehab_force_v2.0.h5, initialize the buffers of each module.

[0081] Signal acquisition and preprocessing, EMGSignal: Acquired at 1kHz frequency, filtered by a 50Hz notch filter to remove mains interference, and then bandpass filtered from 20-500Hz. Passed through a 200ms sliding window with a 50ms step, 12 features including the root mean square (RMS) and waveform length (WL) are calculated, generating a 128-dimensional vector. Example: RMS = 0.025mV, WL = 1.2mV•ms. Force signal: Acquired at 2kHz frequency, filtered by a 100Hz anti-aliasing filter, and then passed through a 100ms sliding window with a 20ms step, 8 features including the instantaneous force value (3.2N) and force change rate (0.08N / ms) are calculated, generating a 64-dimensional vector.

[0082] Feature extraction and enhancement EMG Channel: 128-dimensional feature input CNN The first layer uses a 3×3 convolutional kernel to extract local features (output 32×126×126), the second layer uses a 5×5 convolutional kernel to compress the dimensions (output 64×61×61), and the fully connected layer outputs 32-dimensional muscle strength features (including a spasticity warning value of 0.32 and a threshold of 0.5). Force channel: 64-dimensional feature input. LSTM The first layer processes short-term fluctuations (outputting 64 dimensions), and the second layer extracts long-term trends (outputting 32-dimensional contact force features with a force stability coefficient of 0.85). Adversity enhancement: GAN The generator enhances the 32-dimensional muscle strength features with a similarity of s=0.89 and outputs enhanced features (original features × 0.9 and generated features × 0.1).

[0083] Dynamic fusion and status assessment: At the 15th second, the patient attempted active dorsiflexion of the ankle joint. EMG The signal muscle tension M=1.3M0 (exceeding the 1.2M0 threshold) triggers weighted adjustment. W 1 = 0.4 + 0.2 × (1.3) M 0-1.2 M 0) / M 0 + 0.1 × (5 - 2) / 5 = 0.4 + 0.02 + 0.06 = 0.48. Within 10ms, a smooth transition using an exponentially weighted average occurs, resulting in W1 = 0.48 and W2 = 0.52. (This is the value of the perceived fusion force.) F 融合 =0.48×1.8N+0.52×3.2N=0.864+1.664=2.528N. Output status assessment results: Muscle strength level 2, perception confidence 96%, no abnormal warning.

[0084] Cross-layer interaction and prediction correction: Send to the decision layer every 10ms: F 融合=2.528N, predicted trend over the next 100ms, gradually increasing to 2.8N, reliability score 96%. Received decision-making layer instructions: Training mode = resistance training, m=2, adjustment step size s=0.3N, trigger force sensor weight lower limit adjusted to 0.5. At the 20th second, the deviation ΔE between the actual force value of 2.7N and the predicted value of 2.8N is 0.1N > 0.05N, calculate the correction coefficient. k c =1-0.1 / 0.5=0.8, the next predicted value is revised to 2.8×0.8=2.24N.

[0085] Calibration compensation execution: After running for 30 minutes, start periodic calibration: The robot end effector moves to the calibration station, sequentially contacts 5N and 10N weights, and records the results. F 5 = 4.8N F 10 =9.7N. Calculate. k =(10-5) / (9.7-4.8)=5 / 4.9≈1.02, b =5 - 1.02 × 4.8 ≈ 5 - 4.896 = 0.104. Force signal after correction: F 校正 =1.02× F 原始 +0.104. Environmental compensation: Temperature and humidity sensor readings T=28℃, H=60%, substituting into the formula: F 校正 = F 原始 ×(1+0.002×(28-25)+0.001×(60-50) / 50)= F 原始 ×(1+0.006+0.0002)= F 原始 ×1.0062. Time-varying drift compensation: runtime t =1h, β=0.001+0.0005×2 / 5=0.0012, F 补偿 = F 校正 -0.0012×1= F 校正 -0.0012.

[0086] Example of abnormal management: At the 45-minute mark, the patient experienced mild spasms. EMG Signal Muscle Tension M =1.5 M 0, force fluctuation coefficient F v =0.18N / ms: W 1Automatically adjusted to 0.65. W 2 =0.35. The multimodal filtering unit detected an anomaly in the force signal (correlation coefficient). r =0.2<0.3), initiate correction: F 修正 =0.7×2.1N+0.3×1.9N=1.47+0.57=2.04N. An anomaly warning is sent to the decision-making level, triggering a rapid reduction in the robot's output force (from 3N to 1N within 10ms).

[0087] In this optional embodiment, the steps of acquiring the original electromyography (EMG) signal and the original force signal, and using a generative adversarial network to perform feature enhancement on the original EMG signal and the original force signal respectively; and fusing the feature-enhanced original EMG signal and the original force signal based on a muscle strength grading adaptive mechanism to obtain a fused force perception value include:

[0088] An eight-channel surface electromyography (SEMG) sensor and an S-shaped force sensor were used to acquire raw EMG and raw force signals, respectively. Generative adversarial networks (GANs) were used to enhance the features of the raw EMG and raw force signals. Second-order Butterworth filters and fourth-order Bessel filters were then used to process the enhanced raw EMG and raw force signals, respectively, to construct a first-level feature vector (a 32-dimensional muscle force feature vector) and a second-level characteristic vector (a 32-dimensional contact force feature vector). Convolutional neural networks and long short-term memory networks were used to extract features from the first-level and second-level characteristic vectors, respectively, to obtain the muscle force feature vector and the contact force feature vector. An attention fusion mechanism was then used to compare the muscle force feature vector and the contact force feature vector with a pre-configured muscle force level (i.e., muscle force level). L The fusion is performed to obtain the fusion force perception value.

[0089] Specifically, such as Figure 4 As shown, a dual-channel parallel processing and attention-based dynamic fusion architecture is adopted, and a new muscle strength grading adaptive mechanism is added. The detailed formula algorithm and implementation steps are as follows:

[0090] set up EMG The muscle tension of the signal is M (Base value) M 0 The force fluctuation coefficient is F v The patient's muscle strength grade is L (Levels 0-5) EMG Feature weights W 1 Sum of force sensor feature weights W 2 The calculation formula is: ;whenF v ≥0.15 N / ms hour, Under normal circumstances, And always satisfy W 1 + W 2 =1.

[0091] Calculation of fusion force perception value: F 融合 = W 1 × F EMG + W 2 × F 力传感器 ,in F EMG for EMG The force-sensing value output by the channel. F 力传感器 This represents the force-sensing value output from the force sensor channel. An 8-channel surface electromyography (EMG) sensor with a 16-bit sampling resolution and a common-mode inhibition ratio >110dB is used. It is attached to the target muscle group on the patient, such as the gastrocnemius muscle, and data is acquired at a frequency of 1kHz. EMG The signal is transmitted to the processing unit via Bluetooth 5.0. A second-order Butterworth filter is used for 50Hz notch filtering with a quality factor Q=5 to remove power frequency interference; then, a 20-500Hz bandpass filter with a second-order Butterworth filter and an attenuation rate of 40dB / dec is applied to extract the effective electromyographic components; a sliding window with a window length of 200ms and a step size of 50ms is used to calculate 12 time-domain features, including the root mean square value and waveform length, to construct a 128-dimensional feature vector. (First layer) CNN Input a 128-dimensional feature vector, use 3×3 convolution kernels (32 in total), stride 1, activation function ReLU The first layer performs local temporal feature extraction, outputting a 32×126×126 feature map; the second CNN layer uses 5×5 convolutional kernels (64 kernels), stride 2, and activation function... ReLU The first layer captures the correlation features of muscle activity and outputs a 64×61×61 feature map. The third fully connected layer flattens the feature map into a 124032-dimensional vector and uses a dropout layer (dropout rate 0.3) to prevent overfitting, finally outputting a 32-dimensional muscle strength feature vector. F EMG .Will F EMG , F 力传感器 and muscle strength level L Input the attention fusion layer, call the dynamic weight calculation function, and update it every 10ms. W1 and W 2 .

[0092] An S-type force sensor with a range of 0-50N and a nonlinearity error of <0.1%FS is used, installed on the robot's end effector. Contact force signals are acquired at a frequency of 2kHz, converted by a 16-bit AD converter, and then input into the processing unit. The patient's muscle strength level (L, 0-5) is assessed by a rehabilitation physician and stored in the system parameter database. A fourth-order Bessel filter is used for anti-aliasing filtering with a cutoff frequency of 100Hz and a ripple coefficient of <0.1dB to remove high-frequency noise. A sliding window with a window length of 100ms and a step size of 20ms is used to calculate eight features, including instantaneous force value and rate of change of force, generating a 64-dimensional feature vector. (First layer) LSTM Input a 64-dimensional feature vector, 64 hidden units, a forget gate bias of 0.1, process short-term temporal dependencies of force fluctuations within 100ms, and output a 64-dimensional intermediate vector. Second layer. LSTM The system uses 32 hidden units, a forget gate bias of 0.2, and extracts the long-term trend of cumulative force changes over 500ms, outputting a 32-dimensional contact force feature vector. F 力传感器 .

[0093] In this optional embodiment, the feature enhancement of the original electromyography signal and the original force signal using a generative adversarial network includes:

[0094] The raw electromyography (EMG) signal and raw force signal are constructed with pre-configured clinical assessment results to obtain a training dataset. The generator network and the judge network are trained on the training dataset simultaneously. The key features of the generator network and the judge network are extracted using a knowledge distillation mechanism, and feature enhancement is performed on the input raw EMG signal and raw force signal.

[0095] Specifically, such as Figure 8 As shown, generative adversarial networks (GANs) are introduced. GAN Feature enhancement is performed to improve network robustness. The detailed formula, algorithm, and implementation steps are as follows:

[0096] Generator loss function: L G =- E zPz(z) [ log ( D ( G ( z )))],in G For generator, D For discriminator, z It is random noise. P z (z)For noise distribution. Discriminator loss function:

[0097] L D =- E x P data ( x )[ log ( D ( x ))]- E z P z ( z )[ log (1- D ( G ( z )))],in, x For real feature data, P data ( x The data distribution represents the actual data distribution. Training data from 1000 different patients (covering muscle strength grades 0-5) were collected, including... EMG A training dataset of 500,000 samples was constructed using signals (1kHz sampling), force sensor signals (2kHz sampling), and corresponding clinical assessment results. The dataset was then expanded to 2 million samples by adding Gaussian noise (signal-to-noise ratio 30dB), stretching the time axis (±10%), and scaling the amplitude (±5%) of the additive multimodal sensor data, thereby enhancing the model's generalization ability. The dataset was divided into a training set (80%), a validation set (10%), and a test set (10%), with stratified sampling based on muscle strength levels to ensure uniform data distribution.

[0098] In this optional embodiment, the step of simultaneously training the training dataset using a generator network and a judge network, extracting key features from the generator network and the judge network using a knowledge distillation mechanism, and performing feature enhancement on the input raw electromyography signal and raw force signal includes:

[0099] Random noise is input into the generator network to generate simulated feature vectors. A discriminator network distinguishes between real and generated first-order simulated features (i.e., 32-dimensional simulated feature vectors). The loss function of the first-order simulated features is optimized using local feature matching terms, and a knowledge distillation mechanism is used to compress the first-order simulated features into second-order key features (i.e., 16-dimensional key features). An adaptive moment estimation optimizer (i.e., the Adam optimizer) is used to input the second-order key features into the original electromyography signal and the original force signal for feature enhancement.

[0100] Specifically, the generator employs a 3-layer deconvolutional network, with 100-dimensional random noise as input. The first deconvolutional layer has a 4×4 kernel (64 output channels), the second layer has a 4×4 kernel (32 output channels), and the third layer has a 4×4 kernel (1 output channel). The activation function is... LeakyReLU (Negative slope 0.2) and Tanh Output a 32-dimensional simulated feature vector.

[0101] Discriminator: Employs a 3-layer convolutional network, taking a 32-dimensional feature vector as input (real or generated). The first layer has a 4×4 kernel (32 output channels), the second layer has a 4×4 kernel (64 output channels), and the third layer has a 4×4 kernel (1 output channel). The activation function is... LeakyReLU (Negative slope 0.2) and Sigmoid The output discrimination probability is (0-1). A distillation layer is added between the generator and the discriminator, using knowledge distillation to compress the 32-dimensional generated features into 16-dimensional key features. Local salient features (such as...) are preserved. EMG The instantaneous impulse of energy and force signals in the 50-100Hz frequency band is filtered by attention weights (local feature weights ≥ 0.7). Distillation loss function: W is a 16-dimensional to 32-dimensional mapping matrix, ensuring that over 90% of the key information is retained after compression. The discriminator loss function is optimized, and a local feature matching term is added. Local features include EMG The peak factor of the signal, the rising slope of the force signal, etc.

[0102] In this optional embodiment, the process of fusing the muscle strength feature vector and the contact force feature vector with the pre-configured muscle strength level based on the attention fusion mechanism to obtain the fused force perception value includes: using a dynamic weight calculation function to calculate the weights of the muscle strength feature vector and the contact force feature vector with the pre-configured muscle strength level to obtain a primary weight and a secondary weight; using a weight smoothing transition mechanism to adjust the weights of the primary weight and the secondary weight; and obtaining the fused force perception value through the fused force perception value calculation formula.

[0103] Specifically, F EMG , F 力传感器 and muscle strength level L Input the attention fusion layer, call the dynamic weight calculation function, and update it every 10ms. W 1 and W 2 When detected M ≥1.2 MAt time 0, the weight smoothing transition mechanism is activated, using an exponentially weighted average with a smoothing coefficient of 0.8, and the weight adjustment is completed within 10ms.

[0104] In this optional embodiment, the step of using a dynamic weight calculation function to calculate the weights of the muscle strength feature vector and the contact force feature vector with the pre-configured muscle strength level to obtain the primary weight and secondary weight includes: using the dynamic weight calculation function to monitor muscle tension and force fluctuation coefficient in real time, triggering different target weight calculation paths according to whether they exceed a preset threshold, and generating an initial target weight value; constraining the initial target weight value by limiting the range of the initial target weight value, and truncating the initial target weight value that exceeds the limit to obtain the primary weight and secondary weight.

[0105] Specifically, such as Figure 5 As shown, the fusion force perception value is obtained through the formula. F 融合 The data is stored in shared memory for subsequent modules to access. This network enables force sensing accuracy to reach ±0.08N, a 75% improvement over traditional methods.

[0106] In this optional embodiment, the step of processing the acquired temperature and humidity data and calibration signal using an adaptive calibration mechanism and an adaptive compensation mechanism, extracting baseline drift correction values ​​and temperature compensation coefficients, and inputting the baseline drift correction values ​​into the muscle strength grading adaptive mechanism for optimization includes:

[0107] By triggering a periodic calibration process, the robot comes into contact with a standard load and collects sensor readings. The raw data is input into a linear correction formula to calculate proportional coefficients and offset parameters, which are then stored in a calibration parameter table, and the calibration time is recorded. Environmental data is continuously collected through temperature and humidity sensors. When environmental changes exceed a preset threshold, a compensation mechanism is triggered. Real-time data is input into a temperature compensation formula for correction, and a temperature compensation coefficient is calculated and stored in an environmental compensation parameter table. Combining the calibration time and pre-configured muscle strength level data, the original force signal is compensated by a time-varying drift compensation formula, and a baseline drift correction value is calculated. This baseline drift correction value is then input into a muscle strength grading adaptive mechanism for optimization. A three-dimensional coupling matrix is ​​constructed, and the temperature compensation coefficient and drift correction value are dynamically corrected in conjunction using an adaptive compensation triggering mechanism.

[0108] Specifically, such as Figure 7 As shown, a built-in adaptive calibration mechanism and a new environmental adaptive compensation structure are added. 1) Linear correction formula: F 校正 = k × F 原始 + b ,in k This is the proportionality coefficient. bThe offset is obtained through calibration using standard loads (5N, 10N). k =(10-5) / ( F 10 -F5), b =5- k × F 5 ( F 5. F 10 (These are the original readings corresponding to 5N and 10N loads). 2) Temperature compensation formula: F 校正 = F 原始 ×(1+0.002×( T -25)+0.001×( H -50) / 50), where T The ambient temperature is (°C). H 3) Environmental humidity (%). Time-varying drift compensation formula: F 补偿 = F 校正 - β × t ,in β =0.001+0.0005× L / 5 (drift coefficient, related to muscle strength grade), t Runtime (hours).

[0109] The calibration process is triggered every 30 minutes. The robot's end effector automatically moves to the calibration station, sequentially contacts 5N and 10N standard weights, and records the original sensor readings. F 5 and F 10 Each load was kept in contact for 500ms, and 10 sets of data were continuously collected and averaged to reduce random errors. The proportional coefficient was then calculated using the linear correction formula. k and offset b If the calculated k If the deviation exceeds 1 ± 0.05 (the normal linearity deviation threshold), a secondary calibration is triggered. Data is re-acquired and verified. If the deviation still exceeds the range, a calibration anomaly alarm is generated. The calculated... k and b The calibration parameters are stored in the calibration parameter table, overwriting historical parameters, and a calibration timestamp is recorded for traceability analysis. After calibration, the robot's end effector automatically resets to its working position. The calibration process takes less than 10 seconds and does not affect normal training procedures. 1. Drift compensation and real-time monitoring. EMGSignal baseline: Baseline values ​​are calculated using a sliding window (1s window length, 0.5s step). When the baseline drift exceeds 15% for three consecutive windows, a baseline reset procedure is initiated. Baseline reset: The stable segments from the most recent 30 minutes are retrieved. EMG Data (collected while the patient is at rest), baseline values ​​were recalculated, and adjustments were made. CNN The input layer bias parameters of the subnetwork (bias correction = new baseline - old baseline) ensure the stability of feature extraction. 2. Time-varying drift compensation: Timing starts from system startup, and the runtime is read every hour. t According to muscle strength level L Calculate the drift coefficient β The force signal is compensated by substituting the time-varying drift compensation formula, and the compensation value is updated every 10ms. 3. Environmental compensation: An integrated SHT30 temperature and humidity sensor (temperature measurement range -40~125℃, accuracy ±0.3℃; humidity measurement range 0~100%RH, accuracy ±2%RH) collects environmental temperature and humidity data T and H every 500ms. When the temperature change exceeds 2℃ or the humidity change exceeds 10%RH, the environmental compensation mechanism is triggered, and T and H are substituted into the temperature compensation formula to correct the force sensor output. An environmental compensation parameter table is established to record the compensation coefficients under different temperatures and humidity levels. The compensation accuracy is optimized by linear interpolation to ensure that the compensation error is ≤0.05N under extreme environments (temperature 15℃ / 35℃, humidity 30%RH / 70%RH). A three-dimensional coupling matrix is ​​constructed. K ( L,T,H ),in L Muscle strength grade (0-5) T For ambient temperature (15-35℃) H The humidity level (30%-70%RH) was obtained through an environmental adaptability experiment involving 1000 patients. K ( L,T,H ) = 1 + 0.02 × ( L / 5)+0.003×( T -25)-0.001×( H -50); This matrix is ​​used for dynamic correction. EMG Signal baseline: M 0 ' = M 0 × K ( L,T,H ); M 0 (As the initial baseline), the baseline changes in tandem with muscle strength and environment.

[0110] When a change in muscle strength grade ≥1 and a temperature change ≥3℃ are detected, coupling compensation is automatically activated to avoid overcorrection caused by a single factor. The compensation cycle is dynamically adjusted according to the coupling strength: strong coupling (e.g., L =Level 5+ T When the temperature is 35℃, compensation is performed every 5 minutes; when the coupling is weak, compensation is performed every 30 minutes.

[0111] In this optional embodiment, the step of fusing the extracted fusion force sensing value and temperature compensation coefficient through a multi-scale feature fusion mechanism to obtain the fused multi-scale features includes:

[0112] The force sensing value and temperature compensation coefficient are fed into an adaptive Kalman filter and a multi-scale fusion module for processing, respectively, to generate a preliminary filtered force signal estimate and a multi-scale fused force signal. By calculating the electromyographic muscle activation index and performing correlation analysis with the multi-scale force signal, a correction mechanism is activated to generate a corrected force value when an anomaly is detected. At the same time, physiological delay is calculated through dynamic time warping and the signal is corrected for time offset. The time-corrected multi-modal signals are weighted and fused using a cross-modal spatiotemporal weighted fusion formula, and finally, time-aligned multi-scale features are output.

[0113] Specifically, such as Figure 6 As shown, an innovative master-slave module reconfigurable filtering structure is designed, a new multi-scale feature fusion structure is added, and the adaptive Kalman filter formula is used for prediction. ;in, This is the state estimate. k Indicates the first k The observed value at time, T This represents the matrix transpose operation. Covariance prediction: Kalman gain: ;renew: ;in, This is the state estimate. A =[10.01;01] is the state transition matrix. B =[0.5;1] is the control input matrix, u To control the input, P Let covariance matrix be the variance matrix. Q =diag([0.01,0.001]) represents the process noise covariance. H =

[10] is the observation matrix, R =0.001 represents the observation noise covariance. z For the observed values, K Kalman gain. Multi-scale feature fusion formula: ,in F 高频 This is the force signal after high-frequency filtering.F 低频 The force signal is after low-frequency filtering, α = 0.5 + 0.2 × sin(2πf / 10) (f is the main frequency of the signal, with a dynamic adjustment range of 0.3-0.7).

[0114] Main module: Loading adaptive Kalman filter parameters ( A, B, Q, H, R Initial values), initial state estimates covariance matrix From the module: Initialize the feature mapping interface (using TCP / IP protocol, port number 5000), set... EMG The feature transmission frame rate is 100Hz. Multi-scale fusion module: Initial α is set to 0.5, and the high-frequency filter (50Hz Butterworth high-pass) and low-frequency filter (10Hz Butterworth low-pass) are initialized. Signal input: The master module receives the raw force sensor signal (2kHz) via the SPI interface, buffering data every 5ms. The slave module receives data via the feature mapping interface. EMG The 32-dimensional muscle force feature vector (100Hz) of the signal is verified (CRC32 checksum) and then stored. Filtering: Main module: Performs adaptive Kalman filtering on each frame of force signal, first calculating the predicted value. Covariance Prediction Then based on the observed values z k Calculate Kalman gain K k Last updated state estimate Extraction force signal estimate .

[0115] Multi-scale fusion module: F 卡尔曼 Inputting the high-frequency and low-frequency filters respectively, we obtain F 高频 and F 低频 The signal's dominant frequency f is calculated using Fourier transform, and the α value is updated; then, it is substituted into the multi-scale feature fusion formula to obtain... F 多尺度 From module: EMG Muscle strength features were normalized (mapped to the 0-1 range) and muscle activation index was calculated. ; ( w i For feature weights, f i (As feature values), which are then transmitted to the main module. Cross-validation: The main module calculates... F 多尺度 and EMG Activation Index A correlation coefficient r(Using Pearson correlation coefficient). When | r When |<0.3 (signal anomaly threshold), the correction mechanism is activated: F 修正 =0.7× F 多尺度 +0.3×F EMG映射 (F) EMG映射 for EMG (Force value of feature mapping). Module switching: Real-time monitoring of module health status (CPU utilization <80%, communication latency <5ms). When the main module is abnormal, hot switching is triggered: the slave module switches to the main mode, loads backup filtering parameters, and the switching process is seamlessly connected through a double buffering mechanism (switching time ≤10ms).

[0116] The Dynamic Time Warping (DTW) correction module calculates the time difference Δ between the peak time of the EMG signal and the peak time of the force signal in real time. t (Locating peak values ​​using cross-correlation coefficients), establish a delay mapping table: when muscle strength level L When the level is 0-1, Δ t ≈80ms; L When the level is 4-5, Δ t ≈50ms. Based on Δ t Time offset correction for force sensor signals: advance the force signal by Δ on the time axis. t ,make EMG Features and force features are aligned in the physiological response timeline. Cross-modal attention spatiotemporal weights, in addition to the original dynamic weights, include spatiotemporal alignment weights. The formula is: (60ms is the optimal physiological delay baseline); the final fusion formula is optimized as follows: The further Δt deviates from 60ms, the more... right EMG The more significant the weight decay of a feature, the more likely the force sensor signal with a better time-series match will be trusted.

[0117] In this optional embodiment, the step of using a cross-layer feedback adaptation mechanism to predict and correct multi-scale features, obtaining a perceptual reliability score and weight adjustment parameters, and inputting the weight adjustment parameters into a muscle strength grading adaptive mechanism for secondary optimization includes: constructing a bidirectional interaction mechanism based on a pre-configured decision layer through an initial communication interface, receiving adjustment parameters and training modes issued by the decision layer, and establishing a parameter mapping table and threshold adjustment strategy; reading historical fused force data and inputting it into a Long Short-Term Memory (LSTM) prediction network to generate a future force change trend curve, and generating a dynamic correction coefficient by comparing the deviation between the actual measured value and the predicted value to calibrate the predicted trajectory; calculating the perceptual reliability score to obtain the score result, generating a data frame from the real-time force signal, the calibrated predicted trajectory, and the reliability score result, and feeding it back to the decision layer for optimization.

[0118] Specifically, a two-way interaction mechanism with the decision-making level is constructed, a new prediction-correction closed-loop structure is added, the interaction interface is initialized, and a CAN bus (1Mbps baud rate) is used as the communication interface with the decision-making level, defining the data frame format (ID=0x123, data length 8 bytes). Prediction correction module initialization: settings... LSTM Predict subnetwork parameters (input dimension 32, 2 hidden layers × 64 units, output dimension 1, learning rate 0.001), and correct initial values ​​for coefficients. k c =1. Prediction correction: Acquire the actual force value in the next 10ms. f 实际 Calculate the corresponding prediction points f Deviation Δ of 1 E =| f 实际 - f 1|. When Δ E When the value is greater than 0.05N, update the correction factor. k c =1-Δ E / 0.5 (limited range 0.5-1); otherwise, keep k c =1. Adjust the remaining predicted points: Information output: Calculate the perceived reliability score S = 100 × (1 - Δ) E / 0.2) (Limited range 0-100). Packaging F 融合 Corrected trend curve F 预测The score S is a CAN data frame sent to the decision layer. Parameter reception: Decision layer data frames are received every 10ms, and parsed to obtain the adjustment step size s (in N) and training mode m (0 = passive training, 1 = active assistance, 2 = resistance training). A parameter mapping table is established: When m = 2 (resistance training), the lower limit of the force sensor weight is set. W 2min =0.5; When s≥0.2N, weight locking mode is activated. Threshold adjustment: Read the current... W 1 and W 2 Value, if W 2 < W 2min Then force adjustment W 2 = W 2min , W 1 =1- W 2min The anomaly detection threshold is updated based on the training mode: during passive training, the force fluctuation coefficient threshold is lowered to 0.1 N / ms; during resistance training, it is raised to 0.2 N / ms. Trend prediction: Reading data from shared memory over a period of nearly 500 ms... F 融合数据 (50 samples in total), input LSTM Predictive subnetwork. The network outputs a force change trend curve (10 prediction points) for the next 100ms, stored as an array. F 预 =[ f 1, f 2,..., f 10 ].

[0119] Figure 2 An embodiment of the force sensing network system for rehabilitation robots based on dynamic weight fusion of the present invention is shown.

[0120] In this optional embodiment, the force sensing network system for rehabilitation robots based on dynamic weight fusion includes:

[0121] The perception fusion module 201 is used to acquire the original electromyography (EMG) signal and the original force signal, and to perform feature enhancement on the original EMG signal and the original force signal using a generative adversarial network. Based on the muscle force grading adaptive mechanism, the feature-enhanced original EMG signal and the original force signal are fused to obtain the fused force perception value.

[0122] The correction and compensation module 202 is used to process the acquired temperature and humidity data and calibration signal using an adaptive calibration mechanism and an adaptive compensation mechanism, extract baseline drift correction values ​​and temperature compensation coefficients, and input the baseline drift correction values ​​into the muscle strength grading adaptive mechanism for optimization.

[0123] The feature fusion module 203 is used to fuse the extracted fusion force perception value and temperature compensation coefficient through a multi-scale feature fusion mechanism to obtain the fused multi-scale features.

[0124] The prediction correction module 204 is used to predict and correct multi-scale features using a cross-layer feedback adaptation mechanism, obtain the perceptual reliability score and weight adjustment parameters, and input the weight adjustment parameters into the muscle strength grading adaptive mechanism for secondary optimization.

[0125] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A force sensing network method for rehabilitation robots based on dynamic weight fusion, characterized in that, include: The original electromyography (EMG) signal and the original force signal were acquired, and the original EMG signal and the original force signal were enhanced using a generative adversarial network. Based on the muscle strength grading adaptive mechanism, the original electromyography signal after feature enhancement and the original force signal are fused to obtain the fused force perception value; The acquired temperature and humidity data and calibration signals are processed using adaptive calibration and adaptive compensation mechanisms to extract baseline drift correction values ​​and temperature compensation coefficients. The baseline drift correction values ​​are then input into the muscle strength grading adaptive mechanism for optimization. The extracted fusion force perception value and temperature compensation coefficient are fused using a multi-scale feature fusion mechanism to obtain the fused multi-scale features. A cross-layer feedback adaptation mechanism is used to predict and correct multi-scale features, obtain perceptual reliability scores and weight adjustment parameters, and input the weight adjustment parameters into the muscle strength grading adaptive mechanism for secondary optimization.

2. The force sensing network method for rehabilitation robots based on dynamic weight fusion according to claim 1, characterized in that, The original electromyography (EMG) signal and the original force signal are acquired, and a generative adversarial network is used to enhance the features of the original EMG signal and the original force signal respectively. Based on a muscle strength grading adaptive mechanism, the original electromyography signal after feature enhancement and the original force signal are fused separately to obtain fused force sensing values, including: An eight-channel surface electromyography (EMG) sensor and an S-shaped force sensor were used to acquire raw EMG signals and raw force signals, respectively. Generative adversarial networks were used to enhance the features of the original electromyography and force signals, respectively. The original electromyography signal and the original force signal after feature enhancement were processed by a second-order Butterworth filter and a fourth-order Bessel filter, respectively, to construct a first-level feature vector and a second-level feature vector. Convolutional neural networks and long short-term memory networks are used to extract features from the first-level feature vector and the second-level feature vector, respectively, to obtain the muscle force feature vector and the contact force feature vector. Based on the attention fusion mechanism, the muscle force feature vector and contact force feature vector are fused with the pre-configured muscle force level to obtain the fused force perception value.

3. The method for force sensing network of rehabilitation robot based on dynamic weight fusion according to claim 2, characterized in that, The feature enhancement of the original electromyography signal and the original force signal using a generative adversarial network includes: The original electromyography signals and original force signals were constructed and compared with pre-configured clinical assessment results to obtain the training dataset; The generator network and the judge network are trained on the training dataset simultaneously. The key features of the generator network and the judge network are extracted by knowledge distillation mechanism, and feature enhancement is performed on the input raw electromyography signal and raw force signal.

4. The force sensing network method for rehabilitation robots based on dynamic weight fusion according to claim 3, characterized in that, The method of simultaneously training the training dataset using a generator network and a judge network, extracting key features from the generator network and the judge network using a knowledge distillation mechanism, and performing feature enhancement on the input raw electromyography signal and raw force signal includes: Random noise is input into the generator network to generate simulated feature vectors, and a discriminator network is used to distinguish between real and generated first-order simulated features. The loss function of the first-order simulated features is optimized by using local feature matching terms, and the first-order simulated features are compressed into second-order key features by using a knowledge distillation mechanism. The second-order key features were input into the original electromyography signal and the original force signal using an adaptive moment estimation optimizer to enhance the features.

5. The method for force sensing network of rehabilitation robot based on dynamic weight fusion according to claim 2, characterized in that, The attention-based fusion mechanism fuses the muscle force feature vector and contact force feature vector with a pre-configured muscle force level to obtain a fused force perception value, including: The dynamic weight calculation function is used to calculate the weights of the muscle force feature vector and the contact force feature vector with the pre-configured muscle force level to obtain the first-level weight and the second-level weight. The weights of the first-level and second-level weights are adjusted using a weight smoothing transition mechanism, and the fusion force perception value is obtained through the fusion force perception value calculation formula.

6. The force sensing network method for rehabilitation robots based on dynamic weight fusion according to claim 5, characterized in that, The dynamic weight calculation function is used to calculate the weights of the muscle strength feature vector and the contact force feature vector with the pre-configured muscle strength level, resulting in primary weights and secondary weights, including: The dynamic weight calculation function is used to monitor muscle tension and force fluctuation coefficient in real time. Different target weight calculation paths are triggered according to whether the preset threshold is exceeded, and an initial target weight value is generated. The initial target weight value is constrained by limiting conditions, and the initial target weight value that exceeds the limit is truncated to obtain the first-level weight and the second-level weight.

7. The force sensing network method for rehabilitation robots based on dynamic weight fusion according to claim 1, characterized in that, The process of using adaptive calibration and adaptive compensation mechanisms to process the acquired temperature and humidity data and calibration signals, extracting baseline drift correction values ​​and temperature compensation coefficients, and inputting the baseline drift correction values ​​into the muscle strength grading adaptive mechanism for optimization includes: By triggering a periodic calibration process, the robot is brought into contact with a standard load and sensor readings are collected. The raw data is then input into a linear correction formula to calculate and generate proportional coefficients and offset parameters, which are then stored in a calibration parameter table, and the calibration time is recorded. The system continuously collects environmental data through temperature and humidity sensors. When environmental changes exceed a preset threshold, a compensation mechanism is triggered. Real-time data is input into the temperature compensation formula for correction, and a temperature compensation coefficient is calculated and stored in the environmental compensation parameter table. Combining calibration time and pre-configured muscle strength level data, the original force signal is compensated by inputting into the time-varying drift compensation formula, a baseline drift correction value is calculated, and the baseline drift correction value is input into the muscle strength grading adaptive mechanism for optimization. A three-dimensional coupling matrix is ​​constructed, and an adaptive compensation triggering mechanism is used to dynamically correct the temperature compensation coefficient and drift correction value in tandem.

8. The force sensing network method for rehabilitation robots based on dynamic weight fusion according to claim 1, characterized in that, The extracted fusion force sensing value and temperature compensation coefficient are fused using a multi-scale feature fusion mechanism to obtain the fused multi-scale features, including: The force sensing value and temperature compensation coefficient are fused and input into the adaptive Kalman filter and the multi-scale fusion module for processing, respectively, to generate the preliminary filtered force signal estimate and the multi-scale fused force signal. By calculating the muscle activation index of electromyography and performing correlation analysis with multi-scale force signals, a correction mechanism is activated to generate a correction force value when abnormalities are detected. At the same time, physiological delay is calculated through dynamic time warping and the signal is corrected for time offset. By using a cross-modal spatiotemporal weighted fusion formula, time-corrected multimodal signals are weighted and fused to finally output time-aligned multi-scale features.

9. The method for force sensing network of rehabilitation robot based on dynamic weight fusion according to claim 1, characterized in that, The step of using a cross-layer feedback adaptation mechanism to predict and correct multi-scale features, obtaining perceptual reliability scores and weight adjustment parameters, and then inputting the weight adjustment parameters into a muscle strength grading adaptive mechanism for secondary optimization includes: Based on the pre-configured decision layer, a two-way interaction mechanism is built through the initialization of the communication interface to receive the adjustment parameters and training modes issued by the decision layer, and to establish a parameter mapping table and threshold adjustment strategy. By reading historical fusion force data and inputting it into a long short-term memory prediction network, a future force change trend curve is generated. By comparing the deviation between the actual measured value and the predicted value, a dynamic correction coefficient is generated to calibrate the prediction trajectory. By calculating the perceived reliability score, the score result is obtained. The real-time force signal, the calibrated predicted trajectory and the reliability score result are generated into a data frame and sent back to the decision-making level for optimization.

10. A force sensing network system for rehabilitation robots based on dynamic weight fusion, characterized in that, include: The perception fusion module is used to acquire the original electromyography (EMG) signal and the original force signal. It uses a generative adversarial network to enhance the features of the original EMG signal and the original force signal respectively. Based on the muscle force grading adaptive mechanism, the feature-enhanced original EMG signal and the original force signal are fused to obtain the fused force perception value. The correction and compensation module is used to process the acquired temperature and humidity data and calibration signals using the adaptive calibration mechanism and the adaptive compensation mechanism, extract the baseline drift correction value and temperature compensation coefficient, and input the baseline drift correction value into the muscle strength grading adaptive mechanism for optimization. The feature fusion module is used to fuse the extracted fusion force perception value and temperature compensation coefficient through a multi-scale feature fusion mechanism to obtain the fused multi-scale features. The prediction and correction module is used to predict and correct multi-scale features using a cross-layer feedback adaptation mechanism, obtain the perceptual reliability score and weight adjustment parameters, and input the weight adjustment parameters into the muscle strength grading adaptive mechanism for secondary optimization.

Images