Fault-tolerant reconfiguration method for sensor failure of spatial dexterous hand

Abstract

The space dexterous hand sensor fault tolerance reconstruction method relates to the field of robot control. In order to solve the defects that the dexterous hand has weak fault tolerance ability of on-orbit sensor, insufficient model generalization ability and cannot adapt to the change of complex space environment in the prior art, the technical scheme provided by the space dexterous hand sensor fault tolerance reconstruction method comprises the following steps: a conditional variational autoencoder model is constructed, joint position and rope length sequences are input, time sequence features are extracted, and reconstructed joint position mean and predicted variance are output; the reconstructed joint position mean is compared with actual joint data, a dynamic confidence interval is generated, and it is judged whether the data is faulty; for the data judged as normal, the information entropy is calculated based on the predicted variance, and the data samples with information entropy exceeding the set threshold are screened and written into a dynamic buffer area; when the buffer area data is full, incremental learning is triggered, and the model parameters are updated by using the buffer area data. The space dexterous hand sensor fault tolerance reconstruction method is suitable for real-time detection and fault tolerance reconstruction of sensor faults of the space dexterous hand in the long-term operation process on orbit.

Description

Technical Field

[0001] This relates to the field of robot control and belongs to the fault-tolerant reconfiguration method of on-orbit space dexterous hand. Background Technology

[0002] As the complexity of on-orbit operations continues to increase, space robots are developing towards greater intelligence and flexibility. As an important form of robot end effector, the space dexterous hand possesses multi-degree-of-freedom and multi-finger control capabilities, enabling it to perform a variety of high-precision on-orbit operations such as grasping, assembly, and maintenance, and has become one of the key components in the development of space service robots.

[0003] However, the space environment differs significantly from the terrestrial environment, exhibiting complex characteristics such as extreme temperature differences, strong radiation interference, magnetic field fluctuations, and microgravity. These factors pose a severe challenge to the stability of dexterous hand sensors. Especially in dexterous hand systems with high degrees of freedom and multiple redundancies, a failure in any joint or sensor can trigger a chain reaction of failures, severely impacting the robot's overall performance and task execution capabilities.

[0004] To ensure mission safety and system robustness, scholars both domestically and internationally have conducted extensive research on the detection and fault-tolerant control of robot sensor faults. Common research methods include: Model-based methods, such as sliding mode observers and Kalman filters, achieve fault detection and variable estimation by constructing a system dynamics model. These methods rely on model accuracy and are difficult to adapt to environmental changes or nonlinear systems.

[0005] Data-driven methods, such as neural networks and autoencoders, learn historical data patterns to achieve fault prediction and variable reconstruction. For example, some studies have proposed using Long Short-Term Memory (LSTM) networks to predict robot joint angles and detect abnormal sensor outputs. However, most of these methods are trained based on ground environment data and do not fully consider the non-stationary disturbances and data scarcity issues in the on-orbit environment.

[0006] Fault tolerance methods based on redundant hardware: This involves adding multiple sensors to achieve backup redundancy and comparing the outputs to determine the fault. However, this approach introduces additional load and structural complexity in space missions, making it difficult to adapt in the long term.

[0007] In recent years, some studies have attempted to apply generative models such as variational autoencoders (VAEs) and conditional VAEs (CVAEs) to robot state reconstruction, but they still have shortcomings in areas such as high-degree-of-freedom systems, dynamic environment adaptability, and online update capabilities. Furthermore, existing fault diagnosis techniques mostly use static or sliding thresholds to determine anomalies, making it difficult to address multi-source disturbances in the space environment and system drift problems caused by long-term operation.

[0008] In summary, existing technologies suffer from drawbacks such as weak fault tolerance of on-orbit sensors, insufficient model generalization ability, and inability to adapt to the complex changes in the space environment. Summary of the Invention

[0009] To address the shortcomings of existing technologies, such as weak fault tolerance of on-orbit sensors, insufficient model generalization ability, and inability to adapt to the complex changes in the space environment, the technical solution provided by this invention is as follows: Fault-tolerant reconfiguration methods for sensor failures in spatial dexterous hands include: The steps of constructing a conditional variational autoencoder model are: inputting joint position and rope length sequences, extracting temporal features, and outputting the reconstructed joint position mean and prediction variance. The steps include comparing the reconstructed joint position mean with the actual joint data, performing Bayesian inference based on the reconstruction error and prediction variance, generating dynamic confidence intervals, and determining whether the data is faulty. For data determined to be normal, the information entropy is calculated based on the prediction variance, and data samples whose information entropy exceeds a set threshold are written into a dynamic buffer. When the buffer is full, incremental learning is triggered, which involves updating the model parameters using the buffer data.

[0010] Furthermore, a preferred embodiment is provided, wherein the conditional variational autoencoder model includes a bidirectional temporal encoder and a conditional decoupled decoder, wherein the bidirectional temporal encoder is used to extract the spatiotemporal features of the input data, and the decoder outputs the mean and variance of the reconstructed joint positions respectively.

[0011] Furthermore, a preferred implementation is provided, in which Bayesian inference constructs a normal distribution model from the reconstruction error and the prediction variance, and generates a dynamic confidence interval based on the model to determine whether the actual sensor data is an outlier.

[0012] Furthermore, a preferred implementation is provided in which information entropy is calculated based on prediction variance to measure the uncertainty of the model for the current sample, and data with information entropy exceeding a preset threshold are considered representative and written into a buffer.

[0013] Furthermore, a preferred implementation is provided in which the capacity of the dynamic buffer is set via grid search.

[0014] Furthermore, a preferred implementation method is provided, in which an elastic weight consolidation mechanism based on the Fisher information matrix is ​​used during the incremental learning process to constrain the changes in key parameters.

[0015] It also provides a fault-tolerant reconfiguration device for spatial dexterity hand sensor failures, including: A conditional variational autoencoder model is constructed, which takes the joint position and rope length sequence as input, extracts temporal features, and outputs a module that reconstructs the mean and prediction variance of the joint position. The module compares the reconstructed joint position mean with the actual joint data, performs Bayesian inference based on the reconstruction error and prediction variance, generates dynamic confidence intervals, and determines whether the data is faulty. For data that is determined to be normal, the module calculates the information entropy based on the prediction variance and writes data samples whose information entropy exceeds a set threshold into a dynamic buffer. When the buffer is full, incremental learning is triggered, and the module updates the model parameters using the buffer data.

[0016] A computer storage medium is also provided for storing a computer program, which, when read by the computer, executes the method.

[0017] A computer is also provided, including a processor and a storage medium, wherein the computer executes the method when the processor reads a computer program stored in the storage medium.

[0018] A computer program product is also provided, which, when executed, implements the method described.

[0019] Compared with the prior art, the advantages of the technical solution provided by the present invention are as follows: This invention improves the joint angle reconstruction capability of high-degree-of-freedom spatial dexterity hands under conditions of damaged or distorted sensor data by constructing an improved CVAE architecture that includes a bidirectional temporal encoder and a conditionally decoupled decoder. Bidirectional LSTM extracts temporal features, improving the modeling accuracy of the nonlinear relationship between joint position and rope length. Compared with traditional VAE methods using static encoders, this approach can more accurately capture motion continuity and temporal logic, thereby improving the accuracy of joint reconstruction.

[0020] This invention introduces a dual-path output structure at the decoder end, using both the position mean and variance, enabling the model to estimate prediction uncertainty while outputting prediction results. Unlike traditional neural networks that only output point estimates, this approach can dynamically adjust the system's confidence in prediction results when perception is impaired, providing a reliable basis for subsequent fault detection and fault-tolerant control, and enhancing the system's adaptability to unstable factors in the space environment.

[0021] The loss function designed in this invention comprehensively considers the reconstruction error, KL divergence, and the regularization term of the prediction variance, effectively avoiding the speculative training behavior of traditional CVAE models that masks errors by increasing the prediction variance. While ensuring the accuracy of the model output, it enhances the physical interpretability of the prediction distribution. Compared to CVAE methods using fixed regularization terms, this design exhibits stronger generalization ability in spatially dynamic scenarios.

[0022] This invention employs a Bayesian inference-based adaptive thresholding method (BAT), which dynamically generates confidence intervals using the predicted mean and variance of a CVAE model to achieve real-time detection of sensor faults. Compared to common fixed threshold and sliding window strategies, this method dynamically adjusts the detection sensitivity based on model uncertainty, significantly reducing false alarm and false negative rates, and exhibiting higher diagnostic accuracy when dealing with long-term changing sensor characteristics in the space environment.

[0023] This invention introduces a dynamic incremental buffer (DIB) management strategy based on information entropy. By combining the information entropy of the data calculated from the model's prediction variance, it selects only new data with high credibility and representativeness for model updates, reducing interference from invalid data. Compared with traditional methods without a selection mechanism or based on mean difference selection, this strategy can control learning costs and memory usage while ensuring the effectiveness of incremental model learning.

[0024] This invention employs an EWC (Elastic Weight Consolidation) strategy for incremental learning, effectively avoiding catastrophic forgetting and ensuring that the model retains its performance on older tasks while learning new data. Compared to conventional SGD or Adam training methods, this approach maintains model stability and long-term memory when facing slow but long-term changes in data distribution in spatial environments, enhancing the model's ability to adapt to complex task environments.

[0025] This invention introduces a dynamically adjusted learning rate mechanism in incremental training, adjusting the parameter update rate based on sample variance, thereby improving the model's sensitivity to sample changes and its convergence speed. Compared to a fixed learning rate mechanism, this approach is more suitable for space scenarios where sensor accuracy varies with the mission cycle, enabling the model to respond more quickly to input disturbances caused by the dynamic environment.

[0026] It is suitable for real-time detection and fault-tolerant reconfiguration of sensor faults during long-term on-orbit operation of space dexterous hands. Detailed Implementation

[0027] To make the advantages and benefits of the technical solution provided by the present invention clearer, the technical solution provided by the present invention will now be described in further detail, specifically: Implementation Method 1: This implementation method provides a fault-tolerant reconfiguration method for sensing faults in a spatial dexterous hand, including: The steps of constructing a conditional variational autoencoder model are: inputting joint position and rope length sequences, extracting temporal features, and outputting the reconstructed joint position mean and prediction variance. The steps include comparing the reconstructed joint position mean with the actual joint data, performing Bayesian inference based on the reconstruction error and prediction variance, generating dynamic confidence intervals, and determining whether the data is faulty. For data determined to be normal, the information entropy is calculated based on the prediction variance, and data samples whose information entropy exceeds a set threshold are written into a dynamic buffer. When the buffer is full, incremental learning is triggered, which involves updating the model parameters using the buffer data.

[0028] The conditional variational autoencoder model includes a bidirectional temporal encoder and a conditional decoupled decoder. The bidirectional temporal encoder is used to extract the spatiotemporal features of the input data, and the decoder outputs the mean and variance of the reconstructed joint positions, respectively.

[0029] Bayesian inference constructs a normal distribution model from the reconstruction error and the prediction variance, and generates dynamic confidence intervals based on this model to determine whether the actual sensor data are outliers.

[0030] Information entropy is calculated based on prediction variance and is used to measure the uncertainty of the model for the current sample. Data with information entropy exceeding a preset threshold is considered representative and is written into a buffer.

[0031] The capacity of the dynamic buffer is set via grid search.

[0032] During incremental learning, an elastic weight consolidation mechanism based on the Fisher information matrix is ​​used to constrain changes in key parameters.

[0033] Implementation Method Two: This implementation method is a further detailed description of the technical solution provided in Implementation Method One, specifically: This embodiment provides a sensor fault-tolerant reconfiguration method suitable for on-orbit operation of a space dexterous hand. It combines an improved conditional variational autoencoder (CVAE) model, a Bayesian adaptive threshold fault detection mechanism, and a dynamic incremental learning strategy to improve the robustness and adaptability of the dexterous hand system in extreme environments.

[0034] Step 1: Construct an improved CVAE model and complete ground pre-training. First, a CVAE network structure comprising a bidirectional temporal encoder and a conditionally decoupled decoder is designed. The input to this structure includes joint position data as the target variable and related rope length data as the conditional variable, which are fed into the bidirectional LSTM encoder to extract temporal features. The extracted features are mapped to the latent space through a fully connected layer, outputting the mean and variance of the latent variables. After sampling using a reparameterization technique, the sampling results are input into the decoder. The decoder employs a dual-path structure to output the predicted mean and predicted variance of the joint positions, respectively, to construct the model's reconstruction result and its uncertainty estimate. Based on this structure, a loss function is designed, fusing the reconstruction loss, the KL divergence term, and the regularization term of the predicted variance to guide the model in pre-training on a ground simulation dataset.

[0035] Step 2: Perform Bayesian adaptive threshold fault detection using model output. The pre-trained CVAE model is deployed on the in-orbit system, receiving real-time cable length input and outputting the corresponding reconstructed joint position and prediction variance. The mean of the model output is used as the reconstruction result, and the prediction variance as the error estimate. This is compared with actual joint data from the sensor to obtain the reconstruction error. An error distribution model is constructed based on the reconstruction error and prediction uncertainty, and a dynamic confidence interval is calculated using Bayesian inference. An adjustable interval coefficient is introduced to construct a flexible detection boundary, and the threshold width is dynamically adjusted according to the model's confidence level of the current prediction, achieving adaptive control for fault detection. When the actual joint data falls outside the confidence interval, it is considered abnormal; otherwise, it is considered normal.

[0036] Step 3: Perform information entropy filtering and construct a dynamic incremental buffer. For data samples deemed normal, the information entropy of the current data is further calculated based on the prediction variance of the model output. This entropy measures the model's cognitive uncertainty regarding the current input. High-value samples with information entropy exceeding a set threshold are written into a dynamic incremental buffer. This buffer is a first-in, first-out queue, storing a limited number of joint position and rope length sample pairs. To prevent the model from being interfered with by invalid data, the buffer capacity is set to an optimal value using a grid search method.

[0037] Step 4: Trigger online incremental learning and adopt an elastic weight consolidation strategy. When the dynamic buffer data volume reaches a set upper limit, the incremental learning process is triggered. First, the importance of each parameter to the original task is calculated from the pre-trained model parameters, forming a parameter importance evaluation matrix. A regularization term is introduced during incremental training to constrain the parameter adjustment range and prevent the forgetting of old knowledge. The buffer data is divided into several mini-batches, and the model output is calculated through forward propagation for each batch. The gradient is calculated based on the aforementioned multinomial loss function, and the parameters are updated using the AdamW optimizer. The learning rate is dynamically adjusted according to the prediction variance to adapt to the rhythm of changes in sample distribution, ensuring stable and convergent updates.

[0038] Step 5: Continuously iterate and update the CVAE model to adapt to changes in the orbital environment. Through the aforementioned online incremental learning process, the model continuously absorbs new data features, gradually adapting to the dynamic changes in the space environment, such as sensor aging, performance drift, and environmental interference, maintaining a highly robust state awareness capability. Simultaneously, relying on the elastic weight consolidation mechanism, it ensures that the model retains stable features acquired during ground pre-training while introducing new knowledge, ultimately forming a fault-tolerant reconfiguration system that can operate stably for extended periods in space-based missions.

[0039] Implementation Method 3: This implementation method further describes the technical solution provided above in detail through specific embodiments, specifically: This implementation method is designed to address the problem that existing high-degree-of-freedom dexterous hands cannot meet the on-orbit service requirements in terms of fault diagnosis and fault-tolerant control.

[0040] An online conditional variational autoencoder (Online-CVAE) model combining Bayesian inference is proposed to achieve on-orbit fault-tolerant reconstruction of a dexterous hand. The implementation steps are as follows: Design an improved CVAE architecture comprising a bidirectional temporal encoder, a sampling layer, and a conditional decoupling decoder. The encoder will input joint positions... and rope length The target and condition variables, as inputs, are extracted using a bidirectional LSTM layer for spatiotemporal feature extraction, and then mapped to the latent space through a fully connected layer, ultimately outputting the latent variables. mean and variance Latent variable sampling is achieved by introducing reparameterization techniques;

[0041] To simultaneously learn the reconstructed joint positions and their estimated uncertainties, a dual-path decoupling structure is employed in the decoder network to predict the mean joint positions separately. and variance Location mean It can output the reconstructed joint position corresponding to the rope length; variance This reflects the uncertainty in the model's predictions.

[0042] To maximize the input target variable Given condition variables Log-likelihood variational lower bound (ELBO) and loss function Designed as follows:

[0043] Reconstruction loss measures the decoder's ability to reconstruct input data from latent variables. Its core principle is to maximize the log-likelihood of the data, ensuring that the generated result is as close as possible to the original input. It is assumed that the decoder's output follows a Gaussian distribution. Reconstruction loss It can be defined as the negative log-likelihood:

[0044] KL divergence loss is used to regularize the latent space, making the distribution of latent variables in the encoder output more consistent. Approximating the pre-defined prior distribution To prevent a rapid decrease in KL divergence during the initial training phase, which could lead to a decline in model generation quality, a cosine annealing strategy is employed. The KL divergence is shown below:

[0045] The variance regularization term includes constraints on the magnitude and smoothness of the reconstructed variance, preventing the model from reducing the reconstruction loss by arbitrarily increasing the variance.

[0046] A fault detection method based on Bayesian Adaptive Threshold (BAT) is proposed. This method, together with the previously designed CVAE model, forms an online conditional variational autoencoder model incorporating Bayesian inference.

[0047] Bayesian methods utilize the reconstruction error output by the CVAE model. and variance Bayesian inference is performed to construct a dynamic fault detection framework. Assume reconstruction error... obey The normal distribution is used, and the variance of the reconstruction error is set. Variance of the CVAE model output The mean of reconstruction error The mean of the output of the CVAE model .

[0048] The dynamic confidence interval of the Bayesian method can be expressed as:

[0049] In the formula These are variable interval coefficients. The variance of the CVAE model output reflects the confidence level in the reconstructed values; therefore, an adaptive expansion factor is introduced. Variable interval coefficients It can be represented as

[0050] Posterior parameters They are designed as follows:

[0051]

[0052] In the formula, It is a set forgetting factor. For confidence level weights.

[0053] When the model prediction is reliable, the confidence weights Larger The closer A narrow threshold boundary ensures sensitivity to minor anomalies. Conversely, when model predictions are unreliable, Increasing the threshold margin expands the threshold, thereby reducing the false alarm rate. The hyperbolic tangent function tanh(⋅) ensures the smoothness of threshold adjustment, avoiding the impact of abrupt changes on the detection results.

[0054] The Bayesian fault detection method uses the uncertainty of the CVAE model output for Bayesian inference, adaptively adjusting the error detection interval and data distribution. Therefore, it participates in the angular input signal of the fault-tolerant impedance control system. It can be represented as:

[0055] After the system uses the Bayesian adaptive threshold method to perform anomaly detection on real-time data, if the data is determined to be normal and the information entropy of the data is... Greater than the threshold Then, the motor position and joint position data pairs are stored in a Dynamic Incremental Buffer (DIB). This implementation utilizes the prediction variance output by the decoder network of the CVAE model. Information entropy measures the uncertainty of a model's understanding of current data. The specific formula for calculating information entropy is as follows:

[0056] Information entropy threshold set to ,in and Let be the mean and standard deviation of the entropy of normal samples in the pre-training dataset. Let the capacity of the dynamic incremental buffer DIB be... When the accumulated data reaches this upper limit, the incremental learning process is triggered. The data within the DIB is... ,in For rope length data, This refers to joint angle data. Since a dynamically incremental buffer (DIB) with too small a capacity may lead to slow model convergence, while a capacity that is too large poses challenges to storage space and iteration time, this implementation uses a grid search method to select the most suitable DIB capacity.

[0057] For the incremental learning part, this implementation introduces an Elastic Weight Consolidation (EWC) strategy to avoid the catastrophic forgetting problem during incremental learning. The EWC strategy constrains the changes in model parameters by adding a regularization term to the loss function, allowing it to retain the memory of old knowledge while adapting to new data. First, M batches are sampled from the pre-training data distribution. Then for each parameter Calculate the Fisher matrix :

[0058] Where L is the original loss function. Characterize the importance of parameters to the old task. Define EWC regularization terms during incremental updates. for:

[0059] in This represents the optimal parameters obtained in the pre-trained network; These are the weighting coefficients of the EWC regularization term, controlling the strength of old knowledge retention. To balance the "stability-plasticity tradeoff" in incremental learning, this implementation method... Perform a grid search within the specified range.

[0060] Online incremental updates employ a batch gradient accumulation strategy. Data within the dynamic incremental buffer is divided into several sub-batches according to a preset mini-batch size. For each sub-batch, the network output is calculated using forward propagation. and variance The gradient is calculated based on the loss function. The AdamW optimizer is used for parameter updates, with the momentum parameter set to... and The learning rate is dynamically adjusted based on the data variance. Through this online incremental learning mechanism, the model can continuously receive the latest data while maintaining a memory of the original data distribution, thereby achieving adaptive updates to the dynamically changing space environment.

[0061] This implementation method is suitable for long-term on-orbit service and has strong anti-interference capabilities against complex characteristics such as extreme temperature differences, strong radiation interference, magnetic field fluctuations and microgravity.

[0062] This implementation method can reduce learning costs through ground pre-training and perform incremental learning by collecting actual on-orbit operation data, thereby adjusting the control strategy to adapt to the dynamic deviations generated by the long-term on-orbit operation of the high-degree-of-freedom dexterous hand.

[0063] This implementation uses a Bayesian adaptive threshold algorithm to determine sensor fault conditions, thereby improving the accuracy of system status acquisition and ensuring the effectiveness of online incremental learning. In this implementation, the loss of each fold during ground pre-training decreases continuously with the increase of training epochs. The training loss of all folds converges to the range of 1.09 ± 0.01 within 40 epochs, and this convergence trend verifies the stability and generalization ability of the model. The final average loss of five-fold cross-validation is 1.07, with a relative difference of less than 2.78%, indicating that the model has good robustness to changes in data distribution. This demonstrates that the Online-CVAE model constructed in this implementation can accurately predict joint angles. In space environment simulation experiments, the Bayesian adaptive threshold diagnosis used in this embodiment significantly improved the accuracy of fault diagnosis in the CVAE model compared with other fault diagnosis algorithms such as fixed threshold, sliding threshold, and LSTM prediction diagnosis. This is because on-orbit servicing is a long-term task. Due to equipment aging and damage, the dynamic performance of the manipulator and the detection accuracy of sensors can deviate. Furthermore, special interferences in the space environment, such as cosmic rays and high-energy particles, change over time. These factors challenge the real-time performance and robustness of the on-orbit servicing prediction system. Fixed threshold and LSTM prediction methods pre-set detection mechanisms on the ground and cannot adaptively adjust detection parameters according to changes in the surrounding environment during on-orbit servicing. Dynamic threshold methods are significantly less adaptable to environmental changes than the Bayesian adaptive threshold diagnosis proposed in this embodiment, making it difficult to adapt to complex and changing space environments. Therefore, the Bayesian adaptive threshold diagnosis used in this embodiment is more suitable for on-orbit servicing scenarios and achieves higher accuracy in fault diagnosis.

[0064] The above description of several specific embodiments further details the technical solution provided by the present invention in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, combinations of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault-tolerant reconstruction method for sensing faults in a spatial dexterous hand, characterized in that, include: The steps of constructing a conditional variational autoencoder model are: inputting joint position and rope length sequences, extracting temporal features, and outputting the reconstructed joint position mean and prediction variance. The steps include comparing the reconstructed joint position mean with the actual joint data, performing Bayesian inference based on the reconstruction error and prediction variance, generating dynamic confidence intervals, and determining whether the data is faulty. For data determined to be normal, the information entropy is calculated based on the prediction variance, and data samples whose information entropy exceeds a set threshold are written into a dynamic buffer. When the buffer is full, incremental learning is triggered, which involves updating the model parameters using the buffer data.

2. The fault-tolerant reconstruction method for sensing faults in a spatial dexterous hand according to claim 1, characterized in that, The conditional variational autoencoder model includes a bidirectional temporal encoder and a conditional decoupled decoder. The bidirectional temporal encoder is used to extract the spatiotemporal features of the input data, and the decoder outputs the mean and variance of the reconstructed joint positions, respectively.

3. The fault-tolerant reconstruction method for sensing faults in a spatial dexterous hand according to claim 1, characterized in that, Bayesian inference constructs a normal distribution model from the reconstruction error and the prediction variance, and generates dynamic confidence intervals based on this model to determine whether the actual sensor data are outliers.

4. The fault-tolerant reconstruction method for sensing faults in a spatial dexterous hand according to claim 1, characterized in that, Information entropy is calculated based on prediction variance and is used to measure the uncertainty of the model for the current sample. Data with information entropy exceeding a preset threshold is considered representative and is written into a buffer.

5. The fault-tolerant reconstruction method for sensing faults in a spatial dexterous hand according to claim 1, characterized in that, The capacity of the dynamic buffer is set via grid search.

6. The fault-tolerant reconstruction method for sensing faults in a spatial dexterous hand according to claim 1, characterized in that, During incremental learning, an elastic weight consolidation mechanism based on the Fisher information matrix is ​​used to constrain changes in key parameters.

7. A fault-tolerant reconfiguration device for sensing failures in a spatially dexterous hand, characterized in that, include: A conditional variational autoencoder model is constructed, which takes the joint position and rope length sequence as input, extracts temporal features, and outputs a module that reconstructs the mean and prediction variance of the joint position. The module compares the reconstructed joint position mean with the actual joint data, performs Bayesian inference based on the reconstruction error and prediction variance, generates dynamic confidence intervals, and determines whether the data is faulty. For data that is determined to be normal, the module calculates the information entropy based on the prediction variance and writes data samples whose information entropy exceeds a set threshold into a dynamic buffer. When the buffer is full, incremental learning is triggered, and the module updates the model parameters using the buffer data.

8. A computer storage medium for storing computer programs, characterized in that, When the computer program is read by the computer, the computer executes the method of claim 1.

9. A computer, comprising a processor and a storage medium, characterized in that, When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1.

10. A computer program product, as a computer program, is characterized by: When the computer program is executed, it implements the method of claim 1.

Images