Tensioning overall structure self-sensing system and method based on multi-mode strain sensing

Abstract

The invention discloses a tensegrity structure self-sensing system and method based on multi-mode strain sensing, and the method employs a conductive flexible sensor to replace a tendon in a tensegrity structure, and achieves the dual-mode response of a bending state and a stretching state. Obtaining sensor resistance values in a bending state and a stretching state, and obtaining a sensor strain value in the bending state through polynomial regression fitting; predicting through an LSTM network model to obtain a sensor strain value in a tensile state; and calculating a corresponding sensor length based on the obtained sensor strain value so as to determine space coordinates of nodes in the tensegrity structure and complete shape reconstruction.

Description

Technical Field

[0001] This invention relates to the field of shape reconstruction technology for tensioned integral structures, and in particular to a self-sensing system and method for tensioned integral structures based on multimodal strain sensing. Background Technology

[0002] In recent years, tensioned monolithic structures have been widely used in various fields such as disaster relief, space exploration, and biomimetic design due to their strong environmental adaptability. Focusing on deep space exploration, foreign research institutions have developed a six-bar tensioned monolithic robot to address space exploration missions to Titan. Driven by 12 brushless DC motors, the robot's tumbling is achieved by controlling the length of the ropes using these motors. Furthermore, dedicated simulation software NTRT has been developed to simulate and analyze the robot's dynamics and kinematic characteristics. After being launched into a predetermined orbit by a rocket, the robot is deployed directly from a high altitude, unfolding from a flat, folded state into a spherical shape, and then conducting surface exploration upon landing. However, during this process, collisions with the ground may damage the robot's mechanisms, rendering it immobile. Further research is needed to accurately assess the motion performance of tensioned monolithic structure robots and to develop their own state-sensing technology.

[0003] Existing research on state perception of tensioned monolithic structures is limited, and all approaches have significant limitations. Jonathan Bruce et al. precisely calculated the spatial distance between two nodes in the design by measuring the length changes of the tension elements in the connecting rods, but sensors capable of measuring such large deformations are very limited. Ken Caluwaerts et al. developed a technique based on unscented Kalman filtering (UKF) that integrates inertial measurement, ultra-wideband time-of-flight ranging data, and actuator state information. However, relying on external sensors, markers, or designated base stations to improve resolution is cumbersome and limits the flexibility of tensioned monolithic structures. Joran W. Booth et al. used a robot skin equipped with pneumatic actuators and embedded strain sensors to achieve real-time state perception of tensioned monolithic structures. However, due to the connection method between the sensors and the structure in the prototype, there was a deviation between the node spacing and the effective length of the sensors. Wen-Yung Li et al. achieved shape recognition using a tensioned monolithic structure equipped with flexible sensors and a recurrent neural network method, but this method involves acquiring data first and then reconstructing it, which is not a real-time process. These physics-based models are not well-suited for state reconstruction because they rely on consistent spring forces, cable tensions, and rod torques across system components. This results in a significantly high root mean square error (RMSE) for shape recognition, which can lead to serious biases and potential errors in applications requiring precise measurements, such as confined spaces. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a self-sensing system and method for tensioned monolithic structures based on multimodal strain sensing. This aims to solve a series of key technical problems in the field of tensioned monolithic structures: It achieves autonomous shape recognition without relying on external sensors (such as inertial measurement units or ultra-wideband base stations) and fixed markers, overcoming the limitations of complex environments on structural flexibility and autonomy; it also improves the accuracy of traditional physical models, avoiding shape reconstruction errors caused by the difficulty in obtaining mechanical parameters such as spring force, cable tension, and member torque. The dynamic response performance of the multimodal sensors under complex deformation is optimized, enhancing their durability and overcoming the nonlinear response and dynamic hysteresis problems during bending and tensile deformation. It also solves the delay problem caused by the complexity of recurrent neural network models, achieving a good balance between real-time performance and computational efficiency; furthermore, it enhances the robustness of the tensioned monolithic system in complex and extreme environments such as temperature and humidity changes, reducing the impact of sensor performance fluctuations.

[0005] The objective of this invention is achieved through the following technical solution: a self-sensing system for tensioned integral structures based on multimodal strain sensing, the system comprising:

[0006] The sensing system design module is used to replace the tendons in the tensioned overall structure with conductive and flexible sensors to achieve dual-mode response in bending and stretching states.

[0007] The polynomial curve fitting module is used to obtain strain values ​​based on the sensor resistance values ​​in a bending state through polynomial regression fitting.

[0008] The LSTM network model module is used to train an LSTM network model based on the historical resistance vector sequence under tensile conditions and its corresponding strain data sequence. The trained model is used to predict the sensor strain value under tensile conditions.

[0009] The 3D shape reconstruction module is used to calculate the corresponding sensor length based on the strain values ​​obtained from the polynomial curve fitting module and the LSTM network model module, thereby determining the spatial coordinates of the nodes in the overall tensioned structure and completing the shape reconstruction.

[0010] Furthermore, the sensor has a dual-mode response, capable of simultaneously sensing the bending and stretching states of the tensioned overall structure, including a conductive flexible diene rubber material sensor, a liquid metal sensor, or a capacitive flexible sensor.

[0011] Furthermore, the instantaneous resistance values ​​of all sensors are obtained based on the Arduino Mega data acquisition module, forming a set of real-time updated resistance vectors.

[0012] Furthermore, in the LSTM network model module, the current deformation is predicted by recursively calculating the node position at the previous time step, supporting autonomous monitoring in extreme environments.

[0013] Furthermore, the sensor length L is calculated as follows:

[0014]

[0015] in, The strain value of the sensor. For the pre-measured initial length vector of the tendon, This indicates the XOR operation.

[0016] Furthermore, in the 3D shape reconstruction module, the spatial coordinates of all nodes are determined by solving a system of nonlinear equations consisting of geometric constraints; the geometric constraints include constant link length constraints and dynamically changing tendon length constraints.

[0017] Furthermore, based on the obtained sensor strain values, the node coordinates are dynamically solved using geometric constraint equations and the least squares method, enabling real-time reconstruction of the 3D shape without the need for external sensors.

[0018] On the other hand, the present invention also provides a self-sensing method for tensioned integral structures based on multimodal strain sensing, the method comprising the following steps:

[0019] (1) A conductive and flexible sensor is used to replace the tendons in the tensioned overall structure to achieve a dual-mode response in bending and stretching states;

[0020] (2) Based on the sensor resistance value in the bending state, the strain value is obtained by polynomial regression fitting;

[0021] (3) Train an LSTM network model based on the historical resistance vector sequence and its corresponding strain data sequence under tensile conditions, and predict the sensor strain value under tensile conditions using the trained model.

[0022] (4) Calculate the corresponding sensor length based on the strain value obtained by the polynomial curve fitting module and the LSTM network model module, and then determine the spatial coordinates of the nodes in the tensioned overall structure to complete the shape reconstruction.

[0023] The beneficial effects of this invention: This invention proposes a self-sensing system and method for tensioned integral structures based on multimodal strain sensing. Compared with existing technologies, it achieves significant breakthroughs in core performance indicators such as accuracy, real-time performance, autonomy, and environmental adaptability. Specific comparisons are as follows:

[0024] 1. Improved accuracy: Significantly reduced node-level and system-level errors.

[0025] The 6-bar tension structure proposed by Jonathan Bruce et al. relies on external sensors (ultra-wideband base station, inertial measurement unit), and the root mean square error of node state estimation is as high as 45.6 mm. The robot skin solution proposed by Joran W. Booth et al. achieves a node root mean square error of 32.3 mm through embedded sensors, but the accuracy is limited due to the mechanical coupling deviation between the sensors and structural nodes, and the system root mean square error reaches 45.8 mm. The six-axis accelerometer solution proposed by J. Kimber et al. has a node root mean square error of more than 30 mm and a system root mean square error of more than 40 mm.

[0026] This invention achieves a root mean square error of 21.2 mm in node position reconstruction through joint optimization of 24 multimodal flexible sensors and geometric constraint equations, which is about 34% better than the scheme of Joran W. Booth et al. The overall structural shape reconstruction error is 39.4 mm, which is about 13.5% smaller than the external sensor scheme (Jonathan Bruce) and better than the error of >40 mm of the traditional physical model.

[0027] Local deformation analysis: Fifth-order polynomial fitting was used for bending and tension states respectively. The LSTM model (99% accuracy) effectively captures the nonlinear characteristics of the sensor (such as material hysteresis and compression / tension coupling effect), resulting in a surface height prediction error of 35.8 mm, which is better than existing soft sensor solutions (such as SpikeBot's SMA actuator).

[0028] 2. Real-time performance: The dynamic response speed meets the requirements of closed-loop control.

[0029] The recurrent neural network method proposed by Wen-Yung Li et al. requires offline data acquisition and processing, and cannot track deformation in real time; the accelerometer scheme proposed by J. Kimber et al. can only record vibration data, lacks shape reconstruction capability, and relies on complex mechanical parameters (such as spring stiffness and rod torque), which are actually difficult to measure.

[0030] In this invention, the online analysis response time of the LSTM model and polynomial curve fitting for time-series resistance data is measured in millimeters, and the combination of the two modes achieves millisecond-level closed-loop feedback. Furthermore, this invention uses an Arduino Mega to collect the resistance values ​​of 24 sensors in real time, and dynamically updates the node coordinates using geometric equations (27 constraints), meeting the rapid decision-making needs in complex terrain exploration.

[0031] 3. Autonomy: No external sensors are required, improving system robustness.

[0032] The UKF method developed by Ken Caluwaerts et al. relies on external base stations and markers, which limits the autonomous adaptability of the tension structure; the 9 kg heavy external sensing system of the "Super Ball Bot" developed by foreign research institutions is difficult to maintain stability in low gravity environments.

[0033] The present invention replaces the traditional tendon with 24 multimodal flexible sensors, and directly reconstructs the shape through the resistance-deformation mapping relationship, without the need for external components such as inertial measurement units and ultra-wideband devices;

[0034] Simultaneously, a data-driven strategy (LSTM + geometric equations) is adopted to avoid the dependence of traditional physical models on unmeasurable parameters such as spring stiffness and cable tension, significantly improving robustness in dynamic environments.

[0035] 4. Multimodal fusion: compatible with complex deformations and long-term stability

[0036] Single-sensor solutions (such as Jonathan Bruce's tensile measurement or J. Kimber's acceleration) cannot simultaneously capture bending and tensile coupled deformation; soft actuators such as Spikebot suffer from signal drift due to material fatigue, resulting in large fluctuations in ΔR / R.

[0037] In this invention, the sensor bending state is fitted with a fifth-order polynomial (to eliminate local nonlinear deformation error), and the stretching state is modeled with LSTM to compensate for the material hysteresis effect. The resistance stability test shows that the ΔR / R fluctuation is < ±0.23% (50 cycles). The diene rubber-based sensor is verified through 50 bending / stretching cycles, and the response time consistency (281-926 ms) is better than that of traditional conductive polymers (such as CNT-based sensors). Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the self-sensing system for tensioned integral structures based on multimodal strain sensing according to the present invention.

[0040] Figure 2 This is a schematic diagram of the fitting strategy of the present invention.

[0041] Figure 3 This is a schematic diagram of the reconstruction strategy of the present invention.

[0042] Figure 4This is a schematic diagram of the experimental results of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the invention.

[0044] like Figure 1 As shown, this invention discloses a self-sensing system and method for tensioned monolithic structures based on multimodal strain sensing. Its core lies in achieving real-time, high-precision reconstruction of the three-dimensional morphology of complex flexible structures through a deeply integrated sensor network and an innovative hybrid data-driven model, thereby completely eliminating reliance on external measurement equipment. This invention uses a classic icosahedral tensioned monolithic structure with six links and twenty-four tendons as a case study, aiming to address the core pain points of inaccurate models, low precision, and limited applications in existing technologies, providing a solid technical foundation for developing intelligent robots capable of autonomous operation in unknown environments.

[0045] To transform the physical deformation of a tensioned structure in reality into a calculable and realistic 3D digital model in a computer, four key steps are required: conversion of physical deformation to resistance signals, conversion of resistance signals to digital data, conversion of digital data to geometric information, and finally, reconstruction of the 3D model from geometric information. The hardware architecture of this invention seamlessly integrates sensing functionality with the structural body. Traditional passively tensioned tendons are replaced by twenty-four flexible sensors made of conductive diene rubber. These sensors serve as both key load-bearing components of the structure and sensing elements responding to mechanical deformation, successfully achieving the conversion of physical deformation to resistance signals. The data acquisition module is driven by two Arduino Mega microcontrollers, providing a stable 5V operating voltage to each sensor. A parallel voltage divider circuit with a known 5.8 MΩ precision resistor is used to monitor the voltage changes of each sensor in real time. Based on this, the microcontrollers calculate the instantaneous resistance values ​​of all twenty-four tendons, forming a real-time updated resistance vector. This vector is the original input data for all subsequent analysis and calculations, completing the transformation of the resistance signal into a real-time resistance time series digital signal.

[0046] The key methodological innovation of this invention lies in identifying and accurately modeling the multimodal response of flexible sensors under different deformation modes. Experiments demonstrate that under compression, regardless of the speed of sensor compression, as long as the same degree of compression is achieved, the resistance change is almost completely consistent. This means that the physical process is quasi-static, requiring no consideration of time or historical information. For this fixed, time-invariant functional relationship, multinomial regression is a classic and effective mathematical tool. It requires less computation than complex deep learning models, has a faster response speed, stronger interpretability, and does not require a large amount of training data or a complex training process. Subsequent validation further confirms its effectiveness. The value is 0.9999. Therefore, its accuracy is high enough.

[0047] like Figure 2 As shown, the specific fitting process is as follows: First, data is collected and a series of calibration experiments are conducted. A professional tensile testing machine is used to precisely bend the sensor at different speeds. During the experiment, two key data pairs are recorded simultaneously: compressive strain X and resistance change rate y. After collecting a large number of such data points, a polynomial model is selected.

[0048] By successively increasing the order, it was found that the fitting effect was better at order 5. Finally, the fitting and parameter solving were performed, and the final formula was obtained by using the least squares method and calculating its coefficient of determination. The final formula is as follows.

[0049]

[0050] Using the above formula, this invention can accurately determine the length of the sensor by the magnitude of the resistance, providing crucial data for the next step of reconstructing the specific structure of the sensor.

[0051] However, in tensile mode, the sensor's response is much more complex, exhibiting a significant dynamic hysteresis effect, meaning its resistance-strain relationship is highly dependent on the tensile rate. To address this challenge, such as... Figure 3 As shown, this invention introduces a Long Short-Term Memory (LSTM) network. LSTM is a special type of recurrent neural network whose core design is to process and learn long-term relationships in time-series data. Its internal "gating mechanism" and "cell state" enable it to "remember past information and use this historical information to assist in current decision-making (prediction)." This perfectly aligns with the event correlation in this invention. LSTM controls the flow of information through a series of gating mechanisms, including forget gates, input gates, cell state updates, and output gates.

[0052] Forget Gate: Determines which information to discard from the cell state. The forget gate checks h... t-1 (The hidden state of the previous moment) and X t(The input vector at the current time step) is passed through a sigmoid function and outputs a value between 0 and 1, representing each value corresponding to the old cell state c. t-1 How much information is retained? The formula is:

[0053]

[0054] in, It is the sigmoid function, W f It is the weight matrix, b f It is a bias term.

[0055] Input Gate: Responsible for updating the cell state. First, the input gate is also based on h. t-1 and X t A value is calculated, and then the new information to be updated is determined. This process involves two steps: first, using the sigmoid function to determine which parts will be updated; and second, using the tanh function to create new candidate values. These can be added to the state. The formula is as follows:

[0056]

[0057]

[0058] in, This indicates the bias of the input gate. This represents the weight matrix of the input gate. The weight matrix represents the state of the candidate input gate.

[0059] Cell state update: Update the cell state based on the results of the forget gate and input gate. First, update the old state C... t-1 Multiply by the input of the forget gate t Forget some information; then add the output of the input gate. They may be added to the state, as shown in the following formula:

[0060]

[0061] The output gate determines the next hidden state h. t It should be like this: the state contains a memory of the previous state and will be passed to the next LSTM unit and possibly subsequent layers. The output gate first utilizes h... t-1 and X t This produces an output, which is then passed through a tanh function and multiplied by the result of a sigmoid gate to output a partial new cell state. The formula is:

[0062]

[0063] in, It is the output of the output gate. It is the weight matrix of the output gate. It is the new hidden state at time t. It is the bias term of the output gate.

[0064] To improve model training efficiency and convergence stability, all input sequences X and target output Y are normalized, typically using min-max normalization to map them to the [0, 1] interval. Subsequently, the generated sample set {(X(t), Y(t))} is randomly divided into training, validation, and test sets. The backpropagation algorithm is used to optimize the weight matrix and bias terms of the LSTM model. Mean squared error (MSE) is used as the loss function during training, and the learning rate is adjusted using the Adam optimizer to accelerate convergence. The trained model achieves a 99% prediction accuracy on the test set. The high accuracy and robustness of the LSTM model are verified by comparing the error distribution between the model's predicted strain values ​​and the actual measured values.

[0065] This invention employs an LSTM model to solve the challenge of dynamic modeling under stretched conditions. Specifically, this is reflected in: Dynamic characteristic capture: The hysteresis effect, which traditional regression methods cannot handle, is accurately modeled by LSTM through learning long-term dependencies in the time series. Real-time performance guarantee: The LSTM model is optimized to achieve efficient computation while ensuring accuracy, and its response time meets the real-time shape recognition requirements in dynamic environments. Accuracy improvement and robustness: The hybrid model combined with traditional regression significantly improves the overall reconstruction accuracy, reducing the node RMSE to 21.2mm, and remaining stable within a large noise range (-0.23 to 0.13), making it suitable for complex and ever-changing real-world environments.

[0066] Based on this hybrid model, the complete 3D shape reconstruction algorithm can be executed in real time on the host computer. First, after receiving the resistance vector R, the algorithm calls the corresponding model for each sensor, converting its resistance value into a strain value, thereby obtaining the complete strain vector. Next, using the formula Based on the calculated strain vector ε and the pre-measured initial tendon length vector The system can accurately calculate the current instantaneous length vector L of all twenty-four tendons, thus completing the conversion from data information to geometric information. Finally, the spatial coordinates of all nodes are determined by solving a system of nonlinear equations composed of geometric constraints. This invention first uses the three nodes at the bottom... , , Establish a local coordinate system, and then use the remaining geometric constraints to locate the other nodes. These constraints include six constant strut length constraints (in the form of...). (where Lr is the strut length) and the remaining twenty-one dynamically changing tendon length constraints (in the form of...) (where Lm is the tendon length calculated in real time). This overdetermined or well-determined system of equations, consisting of 30 equations, is solved numerically using an iterative least squares method, ultimately outputting the precise three-dimensional coordinates of all nodes and completing the shape reconstruction.

[0067] like Figure 4 As shown, this invention successfully implemented a high-precision self-sensing system through deep integration of hardware and software and the application of advanced algorithms. Experimental results verified the superior performance of this scheme, with a root mean square error (RMSE) of only 21.2 mm for the localization of a single node and a reconstruction error of 39.4 mm for the overall system morphology. These achievements were all accomplished without any external sensor assistance, not only surpassing most existing technologies in accuracy but, more importantly, endowing the tensioning robot with the core capability to autonomously explore and operate in complex and unknown environments, demonstrating its broad application prospects.

[0068] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A self-sensing system for a tensioned integral structure based on multimodal strain sensing, characterized in that, The system includes: The sensing system design module is used to replace the tendons in the tensioned overall structure with conductive and flexible sensors to achieve dual-mode response in bending and stretching states. The polynomial curve fitting module is used to obtain strain values ​​based on the sensor resistance values ​​in a bending state through polynomial regression fitting. The LSTM network model module is used to train an LSTM network model based on the historical resistance vector sequence under tensile conditions and its corresponding strain data sequence. The trained model is used to predict the sensor strain value under tensile conditions. The 3D shape reconstruction module is used to calculate the corresponding sensor length based on the strain values ​​obtained from the polynomial curve fitting module and the LSTM network model module, thereby determining the spatial coordinates of the nodes in the overall tensioned structure and completing the shape reconstruction.

2. The self-sensing system for tensioned integral structures based on multimodal strain sensing according to claim 1, characterized in that, The sensor has a dual-mode response and can simultaneously sense the bending and stretching states of the tensioned overall structure, including a conductive flexible diene rubber material sensor, a liquid metal sensor, or a capacitive flexible sensor.

3. The self-sensing system for tensioned integral structures based on multimodal strain sensing according to claim 1, characterized in that, The instantaneous resistance values ​​of all sensors are obtained using the Arduino Mega data acquisition module, forming a set of real-time updated resistance vectors.

4. The self-sensing system for tensioned integral structures based on multimodal strain sensing according to claim 1, characterized in that, In the LSTM network model module, the current deformation is predicted by recursively calculating the node position at the previous time step, supporting autonomous monitoring in extreme environments.

5. The self-sensing system for tensioned integral structures based on multimodal strain sensing according to claim 1, characterized in that, The sensor length L is calculated as follows: in, The strain value of the sensor. For the pre-measured initial length vector of the tendon, This indicates the XOR operation.

6. The self-sensing system for tensioned integral structures based on multimodal strain sensing according to claim 1, characterized in that, In the 3D shape reconstruction module, the spatial coordinates of all nodes are determined by solving a system of nonlinear equations consisting of geometric constraints; the geometric constraints include constant strut length constraints and dynamically changing tendon length constraints.

7. The self-sensing system for tensioned integral structures based on multimodal strain sensing according to claim 6, characterized in that, Based on the obtained sensor strain values, the node coordinates are dynamically solved using geometric constraint equations and the least squares method, enabling real-time reconstruction of 3D shapes without the need for external sensors.

8. A self-sensing method for a tensioned integral structure based on multimodal strain sensing, using the system described in any one of claims 1-7, characterized in that, The method includes the following steps: (1) A conductive and flexible sensor is used to replace the tendons in the tensioned overall structure to achieve a dual-mode response in bending and stretching states; (2) Based on the sensor resistance value in the bending state, the strain value is obtained by polynomial regression fitting; (3) Train an LSTM network model based on the historical resistance vector sequence and its corresponding strain data sequence under tensile conditions, and predict the sensor strain value under tensile conditions using the trained model. (4) Calculate the corresponding sensor length based on the strain value obtained by the polynomial curve fitting module and the LSTM network model module, and then determine the spatial coordinates of the nodes in the tensioned overall structure to complete the shape reconstruction.

Images