A tactile sensor device for unmanned equipment based on a large model

Abstract

This invention discloses a contact sensor device for unmanned equipment based on a large model, applied in the fields of artificial intelligence and unmanned equipment technology. It includes: a sensor module for collecting physical signals when the unmanned equipment comes into contact with external objects; a data transmission module for transmitting the collected data to a large model processing module; a large model processing module for performing in-depth analysis and processing of the transmitted data based on the large model, and generating corresponding decision commands to be sent to a control module; wherein the large model is constructed based on the Transformer architecture of the GPT-NeoX model, and the model's input is obtained by feature extraction from the collected data using a combination of CNN and RNN; and a control module for converting the decision commands into control signals and sending them to the execution structure of the unmanned equipment for action control. This invention effectively improves the perception and intelligent decision-making capabilities of unmanned equipment in complex environments.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and unmanned equipment technology, and in particular to a contact sensor device for unmanned equipment based on a large model. Background Technology

[0002] The level of intelligence of unmanned equipment largely depends on its ability to perceive and interact with its surrounding environment, and tactile sensors, as key components for achieving this capability, play a crucial role. Tactile sensors enable unmanned equipment to perceive information such as the contact state with external objects, the magnitude and direction of contact force, etc., enabling unmanned equipment to make more accurate and intelligent decisions in complex environments.

[0003] In industrial manufacturing, collaborative robots rely on tactile sensors to work safely and efficiently with human operators. When a robot accidentally comes into contact with an operator during a task, the tactile sensors can quickly detect the contact and trigger appropriate safety mechanisms to prevent injury and damage to equipment, ensuring smooth production. In the logistics and warehousing industry, automated guided vehicles (AGVs) use tactile sensors for precise cargo handling and retrieval. By sensing the contact force and position information with goods, AGVs can stably grasp goods and accurately transport them in complex warehouse environments, improving logistics efficiency. In the medical field, surgical robots utilize tactile sensors to perform gentle manipulations of tissues, avoiding damage to surrounding healthy tissues and improving surgical precision and safety.

[0004] With the rapid development of artificial intelligence technology, large-scale models, as a significant breakthrough, are profoundly changing the development landscape of various fields. Large-scale models possess powerful learning and generalization capabilities, enabling them to learn from and analyze massive amounts of data, thereby extracting features and patterns. Combining large-scale models with tactile sensors in unmanned equipment presents unprecedented opportunities to enhance the intelligence level of these devices. Large-scale models can perform in-depth analysis and processing of the vast amounts of data collected by tactile sensors, thereby more accurately identifying the shape, material, hardness, and other attributes of objects, as well as determining the interaction state between the unmanned equipment and its surrounding environment. This allows unmanned equipment to better adapt to complex and changing environments and perform more complex and sophisticated tasks.

[0005] Unmanned robots performing search and rescue missions in unknown wilderness environments can move flexibly through complex terrain and obstacles by working in collaboration with tactile sensors and large models. The large model can adjust the robot's movement strategy in real time based on contact information collected by the sensors, guiding the robot to bypass obstacles, find the optimal path, quickly locate trapped individuals, and improve rescue efficiency. In smart home scenarios, intelligent robot assistants, aided by tactile sensors and large models, can understand user intentions and needs. When users interact with the robot, the tactile sensors detect the user's touch, gestures, and other information; the large model analyzes and understands this information to provide corresponding services, such as helping users find items and control home appliances, providing a more convenient and intelligent living experience.

[0006] Research on tactile sensor devices for unmanned equipment based on large-scale models can not only promote the development of unmanned equipment technology and improve its adaptability and task execution capabilities in complex environments, but also has broad application prospects and significant socio-economic value. In the civilian sector, it can promote the development of industries such as intelligent logistics, smart homes, and healthcare, improving people's quality of life and productivity. Therefore, conducting research in this field has important practical significance and far-reaching strategic importance.

[0007] However, despite the progress made in the application of tactile sensors and large-scale models in unmanned equipment, some shortcomings still exist. Existing tactile sensors still need further improvement in terms of sensitivity, resolution, and stability to meet the demands of unmanned equipment in more complex and high-precision tasks. Furthermore, the fusion of large-scale models and tactile sensors is still in its early stages, requiring in-depth research and improvement in data fusion methods, model training optimization, real-time performance, and reliability.

[0008] Therefore, how to provide a large-model-based contact sensor device for unmanned equipment that can effectively combine the knowledge and reasoning capabilities of large models with the perceptual data of contact sensors to achieve more efficient and intelligent decision-making is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention proposes a contact sensor device for unmanned equipment based on a large model.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A contact sensor device for unmanned equipment based on a large model includes: a sensor module, a data transmission module, a large model processing module, and a control module;

[0012] Sensor module: Used to collect physical signals when unmanned equipment comes into contact with external objects;

[0013] Data transmission module: Used to transmit data collected by the sensor module to the large model processing module;

[0014] Large Model Processing Module: This module is used to perform in-depth analysis and processing of the transmitted data based on a large model, and to generate corresponding decision commands to be sent to the control module. The large model is a GPT-NeoX model based on the Transformer architecture, and the input of the GPT-NeoX model is obtained by feature extraction of the collected data based on a combination of CNN and RNN.

[0015] Control module: Used to convert decision commands into control signals and send them to the execution structure of unmanned equipment for motion control of the unmanned equipment.

[0016] Optionally, the sensor module employs a tactile sensor based on quantum tunneling composite materials, and the tactile sensor has been optimized in terms of both structure and materials, specifically as follows:

[0017] In terms of structural design, micro-nano structure manufacturing technology is adopted to construct a sensor surface structure with a micro pyramid array. The effects of different micro pyramid heights, spacings, and cone angles on sensor sensitivity are compared through finite element simulation analysis to determine the optimal parameters of the micro-nano structure.

[0018] Nanoparticles and carbon nanotubes were introduced into the material, and the content and dispersion state of nanoparticles and carbon nanotubes were controlled. The optimal parameters were obtained through long-term stability tests.

[0019] Optionally, the sensor module may also include: signal conditioning circuitry and data acquisition circuitry;

[0020] The signal conditioning circuit is used to amplify, filter, and linearize the weak electrical signal output by the sensor, specifically as follows:

[0021] The signal output by the sensor is first amplified by a low-noise amplifier. Then, the amplified signal is filtered out by a bandpass filter to remove high-frequency noise and low-frequency interference signals. A linearization circuit is added to the signal conditioning circuit to fit and compensate the characteristic curve of the sensor, so that the output signal and the contact force present a good linear relationship.

[0022] The data acquisition circuit is used to convert the conditioned analog signal into a digital signal, specifically:

[0023] An analog-to-digital converter is used to sample the analog signal, and a sample-and-hold circuit is added to the input of the analog-to-digital converter to keep the signal stable at the moment of sampling.

[0024] Optionally, the data transmission module uses a high-speed Ethernet interface as the main communication method. The Ethernet controller is connected to the external network through an RJ45 interface, and an isolation transformer and lightning protection circuit are added to the Ethernet interface. At the same time, the data transmission module also supports multiple communication protocols such as TCP / IP and UDP.

[0025] Optionally, the GPT-NeoX model can be optimized, including: designing a dedicated data encoding module in the input layer of the GPT-NeoX model to convert the multidimensional data collected by the tactile sensor into a vector representation suitable for the model input, specifically:

[0026] A method combining location encoding and feature encoding is adopted to map data of each dimension to a specific vector space, and location encoding is used to distinguish data at different time points and spatial locations.

[0027] Optionally, the GPT-NeoX model can be optimized by: converting the weights and activations in the model from high-precision data types to low-precision data types through model quantization, and removing unimportant connections or neurons from the model through model pruning.

[0028] Optionally, before training the GPT-NeoX model, the following preprocessing is also included: data denoising based on median filtering and Kalman filtering, and data normalization based on min-max normalization and Z-score normalization.

[0029] Optionally, feature extraction can be performed on the collected data using a combination of CNN and RNN, specifically:

[0030] The spatial features extracted by the CNN are used as the input of the RNN. The RNN processes the time series data through a recurrent structure and outputs a feature vector containing spatiotemporal features. An attention mechanism is introduced between the CNN and the RNN to make the model focus on important features.

[0031] Optionally, the training process of the GPT-NeoX model also includes: introducing ensemble learning during the model inference process, fusing the inference results of multiple models with different initializations, and obtaining the final result through a weighted average.

[0032] Optionally, the control module is also used to monitor and manage the operating status of the entire device, including monitoring the working status of the sensor module, data transmission module and large model processing module, as well as fault diagnosis and alarm functions.

[0033] As can be seen from the above technical solution, compared with the prior art, this invention proposes a contact sensor device for unmanned equipment based on a large model. This invention achieves high-precision detection and stable signal output of sub-millinewton level contact force by employing a sensor module based on quantum tunnel composite material and optimized for structure and materials, combined with a design including signal conditioning and data acquisition circuitry. A high-speed Ethernet interface and a data transmission module with lightning protection and anti-interference design ensure reliable real-time data transmission. The GPT-NeoX model based on the Transformer architecture, optimized through input layer encoding and quantization pruning, combined with CNN and RNN fusion feature extraction and attention mechanisms, significantly improves the accuracy of recognizing object material, shape, and contact state (reaching over 90%), and enhances inference reliability through ensemble learning. The control module not only accurately converts decision commands into control signals but also monitors the status of each module in real time and provides fault diagnosis and alarms. Overall, compared with traditional sensors, this device has significant advantages in accuracy (error ≤2%), stability (drift ≤3%), response speed (average 6ms), and reliability (significantly reduced number of failures), effectively improving the perception and intelligent decision-making capabilities of unmanned equipment in complex environments. Attached Figure Description

[0034] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of the device structure of the present invention. Detailed Implementation

[0036] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0037] Example 1

[0038] Embodiment 1 of this invention discloses a contact sensor device for unmanned equipment based on a large model, such as... Figure 1 As shown, it includes: a sensor module, a data transmission module, a large model processing module, and a control module;

[0039] Sensor module: Used to collect physical signals when unmanned equipment comes into contact with external objects.

[0040] In sensor selection, it is crucial to fully consider the specific application scenarios and performance requirements of unmanned equipment. For precision assembly tasks of industrial robots, high-precision, high-sensitivity contact sensors are needed to ensure accurate perception of contact status and force magnitude even under minute forces. Therefore, after comparing and analyzing the performance of various sensors, the sensor module of this invention adopts a contact sensor based on quantum tunneling composite material. Quantum tunneling composite material is a novel smart material with unique resistance-pressure variation characteristics. When subjected to minute pressure, the conductive particles inside rearrange, causing a sharp drop in resistance, thereby achieving highly sensitive pressure detection. This sensor's sensitivity is several times higher than that of traditional conductive rubber sensors, capable of detecting sub-millinerton level contact force changes, meeting the stringent requirements of contact force sensing accuracy for precision assembly in industrial robots.

[0041] To further optimize the sensor's performance, the tactile sensor was optimized in terms of both structure and materials, specifically as follows:

[0042] In terms of structural design, micro-nano structure manufacturing technology is adopted to construct a sensor surface structure with a micro pyramid array. This micro-nano structure can increase the contact area between the sensor and the object being measured, improve the pressure transmission efficiency, and thus enhance the sensitivity of the sensor. The optimal parameters of the micro-nano structure are determined by comparing the effects of different micro pyramid heights, spacings, and cone angles on the sensor sensitivity through finite element simulation analysis.

[0043] By introducing silver nanoparticles and carbon nanotubes into the material, the silver nanoparticles, with their excellent conductivity and stability, can improve the conductivity of the composite material. The carbon nanotubes, with their superior mechanical and electrical properties, can enhance the mechanical strength and electron transport capability of the composite material. Furthermore, by controlling the content and dispersion state of the silver nanoparticles and carbon nanotubes and obtaining the optimal parameters through long-term stability testing, the optimized quantum tunnel composite material maintains high sensitivity while significantly improving stability and durability.

[0044] The sensor module also includes signal conditioning circuitry and data acquisition circuitry.

[0045] The signal conditioning circuit is used to amplify, filter, and linearize the weak electrical signal output by the sensor, specifically as follows:

[0046] The sensor output signal is first amplified by a low-noise amplifier (LNA) to increase the signal amplitude and reduce the impact of noise. The selected LNA has an extremely low noise figure and a high gain-bandwidth product, maintaining signal purity while amplifying the signal. The amplified signal then passes through a bandpass filter to remove high-frequency noise and low-frequency interference, ensuring the signal frequency range meets requirements. A second-order Butterworth bandpass filter is used, with cutoff frequencies set to 1Hz and 100Hz, effectively removing electromagnetic interference and low-frequency drift signals from the environment. To compensate for the sensor's nonlinear characteristics, a linearization circuit is added to the signal conditioning circuit to fit and compensate for the sensor's characteristic curve, resulting in a good linear relationship between the output signal and the contact force.

[0047] The data acquisition circuit is used to convert the conditioned analog signal into a digital signal, specifically:

[0048] A high-precision 16-bit analog-to-digital converter (ADC) is used to sample the analog signal at a sampling rate of up to 100kHz, which meets the requirements for acquiring rapidly changing tactile signals. A high-precision reference voltage source is used for the ADC to ensure conversion accuracy. To improve the stability and reliability of data acquisition, a sample-and-hold circuit is added to the input of the ADC to maintain signal stability at the moment of sampling, avoiding conversion errors caused by signal variations.

[0049] Data transmission module: Used to transmit the data collected by the sensor module to the large model processing module.

[0050] Considering the real-time and reliability requirements of data transmission, the data transmission module uses a high-speed Ethernet interface as the primary communication method. The Ethernet controller employs a high-performance and low-power chip and connects to the external network via an RJ45 interface. To ensure data transmission stability, an isolation transformer and lightning protection circuit are added to the Ethernet interface, effectively preventing damage to the equipment from external electromagnetic interference and lightning strikes. The data transmission module also supports multiple communication protocols, including TCP / IP and UDP, which can be selected and configured according to actual application needs to achieve seamless communication with different external devices. Furthermore, to ensure data transmission reliability, data verification and error correction technologies, such as CRC checksum and Hamming code error correction, are employed during data transmission to ensure the integrity and accuracy of the transmitted data. When the sensor module collects a large amount of data, the data transmission module can dynamically adjust the transmission rate and mode according to the data volume and real-time requirements to meet the real-time data demands of the large model processing module.

[0051] Large Model Processing Module: Used to perform in-depth analysis and processing of the transmitted data based on the large model, and generate corresponding decision instructions to be sent to the control module.

[0052] In selecting the large-scale model, after in-depth analysis and comparison of various mainstream models, and considering the characteristics and requirements of tactile sensor data processing for unmanned equipment, the GPT-NeoX model based on the Transformer architecture was ultimately chosen. The GPT-NeoX model possesses powerful language understanding and generation capabilities, performing excellently in natural language processing tasks. Furthermore, its open-source nature allows for customized development and optimization based on specific application scenarios. The GPT-NeoX model employs the multi-head self-attention mechanism in the Transformer architecture, effectively processing long-sequence data and capturing complex dependencies within it. When processing tactile sensor data, this data typically contains both temporal and spatial information, such as changes in contact force at different times and the distribution of contact points on the sensor surface. The GPT-NeoX model's multi-head self-attention mechanism can simultaneously focus on data from different time points and spatial locations, thereby better understanding the features and patterns in the tactile data. The model's multi-layered structure enables deep feature extraction, mining higher-level semantic information from the raw tactile signals, providing strong support for subsequent object attribute recognition and contact state determination.

[0053] To better adapt the GPT-NeoX model to tactile sensor data, a series of targeted optimizations and adjustments were made to the model, as follows:

[0054] A dedicated data encoding module is designed in the input layer of the GPT-NeoX model to transform the multidimensional data (such as contact force, contact position, contact time, etc.) collected by the tactile sensors into a vector representation suitable for the model input. Specifically:

[0055] A method combining location encoding and feature encoding is employed to map data in each dimension to a specific vector space, and location encoding is used to distinguish data from different time points and spatial locations. In this way, the model can accurately capture the spatiotemporal characteristics of the data when processing the input data.

[0056] To improve the model's inference efficiency and real-time performance, model quantization and pruning techniques were employed. Model quantization converts the weights and activation values ​​in the model from high-precision data types to low-precision data types; for example, converting 32-bit floating-point numbers to 8-bit integers. This reduces the model's storage space and computational load without significantly degrading its performance, while increasing inference speed. Furthermore, model pruning removes unimportant connections or neurons, simplifying the model structure and further reducing computational complexity. The combined use of model quantization and pruning techniques significantly improves the model's inference efficiency while maintaining accuracy, enabling it to meet the stringent real-time requirements of unmanned equipment.

[0057] Before training the GPT-NeoX model, the process includes preprocessing the collected data, including data denoising based on median filtering and Kalman filtering, and data normalization based on min-max normalization and Z-score normalization. Median filtering effectively removes isolated noise points by replacing the value of each point in the data sequence with the median of its neighborhood. Kalman filtering is an optimal estimation method based on a state-space model, which can accurately estimate the true value of the signal even in the presence of noise. After filtering, the data is normalized to a specific interval, such as [0,1] or [-1,1], to eliminate the dimensional differences between data of different dimensions and make the data comparable.

[0058] Feature extraction involves extracting key features from preprocessed data that reflect object attributes and contact states, providing input for subsequent model inference. This invention employs a combination of CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network) to extract features from the collected data, which are then used as input to the GPT-NeoX model. CNN effectively extracts spatial features from data, demonstrating excellent extraction capabilities for information such as contact position distribution in tactile sensor data. By designing convolutional kernels of different sizes and numbers, convolutional operations are performed on the data to extract spatial features at different levels. RNN excels at processing time-series data, capturing temporal dependencies and effectively modeling information such as changes in contact force over time. The spatial features extracted by CNN are used as input to the RNN, which processes the time-series data through a recurrent structure, outputting a feature vector containing spatiotemporal characteristics. To further improve the feature extraction effect, an attention mechanism is introduced between CNN and RNN, enabling the model to focus more on important features in the data and improve feature representativeness.

[0059] Model inference utilizes a trained large model to analyze and judge extracted features, deriving the object's attributes and contact state. The feature vectors extracted are input into the optimized GPT-NeoX model, which infers and predicts based on learned knowledge and patterns. During inference, the model generates a series of intermediate representations based on the input feature vectors. These intermediate representations contain preliminary judgments about the object's attributes and contact state. Further processing and analysis of these intermediate representations ultimately yield information about the object's material, shape, hardness, and other attributes, as well as contact state information such as stable contact, sliding, and collision. To improve the accuracy and reliability of inference, ensemble learning is employed during model inference, fusing the inference results from multiple differently initialized models and obtaining the final result through weighted averaging.

[0060] The output transforms the results obtained from model inference into an easily understandable and applicable form, providing them to the control system of the unmanned equipment or other application modules. Object attributes and contact state information are output in text or numerical form; for example, the object's material might be "metal," its shape "cylinder," its hardness "high," and its contact state "stable contact." To facilitate interaction with the unmanned equipment's control system, the results can also be output as specific control commands, such as adjusting the unmanned equipment's speed and force based on the contact state. During output, the reliability of the results is assessed by calculating indicators such as the confidence level of the model inference results. If the reliability of the results is low, the system automatically triggers a mechanism for re-detection or further analysis to ensure the accuracy and reliability of the output results.

[0061] Control module: Used to convert decision commands into control signals and send them to the execution structure of the unmanned equipment for motion control. For example, when the large model processing module determines that the unmanned equipment is unstable when grasping an object, the control module will immediately adjust the actions of the execution mechanism, such as adjusting the magnitude of the grasping force or changing the grasping position, to ensure stable object grasping.

[0062] The control module is also used to monitor and manage the operating status of the entire device, including monitoring the working status of the sensor module, data transmission module and large model processing module, as well as fault diagnosis and alarm functions.

[0063] The collaborative working mechanism between the functional modules is based on event-driven and task scheduling principles. When the sensor module detects a contact event, it triggers a data acquisition event, sending the acquired data to the large model processing module via the data transmission module. Upon receiving the data, the large model processing module immediately initiates a data processing task, analyzing and processing the data according to preset algorithms and models to generate decision commands. These decision commands are then sent to the actuators of the unmanned equipment via the control module, which perform the corresponding actions. Throughout the process, the control module monitors the working status of each module in real time to ensure stable system operation. If a module malfunctions, the control module will promptly issue an alarm signal and take appropriate fault-handling measures, such as switching to a backup module or performing fault recovery operations.

[0064] To thoroughly evaluate the performance advantages of the large-scale unmanned equipment tactile sensor device, this invention conducts a comprehensive experimental comparison with traditional tactile sensors.

[0065] In terms of accuracy, traditional tactile sensors suffer from measurement errors when measuring contact forces due to their inherent physical characteristics and signal processing limitations. For measuring minute contact forces, traditional sensors have low resolution and struggle to accurately detect minute force changes. When measuring contact forces less than 1N, the relative error of traditional sensors is typically around 5%-10%. However, tactile sensor devices based on large models, through learning and training on large amounts of tactile data and using advanced data processing algorithms, can measure contact forces more accurately. Under the same experimental conditions, for measuring contact forces less than 1N, the relative error of sensors based on large models can be controlled within 2%, demonstrating a significant improvement in accuracy. This is because large models can learn the complex nonlinear relationship between contact force and the sensor output signal, thereby enabling more precise calibration and compensation of the measured values.

[0066] In terms of response time, the signal processing flow of traditional tactile sensors is relatively simple, typically employing hardware circuits for signal amplification, filtering, and conversion. The response time mainly depends on the delay of the hardware circuitry. Generally, the response time of traditional sensors is between 5-10 ms. While large-model-based tactile sensor devices introduce large-model data processing, their response time is not significantly increased through coordinated hardware and software optimization and efficient algorithm implementation. In experiments, the average response time of large-model-based sensors is around 6 ms, comparable to traditional sensors. This is thanks to the use of a high-performance hardware computing platform and optimized algorithm architecture, enabling the large model to complete the analysis and processing of sensor data in a short time.

[0067] In stability comparisons, traditional tactile sensors are prone to measurement drift during prolonged use due to environmental factors (such as changes in temperature and humidity) and component aging. After 8 hours of continuous operation, the measurement drift of traditional sensors can reach 5%-10%. Tactile sensor devices based on large models, utilizing the adaptive learning capabilities of large models, can monitor and compensate for performance drift caused by environmental changes and component aging in real time. Under the same 8-hour continuous operation experimental conditions, the measurement drift of the large-model-based sensor can be controlled within 3%, demonstrating significantly better stability than traditional sensors. The large model can automatically adjust the data processing model based on changes in environmental parameters and historical sensor data, maintaining measurement accuracy.

[0068] In terms of reliability, traditional tactile sensors have relatively simple hardware structures, with failure modes mainly concentrated on damage to sensor components and circuit connection failures. During experiments, traditional sensors experienced an average of 2-3 failures per 100 hours of operation. While the large-model-based tactile sensor device has more complex hardware and software systems, it improves system reliability through redundant design, fault diagnosis and self-healing technologies, and the large model's ability to identify and process abnormal data. Within the same operating time, the average number of failures for the large-model-based sensor can be reduced to less than one. The large model can monitor and analyze the data collected by the sensor in real time. Once abnormal data is detected, it can promptly determine whether it is caused by a sensor failure and take corresponding measures, such as switching to a backup sensor or performing fault repair.

[0069] The high recognition accuracy is a significant advantage of large-model-based tactile sensor devices. Traditional tactile sensors typically only detect the presence or absence of contact and simple force magnitude, with very limited ability to identify object properties such as material and shape. In contrast, large-model-based sensor devices, through learning and training on a large amount of tactile data from objects of different materials and shapes, can accurately identify object properties and contact states. In experiments recognizing objects of various materials and shapes, traditional sensors almost failed to accurately identify the material and shape, while large-model-based sensor devices achieved an accuracy rate exceeding 90%. Large models can extract key features from complex tactile data and match and reason with learned knowledge, thereby achieving accurate judgment of object properties and contact states.

[0070] Large-model-based tactile sensor devices for unmanned equipment exhibit significant advantages over traditional tactile sensors in terms of accuracy, stability, and recognition rate. Although their response time is comparable to traditional sensors, through coordinated hardware and software optimization, the introduction of a large model has not led to a significant increase in response time. In terms of reliability, large-model-based sensor devices have also achieved good performance through the application of various technologies.

[0071] The various embodiments in this specification are described 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. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0072] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A contact sensor device for unmanned equipment based on a large model, characterized in that, include: Sensor module, data transmission module, large model processing module, and control module; The sensor module is used to collect physical signals when the unmanned equipment comes into contact with external objects. The data transmission module is used to transmit the data collected by the sensor module to the large model processing module. The large model processing module is used to perform in-depth analysis and processing on the transmitted data based on the large model, and generate corresponding decision instructions to be sent to the control module; wherein, the large model is constructed based on the GPT-NeoX model of the Transformer architecture, and the input of the GPT-NeoX model is obtained by feature extraction of the collected data based on a combination of CNN and RNN; The control module is used to convert the decision instructions into control signals and send them to the execution structure of the unmanned equipment to perform motion control of the unmanned equipment. The sensor module employs a contact sensor based on quantum tunneling composite materials, and the contact sensor has been optimized in terms of both structure and materials, specifically: In terms of structural design, micro-nano structure manufacturing technology is adopted to construct a sensor surface structure with a micro pyramid array. The effects of different micro pyramid heights, spacings, and cone angles on sensor sensitivity are compared through finite element simulation analysis to determine the optimal parameters of the micro-nano structure. Nanoparticles and carbon nanotubes are introduced into the material, and the content and dispersion state of the nanoparticles and carbon nanotubes are controlled. The optimal parameters are obtained through long-term stability tests. The sensor module further includes: a signal conditioning circuit and a data acquisition circuit; The signal conditioning circuit is used to amplify, filter, and linearize the weak electrical signal output by the sensor, specifically as follows: The signal output by the sensor is first amplified by a low-noise amplifier. Then, the amplified signal is filtered out by a bandpass filter to remove high-frequency noise and low-frequency interference signals. A linearization circuit is added to the signal conditioning circuit to fit and compensate the characteristic curve of the sensor, so that the output signal and the contact force present a good linear relationship. The data acquisition circuit is used to convert the conditioned analog signal into a digital signal, specifically: An analog-to-digital converter is used to sample the analog signal, and a sample-and-hold circuit is added to the input of the analog-to-digital converter to keep the signal stable at the moment of sampling. Optimizing the GPT-NeoX model includes: designing a data encoding module in the input layer of the GPT-NeoX model to convert the multidimensional data collected by the tactile sensor into a vector representation suitable for the model input, specifically: A method combining positional encoding and feature encoding is adopted to map the data of each dimension to a vector space, and to distinguish data at different time points and spatial locations through positional encoding; Feature extraction from the collected data is performed using a combination of CNN and RNN, specifically as follows: The spatial features extracted by the CNN are used as the input of the RNN. The RNN processes the time series data through a recurrent structure and outputs a feature vector containing spatiotemporal features. An attention mechanism is introduced between the CNN and the RNN to make the model focus on important features.

2. The contact sensor device for unmanned equipment based on a large model according to claim 1, characterized in that, The data transmission module uses a high-speed Ethernet interface as the communication method. The Ethernet controller is connected to the external network through an RJ45 interface, and an isolation transformer and lightning protection circuit are added to the Ethernet interface. The data transmission module also supports multiple communication protocols such as TCP / IP and UDP.

3. The contact sensor device for unmanned equipment based on a large model according to claim 1, characterized in that, Optimization of the GPT-NeoX model includes: converting the weights and activation values ​​in the model from high-precision data types to low-precision data types through model quantization, and removing unimportant connections or neurons from the model through model pruning.

4. The contact sensor device for unmanned equipment based on a large model according to claim 1, characterized in that, Before training the GPT-NeoX model, the process also includes: preprocessing the collected data, including: data denoising based on median filtering and Kalman filtering, and data normalization based on min-max normalization and Z-score normalization.

5. The contact sensor device for unmanned equipment based on a large model according to claim 1, characterized in that, The training process of the GPT-NeoX model also includes: introducing ensemble learning during the model inference process, fusing the inference results of multiple models with different initializations, and obtaining the final result through a weighted average.

6. The contact sensor device for unmanned equipment based on a large model according to claim 1, characterized in that, The control module is also used to monitor and manage the operating status of the entire device, including monitoring the working status of the sensor module, data transmission module and large model processing module, as well as fault diagnosis and alarm functions.

Images