A method and device for calibrating a flexible tactile sensor

Abstract

The application discloses a kind of flexible tactile sensor's calibration method and device, the method includes: obtaining the response data of flexible tactile sensor array, feature extraction and data preprocessing are carried out to response data;Based on standardized training data, calibration model is generated by convolutional neural network and long short-term memory network training;Collect the environmental parameters where sensor is located, and carry out self-adapting adjustment using Bayesian optimization algorithm;Sensor is transported to multiple-station parallel calibration unit and is calibrated;The performance of calibrated sensor is tested and generates the unique identification code of quality traceability.The application solves the technical problems such as strong environmental dependence, long calibration time and low efficiency in the existing flexible tactile sensor calibration process, realizes the fast, high-precision, adaptive environment flexible tactile sensor calibration, shortens calibration time from several hours to within 60 seconds, and the calibration error is controlled within ±2%, to meet the large-scale production demand.

Description

Technical Field

[0001] This invention relates to the field of flexible electronics technology, and more specifically to a calibration method and apparatus for a flexible tactile sensor. Background Technology

[0002] Flexible tactile sensors, as core components in emerging fields such as wearable devices, soft robots, and human-computer interaction systems, are characterized by their lightweight, flexibility, and high sensitivity, and have broad application prospects in fields such as medical health, industrial automation, and consumer electronics. Currently, flexible tactile sensors are mainly designed based on different working principles such as resistive, capacitive, and piezoelectric, and are typically composed of multi-layered composite structures consisting of conductive materials, substrate materials, and functional materials.

[0003] Traditional flexible tactile sensor manufacturing technologies mainly rely on methods such as screen printing, inkjet printing, and micro / nano fabrication to prepare sensing elements. For example, screen printing technology transfers conductive ink onto a flexible substrate through a screen to form conductive patterns; while micro / nano fabrication technology uses processes such as photolithography and etching to construct microstructures on flexible substrates to achieve specific tactile sensing functions.

[0004] Existing mass production technologies for flexible tactile sensor arrays primarily employ roll-to-roll continuous printing manufacturing processes, combined with automated assembly and packaging systems, and include sensor performance testing and calibration at the end of the production line. This technology uses a fully automated production line to print conductive patterns onto a flexible substrate, then stacks and assembles functional materials, and finally performs cutting, packaging, and performance testing, forming a complete production process.

[0005] However, existing technologies have significant shortcomings in sensor calibration, mainly manifested in the following ways: the calibration process requires a strictly constant temperature and humidity environment, placing high demands on production conditions; the calibration time is long, severely restricting production efficiency; calibration parameters depend on fixed environmental conditions, resulting in poor performance consistency of sensors in real-world, variable environments; and the calibration equipment is bulky, fixed in design, lacking flexibility and making it difficult to adapt to the needs of large-scale mass production. These problems directly affect the production capacity, yield, and performance stability of flexible tactile sensors, limiting their large-scale application. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a calibration method and apparatus for flexible tactile sensors, enabling fast, high-precision, and environmentally adaptive calibration of flexible tactile sensors, thus meeting the needs of large-scale mass production.

[0007] The technical solution of this invention is:

[0008] A calibration method for a flexible tactile sensor, comprising:

[0009] The response data of the flexible tactile sensor array is acquired, and the response data is subjected to feature extraction and data preprocessing to obtain standardized training data.

[0010] Based on the standardized training data, sensor spatial features are extracted through a convolutional neural network, and temporal response features are obtained through long short-term memory network processing. The obtained sensor spatial features and temporal response features are used as inputs to train and generate a calibration model.

[0011] The environmental parameters of the flexible tactile sensor are collected, and the calibration model is adaptively adjusted based on the environmental parameters using a Bayesian optimization algorithm to obtain optimized calibration parameters.

[0012] The flexible tactile sensor is sent to a multi-station parallel calibration unit, and the optimized calibration parameters are used to perform parallel calibration on the flexible tactile sensor to generate calibration data and obtain the calibrated flexible tactile sensor.

[0013] The calibrated flexible tactile sensor is subjected to performance testing. Based on the calibration data, production parameters and test results are recorded, and a unique identification code for quality traceability is generated.

[0014] Preferably, the step of performing feature extraction and data preprocessing on the response data to obtain standardized training data includes:

[0015] Based on the response data, wavelet transform is used to extract signal features to obtain feature vectors;

[0016] Principal component analysis is performed on the feature vectors to reduce dimensionality and eliminate data redundancy, resulting in dimensionality-reduced feature data.

[0017] The reduced-dimensionality feature data is then standardized to obtain the standardized training data.

[0018] Preferably, the step of extracting sensor spatial features through a convolutional neural network and processing them through a long short-term memory network to obtain temporal response features includes:

[0019] The standardized training data is processed using multiple convolutional layers and pooling layers to obtain spatial feature data;

[0020] The spatial feature data is input into a long short-term memory network unit to extract temporal features, thereby obtaining temporal feature data;

[0021] The spatial feature data and the temporal feature data are fused to obtain the spatiotemporal features of the sensor.

[0022] Preferably, the training to generate the calibration model includes:

[0023] The standardized training data is divided into a training set and a validation set according to a preset ratio;

[0024] The training set is iteratively trained using an optimization algorithm until a preset training accuracy threshold is reached, thereby obtaining an initial calibration model.

[0025] The initial calibration model is evaluated using the validation set. When the evaluation metrics meet preset conditions, the calibration model is obtained.

[0026] Preferably, the acquisition of environmental parameters of the flexible tactile sensor includes:

[0027] Multiple sensor nodes are set up within the calibration area, and environmental data is collected in real time using wireless transmission.

[0028] Each of the sensing nodes is configured with multiple types of sensors, which collect temperature, humidity and air pressure data to obtain raw environmental data;

[0029] The original environmental data is filtered to eliminate noise interference, thereby obtaining the environmental parameters.

[0030] Preferably, the step of adaptively adjusting the calibration model using a Bayesian optimization algorithm based on the environmental parameters includes:

[0031] Based on the environmental parameters, an optimization space for the calibration parameters is generated;

[0032] The optimal parameter combination is searched in the optimization space using the Bayesian optimization algorithm to obtain the optimized parameters;

[0033] The calibration model is adjusted according to the optimization parameters, and the optimized calibration parameters are obtained after filtering.

[0034] Preferably, the step of conveying the flexible tactile sensor to the multi-station parallel calibration unit includes:

[0035] The flexible tactile sensor is positioned using a visual positioning system to obtain position information.

[0036] Based on the location information, the flexible tactile sensor is transmitted to the calibration station by a robotic arm; wherein, a closed-loop control strategy is adopted during the transmission process, and the position of the flexible tactile sensor is adjusted by real-time monitoring of the sensor position and comparing it with the preset calibration position, so that the positional accuracy meets the preset range.

[0037] The flexible tactile sensor is fixed to the calibration station using a clamp.

[0038] Preferably, the parallel calibration of the flexible tactile sensor using the optimized calibration parameters includes:

[0039] The optimized calibration parameters are set as the external stimuli applied to the flexible tactile sensor to excite it with pressure, temperature and humidity, so as to simulate its working state in different real environments.

[0040] The response signals of multiple flexible tactile sensors are simultaneously acquired using a data acquisition module to obtain the acquired response signals;

[0041] The collected response signals are processed and features are extracted to obtain the calibration data.

[0042] Preferably, the performance test of the calibrated flexible tactile sensor includes:

[0043] Based on preset performance evaluation indicators, the performance parameters of the calibrated flexible tactile sensor are measured to obtain parameter test results;

[0044] The response characteristics of the calibrated flexible tactile sensor were tested under various environmental conditions to obtain response test results.

[0045] Based on the parameter test results and the response test results, and according to the preset performance grading standard, the performance level of the calibrated flexible tactile sensor is determined.

[0046] The present invention also provides a calibration device for a flexible tactile sensor, comprising:

[0047] The data acquisition module is used to acquire response data from the flexible tactile sensor array, perform feature extraction and data preprocessing on the response data, and obtain standardized training data.

[0048] The calibration model generation module is used to extract sensor spatial features through a convolutional neural network based on the standardized training data, obtain temporal response features through a long short-term memory network, and use the obtained sensor spatial features and temporal response features as inputs to train and generate a calibration model.

[0049] An adaptive optimization module is used to collect environmental parameters of the flexible tactile sensor and use a Bayesian optimization algorithm to adaptively adjust the calibration model according to the environmental parameters to obtain optimized calibration parameters.

[0050] The calibration execution module is used to send the flexible tactile sensor to the multi-station parallel calibration unit, perform parallel calibration of the flexible tactile sensor using the optimized calibration parameters, generate calibration data, and obtain the calibrated flexible tactile sensor.

[0051] The quality management module is used to perform performance tests on the calibrated flexible tactile sensor, record production parameters and test results based on the calibration data, and generate a unique identification code for quality traceability.

[0052] The beneficial effects of this invention are:

[0053] 1. This invention uses a deep learning model that combines convolutional neural networks and long short-term memory networks to build an intelligent calibration system, which realizes accurate mapping between sensor signals and physical quantities, shortens the traditional calibration time from several hours to within 60 seconds, and controls the calibration error within ±2%.

[0054] 2. This invention relates to a portable, multifunctional integrated calibration device designed using MEMS technology, with a size of only 150×150×50mm. 3 It integrates a miniature excitation source, high-precision data acquisition and wireless communication system, which greatly improves the flexibility and efficiency of the calibration process.

[0055] 3. The adaptive environment simulation and compensation technology of the present invention monitors environmental parameters in real time and dynamically adjusts calibration conditions. Combined with Bayesian optimization algorithm, it automatically generates the optimal calibration path, enabling the sensor to maintain stable performance in the temperature range of -10℃ to 60℃ and the humidity range of 20%-95%RH, which greatly improves the environmental adaptability of the sensor.

[0056] 4. The multi-station parallel calibration unit design of this invention integrates 12-16 independent calibration stations, realizing high-throughput calibration of sensor arrays, with a calibration capacity of 960 pieces / hour per unit time, meeting the production capacity requirement of 3 million pieces / year, and significantly improving production efficiency.

[0057] 5. This invention employs a quality traceability system based on blockchain technology, assigning a unique ID to each sensor and recording production parameters, calibration data, and test results. This enables precise traceability and analysis of quality issues, thereby improving product yield. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of a calibration method for a flexible tactile sensor according to the present invention;

[0060] Figure 2 This is a schematic diagram of the deep learning-based calibration model architecture of the present invention;

[0061] Figure 3 This is a schematic diagram of the calibration device for a flexible tactile sensor according to the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0063] like Figure 1 As shown, the present invention provides a calibration method for a flexible tactile sensor, comprising the following steps:

[0064] Step S1: Obtain the response data of the flexible tactile sensor array, perform feature extraction and data preprocessing on the response data to obtain standardized training data;

[0065] The flexible tactile sensor array was tested under various environmental conditions, collecting a large amount of response data. These environmental conditions included temperatures ranging from 0-60℃, humidity from 20%-90% RH, and different usage states such as static pressure, dynamic pressure, and shear force. For each sensor unit, 100-1000 data points were collected, including input excitation values, output signal values, and corresponding environmental parameters. To ensure data quality and representativeness, Latin hypercube sampling was used for data acquisition to guarantee the uniformity of data distribution under different environmental conditions and usage states.

[0066] After acquiring the response data, effective feature extraction and data preprocessing are required. First, wavelet transform is used to perform multi-scale analysis on the original signal, extracting signal features containing sensor response characteristics to form feature vectors. These features can effectively characterize the sensor's performance parameters such as sensitivity, linearity, hysteresis, drift, and cross-sensitivity. Next, principal component analysis (PCA) is performed on the extracted feature vectors to reduce dimensionality, eliminating redundant information and reducing subsequent computation while retaining the main feature information. After dimensionality reduction, Z-score normalization is used to normalize the feature data, making data of different dimensions comparable. Outlier detection methods are then used to identify and handle abnormal data points, further improving data quality. Finally, standardized training data is obtained, providing a high-quality data foundation for subsequent deep learning model training.

[0067] like Figure 2 As shown, in step S2, based on the standardized training data, the sensor spatial features are extracted through a convolutional neural network, and the temporal response features are obtained through a long short-term memory network. The obtained sensor spatial features and temporal response features are used as inputs to train and generate a calibration model.

[0068] A hybrid deep learning architecture was designed, combining the advantages of Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs) to simultaneously process the spatial and temporal features of sensors. First, standardized training data is input into a CNN structure consisting of 3-5 convolutional and pooling layers. The convolutional layers extract local spatial features of the sensor array by performing convolution operations with the input data using kernels; the pooling layers reduce the dimensionality of the feature maps through downsampling, improving computational efficiency and enhancing the model's robustness. After multiple convolutional and pooling operations, a feature map characterizing the spatial distribution of the sensors is obtained, referred to as spatial feature data. Next, the extracted spatial feature data is input into an LSTM network structure. The LSTM network consists of 2-3 LSTM units, each containing three control units: an input gate, a forget gate, and an output gate, effectively capturing the temporal dependence of the sensor response. The LSTM network learns the sensor's response characteristics at different time points through a memory mechanism, analyzes the temporal evolution of the sensor signal, captures the dynamic response characteristics of the sensor, and finally outputs temporal feature data.

[0069] In the calibration system of this invention, the specific structural design of the Convolutional Neural Network (CNN) fully considers the extraction requirements of spatial features from the sensor array. The CNN structure comprises a deep network consisting of three main convolutional layers and two pooling layers. The first convolutional layer is configured with 32 3×3 convolutional kernels, using a stride of 1 and padding of 1 to maintain the feature map size, and uses the ReLU activation function to introduce non-linear characteristics; the subsequent max pooling layer uses a 2×2 pooling window and a stride of 2 to halve the feature map size while retaining salient features. The second convolutional layer contains 64 3×3 convolutional kernels, also using a stride of 1 and padding of 1 and the ReLU activation function to extract higher-level feature representations; the following second pooling layer uses the same 2×2 max pooling strategy. The third convolutional layer further increases to 128 3×3 convolutional kernels to capture more complex spatial feature patterns, and uses a global average pooling layer to convert the feature map into a fixed-length feature vector. This progressively deepening convolutional structure design can effectively extract spatial distribution features at different scales from sensor array data, providing a reliable foundation for subsequent time series analysis.

[0070] The Long Short-Term Memory (LSTM) network is specifically optimized for the temporal response characteristics of sensors. This structure comprises two bidirectional LSTM layers, each with 128 hidden units, forming a deep recurrent network architecture. The LSTM units use the tanh activation function to handle unit state updates and the sigmoid function to control the on / off states of three gated units (input gate, forget gate, and output gate), effectively modeling long-term dependencies. The bidirectional design of the network allows it to simultaneously consider historical and future state information of the sensor response, improving the accuracy of temporal pattern recognition. The temporal processing length is set to 64 time steps; this parameter has been validated through extensive experiments, effectively balancing computational complexity and the ability to capture the dynamic response characteristics of the sensor. In the feature fusion stage, an attention mechanism is used to weightedly fuse the spatial features extracted by the CNN and the temporal features extracted by the LSTM, enabling the model to adaptively focus on the importance of different features, further improving calibration accuracy.

[0071] The model training data was acquired through meticulously designed data acquisition experiments, including comprehensive response datasets collected from different types and batches of flexible tactile sensors under strictly controlled experimental environments and various environmental conditions (temperature range 0-60℃, humidity range 20%-90%RH). A Latin hypercube sampling method was employed to ensure uniform sampling within the environmental parameter space, with a total of no fewer than 10,000 sampling points covering the sensor characteristics under various operating states and environmental conditions. Each sampling point records multi-dimensional information including the applied precise force value, temperature and humidity conditions, the sensor's raw output signal, and reference measurements, forming a high-quality training data foundation. In the data preprocessing stage, outlier detection, missing value handling, and noise filtering were performed on the raw data. Then, Z-score normalization was used to normalize the data, improving training efficiency and model generalization ability.

[0072] The model training process employed finely tuned hyperparameter settings to achieve optimal training results. The batch size was set to 64, striking a good balance between training efficiency and gradient estimation accuracy. The initial learning rate was 0.001, dynamically adjusted using a cosine annealing strategy. A high learning rate was maintained for rapid convergence in the early stages of training, while the learning rate was reduced for fine-tuning in the later stages. To prevent overfitting, multiple regularization strategies were introduced, including an L2 regularization coefficient set to 0.0001 and a Dropout layer with a ratio of 0.5. The total number of training iterations was 5,000, with an early stopping mechanism that automatically terminated training when validation set performance no longer improved. The entire training process was completed on a high-performance computing cluster equipped with Tesla V100 GPUs, with a typical training time of approximately 8 hours. After training, a comprehensive evaluation was performed using a test dataset to ensure that the model's calibration performance met the ±2% accuracy requirement under various conditions.

[0073] To fully utilize the spatiotemporal characteristics of the sensor, this invention employs feature fusion technology to fuse spatial feature data extracted by CNN and temporal feature data extracted by LSTM, generating comprehensive spatiotemporal features of the sensor. The fusion methods include feature concatenation and weighted fusion. Feature concatenation directly concatenates two types of features to form a higher-dimensional feature vector, while weighted fusion assigns different weights to the two types of features based on their importance, obtaining the fused feature through weighted summation.

[0074] When training the calibration model, the standardized training data is first divided into training and validation sets in an 80%:20% ratio. The hybrid deep learning model is trained using the training set, with the Adam optimizer used to update the model parameters. The mean squared error (MSE) loss function is chosen to measure the difference between the model's predictions and the true values. During training, techniques such as dynamically adjusting the learning rate, introducing batch normalization, and dropout are used to prevent overfitting and improve the model's generalization ability. The training iterations are no less than 5000 times until the loss function value decreases below a preset threshold, resulting in the initial calibration model.

[0075] The initial calibration model was then evaluated using a validation set, with evaluation metrics including mean squared error of prediction (MSE), relative standard deviation (RSD), and computation time. The model was considered to meet performance requirements when the MSE was less than 0.5%, the RSD was less than 1%, and the computation time for a single sensor was less than 50ms. If the performance did not meet the requirements, the model was optimized by adjusting the network structure, optimizing hyperparameters (such as learning rate, batch size, and the number of hidden layer neurons), and introducing regularization methods (such as L1 and L2 regularization) until the performance requirements were met, resulting in the final calibration model. To facilitate deployment in embedded systems, model pruning and quantization techniques were employed to reduce model size, thereby lowering computational complexity and storage requirements while maintaining model performance.

[0076] Finally, the trained deep learning model is combined with traditional temperature compensation algorithms and environmental correction methods to construct a comprehensive calibration model. A reinforcement learning mechanism is introduced, enabling the model to automatically adjust calibration parameters according to changes in environmental conditions, achieving adaptive calibration under a wide range of environmental conditions (such as a temperature range of -10℃ to 60℃ and a humidity range of 20% to 95% RH). This comprehensive calibration model can control the sensor's measurement error under different environmental conditions to within ±2%, significantly exceeding the ±5% accuracy level of traditional calibration methods.

[0077] Step S3: Collect the environmental parameters of the flexible tactile sensor, and use the Bayesian optimization algorithm to adaptively adjust the calibration model according to the environmental parameters to obtain the optimized calibration parameters;

[0078] A comprehensive environmental monitoring system needs to be designed to collect environmental parameters in real time from the location of the flexible tactile sensor. This system will consist of multiple sensor nodes within a calibration area, forming a distributed environmental monitoring network. Each sensor node will be equipped with a high-precision temperature sensor (accuracy ±0.1℃), a humidity sensor (accuracy ±1%RH), a barometric pressure sensor (accuracy ±10Pa), and an electromagnetic field detector (frequency range 10Hz-1GHz). These sensors will collect environmental data at a sampling frequency of 10Hz and transmit the data to the central controller in real time via wireless transmission. The wireless transmission will utilize low-power wide-area network (LPWAN) technologies (such as LoRa or NB-IoT) to ensure transmission stability and low latency.

[0079] After acquiring the raw environmental data, preprocessing is required to eliminate noise interference. First, a Kalman filter is used to filter the temperature, humidity, and air pressure data, eliminating random fluctuations and measurement noise. Then, median filtering and wavelet denoising are used to further purify the data, ensuring the accuracy of the environmental parameters. In addition, the system detects trends in environmental parameters and predicts short-term changes in environmental conditions, providing forward-looking guidance for the calibration process.

[0080] Based on the acquired environmental parameters, this invention employs a Bayesian optimization algorithm to adaptively adjust the calibration model. First, an optimization space for the calibration parameters is generated according to the current environmental parameters. This optimization space includes multiple dimensions, such as temperature compensation coefficient, humidity correction factor, pressure excitation range, and loading rate. Then, the Bayesian optimization algorithm searches for the optimal parameter combination within this multi-dimensional parameter space. Compared to traditional grid search and random search, Bayesian optimization, by constructing a probabilistic model (usually a Gaussian process) of the parameter space, can find the global optimum more efficiently, significantly reducing the number of trials during the search process.

[0081] The Bayesian optimization process mainly includes the following steps: First, initialize the Gaussian process model based on historical calibration data; then, determine the next parameter point to be evaluated using a data acquisition function (such as the expected improvement function EI or the upper confidence bound UCB); next, conduct actual calibration experiments using this set of parameters and record the results; finally, update the Gaussian process model and repeat the above process until the preset number of iterations is reached or the convergence condition is met. In this way, the algorithm can quickly find a near-optimal parameter combination in a finite number of trials.

[0082] After obtaining the optimized parameters, these parameters undergo further processing, including parameter smoothing to avoid abrupt changes, parameter range checking to ensure they remain within reasonable limits, and parameter consistency checks to avoid conflicts. After these processes, the final optimized calibration parameters are obtained, which will guide subsequent calibration processes. In this way, the calibration model can automatically adjust parameters according to changes in environmental conditions, achieving environmentally adaptive calibration and ensuring that the sensor maintains high-precision measurement capabilities under various environmental conditions.

[0083] Step S4: The flexible tactile sensor is sent to the multi-station parallel calibration unit, and the flexible tactile sensor is calibrated in parallel using the optimized calibration parameters to generate calibration data and obtain the calibrated flexible tactile sensor.

[0084] The flexible tactile sensor is transported to the multi-station parallel calibration unit via an automated feeding and conveying system. This process begins with a high-precision vision positioning system identifying the sensor's position. Equipped with a high-resolution industrial camera (resolution better than 5 micrometers), the vision system uses image processing algorithms to identify the sensor's edges, positioning marks, and electrode areas, accurately acquiring the sensor's spatial position and orientation information. The position identification accuracy reaches ±0.05mm, ensuring precise gripping and placement in subsequent operations.

[0085] Based on the acquired position information, a precision robotic arm picks up the sensor from the conveyor belt and transfers it to the appropriate calibration station. The transfer process employs a closed-loop control strategy, using encoder feedback and a vision-assisted system to monitor the sensor's position in real time, compare it with the preset calibration position, and dynamically adjust the robotic arm's trajectory to ensure the sensor is accurately placed at the calibration station. To prevent mechanical damage to the flexible sensor during transfer, the robotic arm uses a controllable pneumatic adsorption method to grasp the sensor. The grasping force can be intelligently adjusted according to the sensor's size and weight, controlled within the range of 0.1-1.0N.

[0086] After the sensor arrives at the calibration station, it is secured using a specialized fixture. The fixture design takes into account the characteristics of flexible sensors, employing an edge-gripping method to avoid applying excessive pressure to the sensor's sensitive areas. Simultaneously, the fixture is equipped with a fine-tuning mechanism, allowing for precise adjustments to the sensor's position within a range of ±0.1mm, ensuring accurate alignment between the sensor and the calibration device. The fixing process is completed collaboratively by a robotic arm and the fixture, taking no more than 3 seconds, significantly improving production efficiency.

[0087] In the multi-station parallel calibration unit, each station is equipped with a complete calibration system, including an excitation source module, a data acquisition module, and an environmental control module. Based on the optimized calibration parameters obtained in step S3, the system automatically sets various operating parameters. First, the excitation source module applies specific pressure, temperature, and humidity conditions to the sensor according to the parameter settings to simulate its working state in a real-world environment. Pressure excitation is applied through a precision piezoelectric actuator or a pressure head driven by a micro-motor, ranging from 0.1N to 50N with an accuracy of ±0.1N; temperature control is achieved through a micro Peltier element, ranging from 0 to 100℃ with an accuracy of ±0.2℃; humidity adjustment is achieved through a micro ultrasonic atomizer and a solid desiccant switching system, ranging from 20% to 95%RH with an accuracy of ±1.5%RH.

[0088] While applying excitation, the data acquisition module simultaneously acquires the sensor's response signal. The acquisition system employs a 24-bit high-precision analog-to-digital converter with a sampling rate of 100kHz and a signal-to-noise ratio better than 100dB, accurately capturing minute changes in the sensor's response. The system supports simultaneous acquisition across multiple channels, enabling simultaneous processing of multiple sensing units in the array. The acquired data is transmitted to the central processing unit via a high-speed data bus for real-time processing and analysis.

[0089] The acquired response signals are processed and their features extracted, including signal filtering, time-domain feature extraction, frequency-domain analysis, and statistical feature calculation. The processed data is compared with preset standard values ​​to calculate compensation coefficients and correction parameters, generating a calibration data package. The entire calibration process is coordinated by a central control system, with multiple workstations operating in parallel, significantly improving calibration efficiency. For a sensor array with 64 sensing units, the calibration time at a single workstation is controlled within 60 seconds. In a system equipped with 16 calibration workstations, the theoretical calibration capacity can reach 960 units / hour, meeting the needs of large-scale production.

[0090] After calibration, the sensor is sent to the next process, and the calibration data is saved in the central database and associated with the unique ID of the corresponding sensor to provide data support for subsequent quality traceability.

[0091] Step S5: Perform performance testing on the calibrated flexible tactile sensor, record production parameters and test results based on the calibration data, and generate a unique identification code for quality traceability.

[0092] After calibration, the flexible tactile sensor undergoes comprehensive performance testing to verify the effectiveness of the calibration and ensure product quality. Performance testing is divided into three levels: basic electrical parameter testing, functional performance testing, and environmental adaptability testing. First, basic electrical parameter testing mainly detects the sensor's resistance, capacitance, leakage current, and other fundamental electrical parameters to ensure that its electrical characteristics meet design specifications. The test uses a high-precision LCR meter and ammeter, achieving a measurement accuracy of 0.1%, providing a fundamental guarantee for sensor performance.

[0093] Next, the functional performance testing focuses on evaluating the sensor's core performance indicators, including sensitivity, linearity, hysteresis, response time, and repeatability. This testing uses standardized testing equipment to apply standard excitation to the sensor under controlled conditions and record its response characteristics. Sensitivity testing evaluates the sensor's responsiveness to external stimuli, requiring a sensitivity deviation of no more than ±5%; linearity testing evaluates the sensor's linear response characteristics within its measurement range, requiring a nonlinearity error of less than 3%; hysteresis testing assesses the sensor's hysteresis effect through load-unload cycles, requiring a hysteresis error of less than 2%; response time testing evaluates the sensor's response speed to rapidly changing stimuli, requiring a response time of less than 10 ms; and repeatability testing evaluates the sensor's stability through multiple measurements under identical conditions, requiring a repeatability error of less than 1%.

[0094] To verify the sensor's performance in real-world applications, the system also undergoes environmental adaptability testing. This test evaluates the sensor's response characteristics under various environmental conditions (such as different temperatures, humidity levels, and air pressures) to verify the calibrated environmental adaptability. The test environment covers a temperature range of -10°C to 60°C and a humidity range of 20% to 95% RH, ensuring the sensor maintains reliable performance in various real-world applications. These tests provide a comprehensive assessment of the calibration's effectiveness and the sensor's performance level.

[0095] Based on the test results, the sensors are graded according to their performance. According to pre-defined performance grading standards, sensors are divided into different levels, such as high-precision, standard, and economical, to meet the needs of different application scenarios. Performance grading information is recorded in the product database and linked to physical products via QR codes or RFID tags for easy customer identification and use.

[0096] Quality traceability is a crucial step in ensuring product reliability and rapid problem localization. This system establishes a full lifecycle quality traceability mechanism based on blockchain technology. The system assigns a globally unique identifier (UUID) to each sensor and records all information related to that sensor on the blockchain, including raw material batches, production process parameters, calibration data, test results, and quality grades. This data is encrypted and tamper-proof, ensuring its authenticity and reliability. By scanning the QR code on the sensor or reading the embedded RFID tag, users or managers can access the sensor's complete historical record, achieving end-to-end traceability.

[0097] Furthermore, statistical analysis is performed on this data to identify key factors and potential risks affecting product quality, providing data support for production process optimization and quality improvement. Through continuous data collection and analysis, combined with Statistical Process Control (SPC) methods, the system can achieve real-time monitoring and early warning of the production process, preventing batch quality problems. This closed-loop quality management mechanism has effectively improved product yield, gradually increasing it from 85% initially to over 95%, exceeding the industry average of 90%.

[0098] This invention's production management platform adopts a modern software architecture design, constructing a fully functional flexible tactile sensor production management system. The platform is based on a B / S (Browser / Server) architecture, with the backend developed using the Spring Boot framework, providing high-performance REST API services. The frontend uses Vue.js to implement a responsive user interface, supporting cross-platform and multi-terminal access. The data storage layer employs a hybrid architecture combining a MySQL relational database and a MongoDB non-relational database, respectively used to process structured production parameters and unstructured sensor characteristic curves, images, and other data. The system adopts a microservice architecture, decoupling functional modules into independent services. Efficient communication and load balancing are achieved through message queues (RabbitMQ) and a service gateway (Spring Cloud Gateway), ensuring stable operation in high-concurrency production environments. The platform is deployed in a containerized environment (Docker), supporting elastic scaling and rapid deployment to meet the needs of production lines of different sizes.

[0099] The production management platform comprises four core functional modules: The equipment monitoring module collects real-time data on the operating status, temperature and humidity parameters, and energy consumption of calibrated equipment, achieving millisecond-level status updates via WebSocket technology and automatically issuing warnings when equipment parameters deviate from normal ranges; the production scheduling module enables production planning, work order breakdown, and resource allocation, employing heuristic algorithms to optimize production scheduling and maximize equipment utilization, while displaying real-time production progress through electronic dashboards; the quality control module integrates test data from automated testing equipment, recording the performance parameters and pass / fail status of each sensor, generating detailed quality reports and statistical analyses; and the data analysis module is the system's intelligent core, using various data mining algorithms to discover patterns and anomalies hidden in production data.

[0100] The data analysis module employs various advanced algorithms, including random forests, XGBoost, and deep neural networks, to comprehensively analyze production process data. The algorithms can identify key factors affecting product quality, such as ambient temperature fluctuations during calibration, the application rate of excitation force, and batch-to-batch material variations. These factors are then ranked according to their degree of influence using methods such as feature importance analysis, correlation heatmaps, and principal component analysis. The system utilizes time series forecasting methods to analyze the changing trends of process parameters over time, predict potential quality fluctuations, and recommend optimal parameter adjustment strategies. For example, analysis revealed that sensor consistency is best when the calibration temperature is within the range of 28±0.5℃, and hysteresis is minimized when the excitation force loading rate is maintained at 0.2 N / s. These findings directly guide the optimization and adjustment of process parameters, resulting in more stable and consistent product performance.

[0101] The yield improvement mechanism is a crucial component of the production management platform, employing a closed-loop quality control strategy to continuously improve production processes. Based on the PDCA (Plan-Do-Check-Act) cycle model, this mechanism first uses Pareto analysis to identify the main defect types affecting yield, such as uneven sensitivity, linearity deviation, and zero-point drift. Then, it uses Fault Tree Analysis (FTA) and Failure Mode and Effects Analysis (FMEA) to deeply analyze the root causes, such as improper material ratios, insufficient environmental control precision, or unreasonable calibration parameter settings. Next, targeted improvement measures are developed and implemented for verification. Finally, effective measures are standardized and incorporated into production specifications. The system also incorporates Statistical Process Control (SPC) to monitor key process parameters such as calibration temperature, applied force accuracy, and signal acquisition noise in real time. Control charts are used to analyze parameter trends, automatically triggering warnings when parameters approach control limits or abnormal patterns occur. Through this comprehensive yield improvement mechanism, the yield of sensor products has gradually increased from 85% at the beginning of the project to over 95%, significantly exceeding the industry average of 90%, while simultaneously improving product performance consistency and reliability.

[0102] Specifically, in step S1, the step of extracting features and preprocessing data from the response data to obtain standardized training data includes: extracting signal features based on the response data using wavelet transform to obtain feature vectors; performing principal component analysis on the feature vectors to reduce dimensionality and eliminate data redundancy, thereby obtaining dimensionality-reduced feature data; and performing standardization on the dimensionality-reduced feature data to obtain the standardized training data.

[0103] In a preferred embodiment, a large amount of sensor response data is collected from the flexible tactile sensor array under various environmental conditions (temperature 0-60°C, humidity 20%-90%RH) and different usage states (static pressure, dynamic pressure, shear force, etc.). Each sensor unit collects 100-1000 sets of data points, including input excitation values, output signal values, and environmental parameters, forming an initial dataset. Latin hypercube sampling is used during data acquisition to ensure uniform data distribution.

[0104] Then, the collected sensor data is subjected to characteristic analysis, including parameters such as sensitivity, linearity, hysteresis, drift, and cross sensitivity. Wavelet transform and Fourier analysis are used to extract sensor signal features, PCA and t-SNE are employed to reduce data redundancy, and Z-score normalization and outlier detection methods are used to clean and preprocess the data to improve data quality.

[0105] In step S2, the step of extracting sensor spatial features through a convolutional neural network and processing them through a long short-term memory network to obtain temporal response features includes: processing the standardized training data using multiple convolutional layers and pooling layers to obtain spatial feature data; inputting the spatial feature data into a long short-term memory network unit to extract temporal features to obtain temporal feature data; and fusing the spatial feature data and the temporal feature data to obtain the spatiotemporal features of the sensor.

[0106] The training to generate the calibration model includes: dividing the standardized training data into a training set and a validation set according to a preset ratio; iteratively training the training set using an optimization algorithm until a preset training accuracy threshold is reached to obtain an initial calibration model; and using the validation set to evaluate the performance of the initial calibration model. When the evaluation index meets preset conditions, the calibration model is obtained.

[0107] Specifically, during the training process of the calibration model, the standardized training data needs to be rationally divided. This invention employs a stratified random sampling method, dividing the data into training and validation sets according to a preset ratio (typically 80%:20%). Stratified sampling ensures that the distribution of various data types in the training and validation sets remains consistent with the original dataset, avoiding model bias that may result from imbalanced sample distribution. In practice, this ratio can be further adjusted depending on the sensor type and data volume. For example, a 75%:25% ratio can be used for larger datasets, while an 85%:15% ratio can be used for smaller datasets to ensure that the training set has sufficient samples to support the model in learning complex features.

[0108] Iterative training of the training set using optimization algorithms is the core step in model building. This invention chooses the Adam optimizer as the main optimization algorithm because it combines the advantages of momentum and RMSProp, adaptively adjusting the learning rate of each parameter, making it particularly suitable for handling non-stationary targets and data with a large amount of noise. Batch training is used during training, with a batch size of 64-128 samples, achieving a good balance between computational efficiency and optimization stability. To prevent overfitting, various regularization techniques are introduced during training, including L2 regularization (weight decay coefficient set to 0.0001), dropout (dropout rate of 0.3-0.5), and early stopping. Furthermore, a dynamic adjustment strategy is adopted for the learning rate, initially set to 0.001. When the loss on the validation set no longer decreases after 5 consecutive epochs, the learning rate is reduced to 50% of its original value, accelerating model convergence. Iterative training continues until the model's mean absolute error (MAE) on the training set drops below a preset threshold (usually 1%), or the maximum number of iterations (5000) is reached, at which point the initial calibrated model is obtained.

[0109] After the initial calibration model training is completed, its performance needs to be comprehensively evaluated using a validation set. The evaluation employs a multi-metric comprehensive evaluation system, including multiple dimensions such as mean squared error (MSE), mean absolute error (MAE), relative standard deviation (RSD), maximum error (Max Error), and computation time. MSE and MAE evaluate the overall predictive accuracy of the model, RSD evaluates the stability of the prediction results, maximum error evaluates performance under extreme conditions, and computation time evaluates the model's practicality. Specifically, the requirements are: MSE < 0.5%, MAE < 0.8%, RSD < 1%, maximum error < 2%, and computation time for a single sensor < 50ms. Only when all evaluation metrics simultaneously meet the preset conditions is the model considered to have passed the performance test, at which point the final calibration model is obtained.

[0110] If the evaluation results fail to meet the preset conditions, model optimization and adjustment are required. Adjustment strategies include: network structure adjustment (e.g., adding or removing network layers, adjusting the number of neurons), hyperparameter optimization (e.g., adjusting the learning rate, batch size, and regularization coefficient), feature engineering improvements (e.g., adding feature combinations, introducing new features), and ensemble learning methods (e.g., model averaging, model fusion). The optimized model is then trained and evaluated again, and this process is iterated until a calibrated model that meets the performance requirements is obtained. After the model is finally determined, its robustness and generalization ability are further verified through k-fold cross-validation (k=5) to ensure stable performance under different data partitioning methods.

[0111] To enable the calibration model to be effectively deployed in real-world production environments, particularly embedded in portable calibration devices, this invention also optimizes the final model's deployment. First, model pruning removes neural connections that contribute little to the prediction results, reducing the number of model parameters. Second, weight quantization converts 32-bit floating-point numbers to 8-bit integer representations, significantly reducing the model's storage requirements. Finally, model compilation and hardware acceleration optimization fully utilize the target platform's computing resources, improving inference speed. After these optimizations, the model size is reduced by approximately 70%, inference speed is increased by approximately 3 times, and accuracy loss is controlled within 0.2%, fully meeting the needs of practical applications. The resulting calibration model not only possesses high accuracy and efficiency but also exhibits excellent deployment adaptability, laying a solid foundation for subsequent adaptive calibration work.

[0112] In a preferred embodiment, a hybrid deep learning architecture comprising a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network is designed. The CNN part, consisting of 3-5 convolutional and pooling layers, is used to extract spatial features from the sensor array; the LSTM part, consisting of 2-3 LSTM layers, is used to capture the temporal response characteristics of the sensor. The model input consists of the raw sensor signals and environmental parameters, and the output consists of calibrated pressure values ​​and temperature compensation coefficients. The Adam optimizer and mean squared error loss function are used, and the model is trained on a GPU cluster for at least 5000 iterations.

[0113] The model performance was then evaluated using cross-validation, with the validation dataset containing 20% ​​sensor samples that were not used in training. Evaluation metrics included mean squared error of prediction (MSE < 0.5%), relative standard deviation (RSD < 1%), and computation time (< 50 ms / piece). Based on the validation results, the model's generalization ability was improved by adjusting the network structure, optimizing hyperparameters, and introducing regularization methods. Model pruning and quantization techniques were also employed to reduce the model size, making it suitable for embedded system deployment.

[0114] A comprehensive calibration model is constructed by combining a trained deep learning model with traditional temperature compensation algorithms and environmental correction methods. A reinforcement learning mechanism is introduced to enable the model to automatically adjust calibration parameters according to environmental changes, achieving adaptive calibration within a temperature range of -10℃ to 60℃ and a humidity range of 20% to 95% RH. The final model can control the sensor's measurement error under different environmental conditions to within ±2%, significantly exceeding the ±5% accuracy level of traditional calibration methods.

[0115] In step S3, the acquisition of environmental parameters of the flexible tactile sensor includes: setting up multiple sensing nodes in the calibration area and acquiring environmental data in real time using wireless transmission; configuring multiple types of sensors for each sensing node, and acquiring temperature, humidity and air pressure data to obtain raw environmental data; filtering the raw environmental data to eliminate noise interference and obtain the environmental parameters.

[0116] The step of adaptively adjusting the calibration model using the Bayesian optimization algorithm based on the environmental parameters includes: generating an optimization space for calibration parameters based on the environmental parameters; searching for the optimal parameter combination in the optimization space using the Bayesian optimization algorithm to obtain optimized parameters; adjusting the parameters of the calibration model according to the optimized parameters; and obtaining the optimized calibration parameters after filtering.

[0117] In a preferred embodiment, an integrated adaptive optimization module is designed, comprising a high-precision temperature sensor (±0.1℃), a humidity sensor (±1%RH), a barometric pressure sensor (±10Pa), and an electromagnetic field detector (frequency range 10Hz-1GHz). A distributed arrangement is adopted, placing multiple sensor nodes within a calibration area to form a three-dimensional distributed monitoring network for environmental parameters. Monitoring data is transmitted wirelessly to the central controller in real time, with a sampling frequency of 10Hz, providing timely feedback for environmental parameter control.

[0118] Based on MEMS technology, an integrated micro-excitation source was designed. Pressure excitation employs a multi-stage pressure application mechanism combining a piezoelectric actuator and a micromotor, with a force range of 0.1N-50N and a resolution of 0.05N. Temperature excitation utilizes a micro Peltier element and a thin-film resistance heater, with a temperature control range of 0-100℃ and a heating / cooling rate of up to 5℃ / s. Humidity control employs a micro ultrasonic atomizer and a solid desiccant switching system, with a humidity regulation range of 20%-95%RH. The entire excitation source module has a volume controlled within 150×150×50mm. 3 Within the range.

[0119] A precise environmental control system was developed. Temperature control employs a combination of thermoelectric coolers (TEC) and microfluidic cooling channels to achieve precise temperature control within ±0.2℃ of the -10℃ to 60℃ range. Humidity control combines ultrasonic humidification and molecular sieve adsorption drying technologies to achieve precise humidity regulation within ±1.5%RH of the 20%-95%RH range. Pressure control utilizes a miniature air pump and a precision pressure reducing valve to simulate a pressure environment of 70kPa-110kPa. The control system employs a PID algorithm combined with fuzzy logic, achieving a response time of less than 10 seconds and a settling time of less than 30 seconds.

[0120] An intelligent calibration strategy generation system was developed to automatically generate the optimal calibration path and parameter settings based on real-time monitored environmental parameters and sensor characteristics. A Bayesian optimization algorithm is employed to select the most suitable parameter combination from a pre-defined calibration parameter space, including excitation force magnitude, loading rate, and temperature gradient. The system can dynamically adjust subsequent calibration steps based on initial test results, reducing redundant test points and optimizing the standard 50-point calibration process to 15-20 key points, significantly improving calibration efficiency while maintaining accuracy.

[0121] Based on Bayesian networks and data fusion technology, calibration data acquired under different environmental conditions are comprehensively analyzed. An environmental dependence model of the sensor response is constructed, quantifying the influence coefficients of environmental factors such as temperature and humidity on the sensor output. Through multi-source data fusion algorithms, a sensor response surface covering the entire environmental range is generated, ensuring the calibration results remain effective under a wider range of environmental conditions. Finally, a calibration data package containing compensation coefficients and correction formulas is generated, controlling the measurement accuracy variation of the sensor within ±3% in a temperature range of -10℃ to 60℃ and a humidity range of 20% to 95% RH.

[0122] In step S4, the step of transporting the flexible tactile sensor to the multi-station parallel calibration unit includes: using a visual positioning system to identify the position of the flexible tactile sensor and obtain position information; based on the position information, using a robotic arm to transfer the flexible tactile sensor to the calibration station; wherein, a closed-loop control strategy is adopted during the transfer process, and the position of the flexible tactile sensor is adjusted by real-time monitoring of the sensor position and comparing it with a preset calibration position to ensure that the position accuracy meets the preset range; and a fixture is used to fix the flexible tactile sensor on the calibration station.

[0123] The parallel calibration of the flexible tactile sensor using the optimized calibration parameters includes: setting pressure, temperature, and humidity excitation conditions as external stimuli applied to the flexible tactile sensor according to the optimized calibration parameters to simulate its working state under different actual environments; using a data acquisition module to simultaneously acquire response signals from multiple flexible tactile sensors to obtain the acquired response signals; and processing and extracting features from the acquired response signals to obtain the calibration data.

[0124] In a preferred embodiment, a high-precision automated sensor array loading system is developed, employing visual positioning and pneumatic gripping technology to precisely place batch-produced sensors onto calibration station trays. The conveyor system utilizes a dust-free conveyor belt and a precision positioning mechanism, achieving a positioning accuracy of ±0.1mm to prevent positional shifts and mechanical damage during transport. The system supports various sensor array sizes (5×5mm). 2 Up to 100×100mm 2Automatic identification and adaptation processing of ).

[0125] The system features a multi-station parallel calibration unit. Each unit comprises 12-16 independent calibration stations, each equipped with a complete excitation, measurement, and environmental control system. An automated robotic loading and unloading mechanism ensures precise docking between sensors and calibration equipment. The calibration process is centrally scheduled by the control system for optimal resource allocation. The complete calibration time for a single sensor array (64 sensing units) is controlled within 60 seconds. With 16 stations operating in parallel, the system's calibration capacity reaches 960 samples per hour.

[0126] In step S5, the performance test of the calibrated flexible tactile sensor includes: measuring the performance parameters of the calibrated flexible tactile sensor based on preset performance evaluation indicators to obtain parameter test results; testing the response characteristics of the calibrated flexible tactile sensor under various environmental conditions to obtain response test results; and determining the performance level of the calibrated flexible tactile sensor based on the parameter test results and the response test results, according to preset performance grading standards.

[0127] In a preferred embodiment, a multi-level quality inspection mechanism is introduced, including three levels: optical appearance inspection, electrical parameter testing, and functional performance testing. High-resolution industrial cameras (5-micron resolution) and automated testing equipment are used to inspect the physical defects and electrical performance of the sensors, and pass / fail standards are set for key indicators such as sensitivity consistency, linearity, and hysteresis. A data traceability system based on blockchain technology is established, assigning a unique ID to each sensor and recording raw material information, production parameters, calibration data, and test results to achieve full lifecycle quality traceability.

[0128] Develop a production management platform based on Industrial Internet of Things (IIoT) technology to achieve integrated management of equipment status monitoring, production planning and scheduling, material management, and quality control. Introduce big data analytics to mine and analyze production process data, identify key factors affecting product quality and capacity, and provide decision support for process optimization. The system provides a web interface and mobile application, supporting remote monitoring and management, enabling managers to grasp production status in real time and respond accordingly.

[0129] Based on big data analytics, a closed-loop mechanism for yield improvement was established. By analyzing the distribution of defective product causes, key quality control points were identified, and production process parameters and calibration procedures were optimized in a targeted manner. Statistical Process Control (SPC) methods were introduced to monitor the fluctuation trends of key process parameters, enabling early warning and intervention. Through continuous improvement and optimization, the yield of sensor products was gradually increased from the initial 85% to over 95%.

[0130] like Figure 3As shown, the present invention also provides a calibration device for a flexible tactile sensor, comprising:

[0131] The data acquisition module 101 is used to acquire response data from the flexible tactile sensor array, perform feature extraction and data preprocessing on the response data, and obtain standardized training data.

[0132] The calibration model generation module 102 is used to extract sensor spatial features through a convolutional neural network based on the standardized training data, obtain temporal response features through a long short-term memory network, and use the obtained sensor spatial features and temporal response features as input to train and generate a calibration model.

[0133] The adaptive optimization module 103 is used to collect the environmental parameters of the flexible tactile sensor and use the Bayesian optimization algorithm to adaptively adjust the calibration model according to the environmental parameters to obtain the optimized calibration parameters.

[0134] The calibration execution module 104 is used to send the flexible tactile sensor to the multi-station parallel calibration unit, perform parallel calibration of the flexible tactile sensor using the optimized calibration parameters, generate calibration data, and obtain the calibrated flexible tactile sensor.

[0135] The quality management module 105 is used to perform performance tests on the calibrated flexible tactile sensor, record production parameters and test results based on the calibration data, and generate a unique identification code for quality traceability.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A calibration method for a flexible tactile sensor, characterized in that, include: The response data of the flexible tactile sensor array is acquired, and the response data is subjected to feature extraction and data preprocessing to obtain standardized training data. Based on the standardized training data, spatial features of the sensor are extracted through a convolutional neural network, and temporal response features are obtained through long short-term memory network processing. The obtained sensor spatial features and temporal response features are used as inputs to train and generate a calibration model. The calibration model, after training, has the ability to output calibration parameters based on environmental parameters, including temperature compensation coefficient, humidity correction factor, pressure excitation range, and loading rate. The process involves collecting environmental parameters of the flexible tactile sensor, and then using a Bayesian optimization algorithm to adaptively adjust the calibration model based on these parameters to obtain optimized calibration parameters. This includes: generating an optimization space for the calibration parameters based on the environmental parameters; searching for the optimal parameter combination within the optimization space using the Bayesian optimization algorithm to obtain optimized parameters; adjusting the calibration model's parameters based on the optimized parameters; and outputting the optimized calibration parameters through a filtered processing based on the mapping relationship between the environmental parameters and the calibration parameters formed during the model's training. The flexible tactile sensor is fed to a multi-station parallel calibration unit, where it is calibrated in parallel using the optimized calibration parameters to generate calibration data and obtain the calibrated flexible tactile sensor. The calibration process includes: setting pressure, temperature, and humidity excitation conditions as external stimuli applied to the flexible tactile sensor according to the optimized calibration parameters to simulate its working state under different real-world environments; simultaneously acquiring response signals from multiple flexible tactile sensors using a data acquisition module to obtain the acquired response signals; and processing and extracting features from the acquired response signals to obtain the calibration data. The calibrated flexible tactile sensor is subjected to performance testing. Based on the calibration data, production parameters and test results are recorded, and a unique identification code for quality traceability is generated.

2. The method according to claim 1, characterized in that, The step of performing feature extraction and data preprocessing on the response data to obtain standardized training data includes: Based on the response data, wavelet transform is used to extract signal features to obtain feature vectors; Principal component analysis is performed on the feature vectors to reduce dimensionality and eliminate data redundancy, resulting in dimensionality-reduced feature data. The reduced-dimensionality feature data is then standardized to obtain the standardized training data.

3. The method according to claim 1, characterized in that, The step of extracting sensor spatial features through a convolutional neural network and processing them through a long short-term memory network to obtain temporal response features includes: The standardized training data is processed using multiple convolutional layers and pooling layers to obtain spatial feature data; The spatial feature data is input into a long short-term memory network unit to extract temporal features, thereby obtaining temporal feature data; The spatial feature data and the temporal feature data are fused to obtain the spatiotemporal features of the sensor.

4. The method according to claim 1, characterized in that, The training generates a calibration model, including: The standardized training data is divided into a training set and a validation set according to a preset ratio; The training set is iteratively trained using an optimization algorithm until a preset training accuracy threshold is reached, thereby obtaining an initial calibration model. The initial calibration model is evaluated using the validation set. When the evaluation metrics meet preset conditions, the calibration model is obtained.

5. The method according to claim 1, characterized in that, The collection of environmental parameters of the flexible tactile sensor includes: Multiple sensor nodes are set up within the calibration area, and environmental data is collected in real time using wireless transmission. Each of the sensing nodes is configured with multiple types of sensors, which collect temperature, humidity and air pressure data to obtain raw environmental data; The original environmental data is filtered to eliminate noise interference, thereby obtaining the environmental parameters.

6. The method according to claim 1, characterized in that, The step of conveying the flexible tactile sensor to the multi-station parallel calibration unit includes: The flexible tactile sensor is positioned using a visual positioning system to obtain position information. Based on the location information, the flexible tactile sensor is transmitted to the calibration station by a robotic arm; wherein, a closed-loop control strategy is adopted during the transmission process, and the position of the flexible tactile sensor is adjusted by real-time monitoring of the sensor position and comparing it with the preset calibration position, so that the positional accuracy meets the preset range. The flexible tactile sensor is fixed to the calibration station using a clamp.

7. The method according to claim 1, characterized in that, The performance test of the calibrated flexible tactile sensor includes: Based on preset performance evaluation indicators, the performance parameters of the calibrated flexible tactile sensor are measured to obtain parameter test results; The response characteristics of the calibrated flexible tactile sensor were tested under various environmental conditions to obtain response test results. Based on the parameter test results and the response test results, and according to the preset performance grading standard, the performance level of the calibrated flexible tactile sensor is determined.

8. A calibration device for a flexible tactile sensor, characterized in that, include: The data acquisition module is used to acquire response data from the flexible tactile sensor array, perform feature extraction and data preprocessing on the response data, and obtain standardized training data. The calibration model generation module is used to extract sensor spatial features through a convolutional neural network based on the standardized training data, and obtain temporal response features through a long short-term memory network. The obtained sensor spatial features and temporal response features are used as inputs to train and generate a calibration model. The calibration model, after training, has the ability to output calibration parameters based on environmental parameters, including temperature compensation coefficient, humidity correction factor, pressure excitation range, and loading rate. An adaptive optimization module is used to collect environmental parameters of the flexible tactile sensor and adaptively adjust the calibration model based on the environmental parameters using a Bayesian optimization algorithm to obtain optimized calibration parameters. This includes: generating an optimization space for calibration parameters based on the environmental parameters; searching for the optimal parameter combination in the optimization space using a Bayesian optimization algorithm to obtain optimized parameters; adjusting the parameters of the calibration model based on the optimized parameters; and obtaining the optimized calibration parameters after filtering. The calibration execution module is used to transmit the flexible tactile sensor to a multi-station parallel calibration unit, perform parallel calibration of the flexible tactile sensor using the optimized calibration parameters, generate calibration data, and obtain the calibrated flexible tactile sensor. The process includes: setting pressure, temperature, and humidity excitation conditions as external stimuli applied to the flexible tactile sensor according to the optimized calibration parameters to simulate its working state under different actual environments; simultaneously acquiring response signals from multiple flexible tactile sensors using a data acquisition module to obtain the acquired response signals; and processing and extracting features from the acquired response signals to obtain the calibration data. The quality management module is used to perform performance tests on the calibrated flexible tactile sensor, record production parameters and test results based on the calibration data, and generate a unique identification code for quality traceability.

Images