Method for improving real-time performance of touch sensor

Abstract

The invention relates to a method for improving real-time performance of a tactile sensor. The method comprises the following steps: acquiring a vibration signal of the tactile sensor, predicting vibration signal distribution by using Gaussian process regression (GPR), predicting a potential range of an effective sensing point in a sampling stage, obtaining a two-dimensional Gaussian probability distribution function, and determining a central point of the two-dimensional Gaussian probability distribution function and variances in x and y directions; setting the number M of randomly sampled points according to the resolution of the tactile sensor, randomly generating a plurality of groups of M two-dimensional points by using a two-dimensional Gaussian probability density function, and calculating the distance from each two-dimensional point in each group to a central point as the radius of a circular ring; each two-dimensional point corresponds to a circular ring radius, and the rings are arranged in sequence to form a differential pressure value matrix; and quantizing the matrix through a DPCM algorithm to obtain a differential pressure value matrix, and then performing Huffman coding compression to realize real-time transmission of the tactile sensor data. The real-time performance can be improved on the premise of ensuring the sampling accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of tactile sensor data transmission, and particularly to a method for improving the real-time performance of tactile sensors. Background Art

[0002] To ensure safe physical human-robot interaction, robots must be capable of perceiving forces. Compared with single-point force sensors, large-area tactile sensors can provide a more comprehensive perception of interaction information, laying a foundation for complex tasks. However, due to information transmission delays, the development and application of tactile sensors have been restricted.

[0003] Currently, research work mainly focuses on improving the resolution and accuracy of tactile sensors, and many methods have been developed and effectively applied in these fields. However, with the improvement of resolution and accuracy, the amount of tactile information also increases, bringing a huge burden to the transmission process. The information transmission process from the sampling stage to the receiving end includes five stages: sampling, compression, transmission, restoration, and presentation. The duration of each stage is directly affected by the amount of data, and the existing framework lacks corresponding optimization for tactile information, resulting in poor real-time performance. Among them, the Differential Pulse Code Modulation (DPCM) algorithm has been widely used in improving the real-time performance of various signals. It is a lossless compression algorithm. However, as the amount of signals increases, the compression ratio of the DPCM algorithm gradually fails to meet the requirements. Considering that the time cost of the transmission and restoration stages is determined by the hardware, the present invention aims to improve the real-time performance of the sampling, compression, and presentation stages. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art and in order to objectively identify effective tactile perception points showing pressure changes and reduce the amount of transmitted data, the present invention proposes a method for improving the real-time performance of tactile sensors. The method uses the Gaussian Process Regression-Differential Pulse Code Modulation (GPR-DPCM) algorithm for sampling, which can improve the real-time performance on the premise of ensuring sampling accuracy.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] A method for improving the real-time performance of tactile sensors, the method comprising the following:

[0007] Obtain the vibration signal of the tactile sensor. First, use Gaussian process regression (GPR) to predict the distribution of the vibration signal, predict the potential range of the effective sensing points (points with pressure changes) in the sampling stage, obtain the two-dimensional Gaussian probability distribution function, and thus determine its center point and the variances in the x and y directions.

[0008] Set the number M of randomly sampled points according to the resolution of the tactile sensor. Randomly generate multiple groups of M two-dimensional points with the two-dimensional Gaussian probability density function, and calculate the distance from each two-dimensional point in each group to the center point as the radius of the ring. The data features of each two-dimensional point include the radius, the variances in two directions, and the pressure difference information.

[0009] Each two-dimensional point corresponds to a ring radius, and each ring is arranged in order to form a pressure difference value matrix.

[0010] The arrangement method is as follows: Use 0XEA as the start index of the overall data to distinguish the start of each frame of data, and 0XEB as the start index of GDP (Gaussian process regression) to distinguish the data between each ring. After the overall data start index, it is the overall characteristic value of the data of each two-dimensional point. Multiple groups of M two-dimensional points form a pressure difference value matrix. When the data features of multiple groups of M two-dimensional points are all filled into the matrix, the remaining end is filled with zeros.

[0011] Then, quantize the matrix through the DPCM algorithm to obtain a differential pressure value matrix, and then use Huffman coding for compression to achieve real-time transmission of the tactile sensor data.

[0012] Furthermore, the data transmission of the sensor is controlled within a few milliseconds, and the processing time of Gaussian process regression GPR is within a few microseconds.

[0013] Furthermore, the dynamic change between adjacent moments of the tactile sensor is described in the differential pressure value matrix of the GPR-DPCM algorithm. Estimate the noise variance according to the variance of the differential pressure value matrix in the quantization stage of the GPR-DPCM algorithm, and only move the local area within the region where the vibration signal changes significantly, and replace the original pressure value with the filtered pressure value for filtering.

[0014] After obtaining the transmitted signal at the receiving end and passing through the improved adaptive local noise reduction (IALNR) algorithm, the final pressure difference signal matrix is obtained through decoding and inverse quantization and superimposed with the previous frame signal to obtain the latest tactile signal.

[0015] Furthermore, the resolution of the tactile sensor is 32*32, and the number of M is 10.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] The present invention first uses vibration signals (pressure changes) as the data for signal transmission. In the prediction stage, to describe the pressure changes, the GPR-DPCM algorithm employs Gaussian Process Regression (GPR) technology to predict the potential range of effective sensing points (points with pressure changes) in the sampling stage. Considering the temporality of the tactile pressure change distribution, the previous information can accurately predict the subsequent tactile pressure change distribution. Therefore, GPR utilizes the powerful capabilities of the Gaussian Process (GP) to predict the upcoming state of the system by effectively using historical data for predicting the pressure change distribution. And a single two-dimensional Gaussian distribution model is adopted instead of the time-consuming Gaussian Mixture Model (GMM). However, the minimum probability threshold defined by the single Gaussian model may lead to omission of a large amount of data of effective sensing points. So, to reduce the problem of omission of effective signals, in the sampling stage, GPR is used to predict the effective tactile sensing points whose pressure changes are described by a single Gaussian distribution, and then the tactile points are undersampled in the form of a Gaussian ring according to the predicted Gaussian Probability Density Function (PDF). Based on this undersampling, random sampling and multiple samplings based on the Gaussian probability density function are carried out. In this way, the selection of sampling points is based on the probability of signal occurrence, and compared with the random selection of ordinary undersampling, it is more likely to cover the key areas of the signal and reduce the possibility of omitting important information. Multiple samplings are carried out, and according to the resolution of the tactile sensor, ten points are generated each time and the sampling is repeated three times. Multiple samplings make the results more accurately reflect the true tactile information characteristics, reduce the influence of biases or outliers generated by single sampling, and further reduce the risk of omitting key information. The difference from the current undersampling lies in the strategy and the degree of retention of data characteristics. The method of the present invention can further retain the key data characteristics. Subsequently, the sampled tactile signals will be subjected to differential quantization, re-indexing, and compression using DPCM technology before transmission.

[0018] In the rendering stage, the present invention uses an improved Adaptive Local Noise Reduction (IALNR) algorithm. This algorithm mainly focuses on accurately calculating the filtering parameters and the filtering space, thereby shortening the filtering time and enhancing the filtering effect. This algorithm no longer evaluates the noise variance based on the overall variance of the signal itself, but uses the pressure changes to estimate the noise variance, thus improving the filtering process. By implementing the adaptive local filtering algorithm, the filtering is only carried out in the areas where there are effective tactile sensing signals, rather than uniform filtering, thereby shortening the filtering time.

[0019] In summary, the experimental results show that the present invention is a brand-new method for transmitting tactile interaction information, aiming to improve the real-time performance and accuracy of tactile information processing. Compared with the traditional direct transmission algorithm and the DPCM algorithm, the GPR-DPCM algorithm proposed by the present invention shortens the transmission time of tactile information and improves the real-time performance. Brief Description of the Drawings

[0020] Figure 1 is a schematic flowchart of the method of the present invention.

[0021] Figure 2 is a comparison diagram of the effects of the improved adaptive local noise reduction (IALNR) algorithm of the present invention and the existing algorithm. Detailed Embodiments

[0022] The present invention will be further explained below in conjunction with embodiments and the accompanying drawings, but this is not intended to limit the protection scope of the present application.

[0023] The present invention mainly proposes a technical method that combines Gaussian process regression (GPR) and differential pulse code modulation (DPCM) for tactile interaction information transmission, and improves the adaptive local noise reduction (ALNR) filter. If the vibration signals of the tactile sensor are directly uploaded, the real-time performance is poor. The present application uses GPR to predict the vibration signals, and processes them according to the prediction results, which can achieve rapid update, enabling data transmission within a few microseconds.

[0024] GPR-DPCM algorithm: In the processing of vibration signals, first use GPR to predict the distribution of vibration signals, predict the potential range of effective sensing points (points with pressure changes) in the sampling stage, fit with a two-dimensional Gaussian probability density function, determine its center point and variances in two directions, and represent the potential range with a Gaussian probability distribution density function; set the number M of randomly sampled points according to the resolution of the tactile sensor, randomly generate multiple groups of M two-dimensional points with a two-dimensional Gaussian probability density function, and calculate the distance from each two-dimensional point in each group to the center point as the radius of the ring; the data characteristics of each two-dimensional point include the radius, variances in two directions, and differential pressure information; then quantize the matrix through the DPCM algorithm to obtain a differential pressure value matrix.

[0025] The present invention can obtain effective signals by combining random sampling based on a two-dimensional Gaussian probability density function. This random sampling method selects sampling points based on the probability of signal occurrence, conducts multiple samplings, generates ten points each time according to the resolution of the tactile sensor, and repeats the sampling three times to form a vibration signal matrix, retaining key data characteristics.

[0026] After obtaining the differential pressure value matrix, Huffman coding compression is then used to achieve real-time transmission of tactile sensor data.

[0027] The dynamic changes between adjacent moments of the tactile sensor are described in the differential pressure value matrix of the GPR-DPCM algorithm. Estimate the noise variance based on the variance of the differential pressure value matrix in the quantization stage of the GPR-DPCM algorithm, and only move the local area within the region where the vibration signal changes significantly, replacing the original pressure value with the filtered pressure value for filtering;

[0028] After obtaining the transmitted signal at the receiving end and passing it through the improved adaptive local noise reduction (IALNR) algorithm, the final pressure difference signal matrix is obtained through decoding and inverse quantization, and is superimposed with the previous frame signal to obtain the latest tactile signal.

[0029] The improved adaptive local noise reduction (IALNR) algorithm is to use a Gaussian filter to estimate the noise variance based on the variance of the differential pressure value matrix in the quantization stage of the GPR-DPCM algorithm, and only move the local area within the region where the vibration signal changes significantly, replacing the original pressure value with the filtered pressure value for filtering.

[0030] Such processing in this application can reduce the computational load, improve the filtering effect, stabilize the pressure values of unsampled points, and enhance the continuity and integrity of the tactile signal characteristics.

[0031] Embodiment 1

[0032] The method for improving the real-time performance of the tactile sensor in the present invention uses the Gaussian process regression-differential pulse code modulation (GPR-DPCM) algorithm for sampling and compression, and uses the improved adaptive local noise reduction (IALNR) algorithm for presentation. The steps of the GPR-DPCM algorithm are as follows:

[0033] 1) Prediction. Considering the temporality of the tactile pressure change distribution, the information of the previous moment can accurately predict the subsequent tactile pressure change distribution. According to the tactile signals of the previous ten moments, a single two-dimensional Gaussian distribution model is used, and Gaussian process regression (GPR) is used to predict the potential range of the effective sensing points in the sampling stage, that is, to obtain the potential distribution of the predicted pressure sensing points.

[0034] 2) Sampling. According to the potential distribution of the predicted pressure sensing points, determine its center point and the variances in the x and y directions. Set the number M of randomly sampled points according to the resolution of the tactile sensor, and randomly generate multiple groups of M two-dimensional points with a two-dimensional Gaussian probability density function. Calculate the distance from each two-dimensional point in each group to the center point as the radius of the ring; the data characteristics of each two-dimensional point include the radius, the variances in two directions, and the pressure difference information. Each ring is a Gaussian ring, and the Gaussian ring is used to undersample the tactile points.

[0035] 3) Quantization. Replace the pressure value of the sampled tactile point with the differential pressure value. Each differential pressure value is assigned one byte. These differential pressure values, together with their corresponding indices, are organized into a pressure difference value matrix of size 128×n, where n depends on the number of sampled tactile points, and the setting of 128 is related to the tactile signal acquisition device used.

[0036] 4) Encoding. Use a Huffman encoder to perform Huffman encoding on the pressure difference value matrix, and then transmit it to the receiving end.

[0037] The specific algorithm is as follows:

[0038] Use a two-dimensional matrix to represent the pressure difference:

[0039] P i =(px i , py i , σ xxi , σ xyi , σ yxi , σ yyi )

[0040] Where i represents the moment, px i , py i represent the positions of the maximum pressure difference values. σ xxi , σ xyi , σ yxi , σ yyi represent the degree of dispersion of the pressure difference values (x, y represent the coordinates of the tactile points, and the coordinate origin can be defined by oneself according to the adopted tactile signal acquisition device). Since σ xyi , = σ yxi , P i actually has five parameters.

[0041] During the prediction process of the GPR-DPCM algorithm, use GPR to predict each future parameter.

[0042] First, calculate and collect the known sequence R t-1 before moment t:

[0043] R t-1 =(P1, P2,..., P t-1 )

[0044] Subsequently, establish the mean function and kernel function of the Gaussian process GP. Assume that the mean function is zero, and the covariance function k(s, s') corresponds to the kernel function δ. The specific expressions are as follows:

[0045]

[0046] s and s' represent time variables, σ f and l are hyperparameters, σ n = 0.3. The hyperparameters can be solved by finding the maximum likelihood estimate of the variables.

[0047] Then, the joint Gaussian probability density function is as follows:

[0048]

[0049] Where R t-1 represents the expected parameter historical values up to moment t - 1. P tDenote the expected parameters at time t. K, K * , K ** is the covariance function matrix. P t The mean and variance of can be determined by the posterior distribution. The specific expressions are as follows:

[0050]

[0051] The parameter at time t: P t =(px t , py t , σ xxt , σ xyt , σ yxt , σ yyt ) can be obtained from the parameter P sequence before time t using GRP t .

[0052] After obtaining the predicted differential pressure distribution, the tactile signals are sampled. To ensure comprehensive sampling of tactile points with pressure changes, annular undersampling based on the Gaussian probability density function is adopted in the sampling stage, which can increase the acquisition of tactile data at effective perception points. In the two-dimensional Gaussian probability density function, points with equal probability form an equiprobability ellipse, and the general two-dimensional Gaussian probability density function is simplified to the following elliptical mathematical equation:

[0053]

[0054] where μ x and μ y are the means of the x and y variables respectively. σ x and σ y are the standard deviations of the x and y variables respectively. ρ is the correlation coefficient between x and y. C is a constant related to x and y.

[0055] According to the above formula, the constant C can be determined by obtaining specific values of x and y. Therefore, initially, a set of ten two-dimensional points (tactile sensor resolution is 32*32) are generated according to the predicted two-dimensional Gaussian probability density function, and 10 fixed ellipse equations can be derived. Given the randomness of the generated points, any point showing potential pressure changes can be randomly selected, thus solving the limitation of uniform sampling. Compared with the fixed-region sampling method, this method can ensure that more effective tactile perception points are sampled while maintaining the same sample size.

[0056] The above random sampling method is repeated three times to ensure comprehensive sampling. The specific number of repetitions needs to be determined through experiments according to the resolution of the tactile sensor used.

[0057] After the sampling stage, in order to adapt to the range of pressure difference values between consecutive time instants, one byte is allocated to store the pressure difference value of each tactile point. For ease of subsequent processing, the pressure difference values and their corresponding index values are systematically arranged in a specific order (the specific order means that they can be transmitted row by row. If there are 128 data in the first row, then the labels of the first row are 0 - 127, the second row is 128 - 255, and so on, and each data has an index. They can also be arranged column by column) to form a pressure difference value matrix. To ensure that different data can be clearly distinguished in the subsequent inverse quantization stage, start indices are set at the beginning of different data. Among them, 0XEA is the start index of the overall data, which distinguishes the start of each frame of data. After the start index of the overall data is the overall data characteristic value (including radius, variance, and pressure difference information). After the overall data characteristic value is 0XEB, and 0XEB is the start index of GDP (Gaussian Process Regression), which distinguishes the data between each ring.

[0058] Since the number of tactile points is unpredictable, zeros are added at the end of the matrix. In this way, if the number of data in the last row does not reach the threshold of 128 bytes, additional values will be added to form a complete row.

[0059] In the last step, the differential pressure value matrix is encoded using Huffman coding as a mechanism to reduce the amount of data during transmission. Utilizing the probability distribution of the transmitted data content, the number of encoding bits is minimized, which further reduces the amount of transmitted data.

[0060] The specific improvement measures of the Improved Adaptive Local Noise Reduction (IALNR) algorithm are as follows:

[0061] 1) Consider the variance of the pressure difference value matrix in the quantization stage of the GPR - DPCM algorithm as the noise variance.

[0062] 2) Filter in the most important area involved in the pressure difference value matrix (here the so - called important area is the transmitted pressure change matrix).

[0063] When applying the adaptive local noise reduction filter to the tactile signal, the local denoising area adopts a 9×9 dimension. The improved filter calculates the filtered pressure value using the following formula:

[0064]

[0065] where x and y represent the signal coordinates, f(x, y) and g(x, y) represent the filtered pressure value and the original pressure value respectively, σ η is the noise variance, σ L is the signal variance of the local area, and m L represents the average value of the local area.

[0066] The dynamic changes between adjacent moments of the tactile sensor are described in the pressure difference matrix of the GPR-DPCM algorithm. In this matrix, the superposition of random noise and the effective pressure change value produces the observed effective pressure change. Therefore, the characteristics of the noise distribution can be better determined by studying the variance of this matrix. Once the noise variance is determined, the local area can be moved within the most important area represented by the pressure difference matrix. As Figure 2 shown in (b) therein, this movement is to replace the original pressure value with the filtered pressure value. Compared with Figure 2 the method of filtering all signals shown in (a) therein, filtering within the area range concentrated in the pressure difference matrix can effectively prevent the amplification of noise in the area without pressure change and shorten the filtering time.

[0067] After the transmission signal is obtained at the receiving end and passed through the improved adaptive local noise reduction (IALNR) algorithm, the final pressure difference signal matrix is obtained through decoding and inverse quantization and superimposed with the previous frame signal to obtain the latest tactile signal.

[0068] The present invention compares the GPR-DPCM algorithm with the DPCM algorithm and direct transmission. During the experiment, assuming that the original signal accurately represents the real-time sampling data of the tactile sensor acquisition system, the sampling time is not considered. Subsequently, the tactile signal displayed at the receiving end is compared with the original signal to evaluate the real-time performance. The bias error is quantified using the mean square error (MSE). To improve the accuracy of MSE calculation, the captured voltage value is directly used as the tactile signal, and the voltage range is expanded to 0-100.

[0069] The tactile sensor adopts two algorithms of Gaussian similar pressure distribution and Gaussian dissimilar pressure distribution, and analyzes its real-time performance. For evaluation, 10 frames are selected from the instances where both types of pressure distributions are applied to the sensor. The MSE values are calculated for these ten frames, and a graph is plotted according to the results. As Figure 1 shown, compared with the DPCM algorithm and direct transmission, the GPR-DPCM algorithm always shows a lower MSE. This ablation study confirms the significance of predicting the sampling area and adopting undersampling for improving real-time performance.

[0070] Compared with the existing commonly used algorithms, the GPR-DPCM algorithm of the present invention has a higher transmission rate and can quickly update the tactile signal.

[0071] The improved ALNR filter of the present invention uses a more personalized variance to achieve the purpose of denoising, and its filtering range is limited to the area of particular interest. The effectiveness of the IALNR filter in reducing the filtering time and enhancing the filtering effect.

[0072] The basis of the present invention is the use of a tactile signal acquisition system designed by our research group. The system includes a capacitive array tactile sensor and a data acquisition (DAQ) board designed specifically for collecting tactile signals. The tactile sensor itself consists of several integral components, including a protective film made of thermoplastic polyurethane rubber, a sponge dielectric layer, an upper electrode layer, and a lower electrode layer. These electrode layers are made of conductive tape or ion layer electrodes. There are 32 electrodes in total in the upper and lower electrode layers, and the spacing between each electrode is 1 mm. In addition, the direction of the upper layer electrodes is perpendicular to that of the lower layer electrodes. The tactile sensor actually forms a 32×32 capacitor matrix, enabling the detection of pressure fluctuations. The basic measurement principle of this capacitive tactile sensor is to evaluate the applied pressure by monitoring the change in the gap between two parallel capacitor plates.

[0073] A capacitance response test was conducted using a material testing system (MTS, ZQ-990B) and an LCR measuring device (IM 3536) to test the electrical and mechanical characteristics of the tactile sensor. At a frequency of 30 kHz, the response of the capacitance value to pressure shows an almost linear change, indicating that the tactile sensor has a wide measurement range.

[0074] What is not described in the present invention applies to the prior art.

Claims

1. A method for improving the real-time performance of a tactile sensor, characterized in that, The method includes the following steps: Obtain the vibration signal of the tactile sensor. First, use Gaussian Process Regression (GPR) to predict the distribution of the vibration signal, predict the potential range of effective sensing points (points with pressure changes) in the sampling stage, obtain the two-dimensional Gaussian probability distribution function, and thus determine its center point and variances in the x and y directions. Set the number M of randomly sampled points according to the resolution of the tactile sensor. Randomly generate multiple groups of M two-dimensional points with the two-dimensional Gaussian probability density function, and calculate the distance from each two-dimensional point in each group to the center point as the radius of the ring. The data characteristics of each two-dimensional point include the radius, variances in two directions, and differential pressure information. Each two-dimensional point corresponds to a ring radius, and each ring is arranged in sequence to form a differential pressure value matrix. The arrangement method is as follows: Use 0XEA as the start index of the overall data to distinguish the start of each frame of data, and 0XEB as the start index of GDP (Gaussian Process Regression) to distinguish the data between each ring. After the start index of the overall data, there are the overall characteristic values of the data of each two-dimensional point. Multiple groups of M two-dimensional points form a differential pressure value matrix. When the data characteristics of multiple groups of M two-dimensional points are all filled into the matrix, the remaining end is filled with zeros. Then, quantize the matrix through the DPCM algorithm to obtain a differential pressure value matrix, and then use Huffman coding for compression to achieve real-time transmission of the tactile sensor data.

2. The method according to claim 1, characterized in that, The data transmission of the sensor is controlled within a few milliseconds, and the processing time of Gaussian Process Regression (GPR) is within a few microseconds.

3. The method according to claim 1, wherein The dynamic changes between adjacent moments of the tactile sensor are described in the differential pressure value matrix of the GPR-DPCM algorithm. Estimate the noise variance according to the variance of the differential pressure value matrix in the quantization stage of the GPR-DPCM algorithm. Only move the local area within the region where the vibration signal changes significantly, and replace the original pressure value with the filtered pressure value for filtering. After obtaining the transmitted signal at the receiving end and passing it through the Improved Adaptive Local Noise Reduction (IALNR) algorithm, the final differential pressure signal matrix is obtained through decoding and inverse quantization, and is superimposed with the previous frame signal to obtain the latest tactile signal.

4. The method according to claim 1, wherein The resolution of the tactile sensor is 32*32, and the number of M is 10.

Images