CN113627337B - Force touch signal processing method based on stack type automatic coding - Google Patents
Force touch signal processing method based on stack type automatic coding Download PDFInfo
- Publication number
- CN113627337B CN113627337B CN202110916247.8A CN202110916247A CN113627337B CN 113627337 B CN113627337 B CN 113627337B CN 202110916247 A CN202110916247 A CN 202110916247A CN 113627337 B CN113627337 B CN 113627337B
- Authority
- CN
- China
- Prior art keywords
- signal
- error
- haptic signal
- reconstruction
- noise
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000003672 processing method Methods 0.000 title abstract description 5
- 238000000034 method Methods 0.000 claims abstract description 47
- 238000007906 compression Methods 0.000 claims abstract description 27
- 230000006835 compression Effects 0.000 claims abstract description 27
- 230000015654 memory Effects 0.000 claims abstract description 22
- 238000012545 processing Methods 0.000 claims abstract description 13
- 239000013598 vector Substances 0.000 claims description 26
- 239000011159 matrix material Substances 0.000 claims description 18
- 238000001914 filtration Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 230000007787 long-term memory Effects 0.000 claims description 6
- 230000006403 short-term memory Effects 0.000 claims description 6
- 238000012549 training Methods 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 5
- 230000000717 retained effect Effects 0.000 claims description 4
- 241000364483 Lipeurus epsilon Species 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 230000005540 biological transmission Effects 0.000 abstract description 4
- 230000008901 benefit Effects 0.000 abstract description 3
- 238000011084 recovery Methods 0.000 abstract 1
- 230000008447 perception Effects 0.000 description 7
- 238000013144 data compression Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 4
- 239000000463 material Substances 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 241000282414 Homo sapiens Species 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 239000004744 fabric Substances 0.000 description 2
- 230000036039 immunity Effects 0.000 description 2
- 238000001228 spectrum Methods 0.000 description 2
- 239000002023 wood Substances 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000002146 bilateral effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000007654 immersion Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000013139 quantization Methods 0.000 description 1
- 230000005236 sound signal Effects 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Complex Calculations (AREA)
- User Interface Of Digital Computer (AREA)
Abstract
The invention provides a power touch signal processing method based on stack type automatic coding, and belongs to the field of digital signal processing. The coding part of the stack type automatic coding is applied to compress the force touch data which is obtained from the physical surface and shows the key information of the shape and the texture of the force touch data in real time, the characteristics with the dimension far smaller than the dimension of the force touch data are extracted from the original force touch signal, the noise which is mixed in the transmission process of the characteristics is considered, the long-short-time memory network is adopted as a filter at the receiving end to filter the noisy characteristics, and the filtered characteristics are reconstructed to the dimension same as the original data by the decoding part of the stack type automatic coding, so that the compression and the recovery of the signals are realized. The method has the advantages that the method can be realized on line through the execution time of millisecond level, improves the compression ratio of the force touch signal, can remove noise brought by data in the channel transmission process to a certain extent, and ensures lower reconstruction error.
Description
Technical Field
The invention belongs to the field of digital signal processing, relates to compression, reconstruction and filtering methods of a force touch signal, and in particular relates to a force touch signal processing method based on stack type automatic coding, which can realize high-efficiency compression of the force touch signal, ensure lower reconstruction error, has a filtering function and can be better applied to the fields of touch data compression, touch remote presentation and the like.
Background
With the development of haptic reproduction technology and the internet, there is an increasing demand for immersive remote operations and interactions with the physical world by internet users. Haptic internet has great potential applications in bilateral teleoperation and online shopping, etc., and additional demands are made of the haptic internet to achieve an immersive haptic experience, including collection, compression, transmission, reconstruction and display of haptic data. The force haptic signals may exhibit the shape and texture of objects through advanced surface haptic devices, which enable the haptic internet to improve interactive immersion with real or virtual environments. With the time-varying of force haptic signals acquired by force sensors, how to achieve reliable online compression and reconstruction is a challenge for the haptic internet.
In 2010, okamoto et al studied the human perception of texture of vibratable materials, focusing on the impact of subthreshold vibration amplitude and amplitude variation on perception, and expected that haptic data would be compressed by the perception characteristics, using lossy data compression. Data compression while maintaining subjective quality, while processing multiple variables, three lossy data compression methods were applied to the surface texture of wood and sandpaper: linear and logarithmic quantization of amplitude in frequency space, and data truncation under the displacement detection threshold curve. The test participants scored the subjective similarity of these compressed textures to the uncompressed original textures. Without significant differences in rating being observed, subjective perception of wood and sandpaper achieved compression ratios approaching 4:1 while guaranteeing quality, indicating that the data size of the material texture can be reduced by about 75%. However, when the data dimension is large, the data compression effect caused by such a compression ratio is very limited, and only the tactile data generated by the material with vibration characteristics can be compressed, so that the universality is lacking.
In 2012, chaudhari et al inspired by the similarity between the vibro-rotational texture signal and the speech signal, proposed a method for compressing the vibro-rotational texture in haptic teleoperation based on a source filter model, and adopted a speech coding technique. The determined grouping rate is formed through constant output bit rate, so that the design of the audio-video tactile multiplexer is simplified to a certain extent, and the texture surface is freely and naturally scanned in real time through frame-by-frame processing. The final texture codec achieves a compression ratio of 8:1 while preserving the spectral characteristics of the original signal and its temporal characteristics. But this approach is no longer applicable once the haptic signal that needs to be compressed has no characteristics similar to the speech signal.
The existing main compression modes of the touch data are two types, namely, the data smaller than a perception threshold are directly discarded to be compressed through the principle that human beings have the smallest perception difference in psychophysics, the compression ratio achieved by the mode is very limited, the smallest perception threshold has larger difference for different users, and accurate and specific standards are difficult to form to compress signals. Another is to compress certain haptic signals by means of speech signal processing, which takes advantage of the amplitude-frequency characteristics of the haptic signals, which are similar to those of the sound signals, and the like, and which are no longer suitable when they do not, most haptic signals do not have such characteristics, and the compression achieved by the compression based on the amplitude-frequency characteristics of the signals is relatively low, similar to the former method. And has limitations.
Disclosure of Invention
The invention provides a force touch signal processing method based on stack-type automatic coding, which aims to solve the problems that the existing touch signal compression can only compress signals with certain characteristics, and finally, the compression ratio is not ideal.
The technical scheme adopted by the invention is that the method comprises the following steps:
(1) Performing dimension reduction on the original tactile signal by using a rectangular weight matrix and a bias vector to obtain the characteristics of the tactile signal;
(2) Reconstructing the characteristics by using a rectangular weight matrix and a bias vector, and recovering the characteristics to the same dimension of the input signal to obtain the reconstruction of the touch signal;
(3) Calculating an average relative error between the original haptic signal and the reconstructed haptic signal
(4) Deriving the error updates weights and offsets in the direction of derivative decrease to obtain encoder and decoder
(5) Mixing Gaussian white noise into the characteristics of the compressed data obtained through the model;
(6) Filtering noise-containing features using long-short-term memory neural network as filter
(7) Calculating the average absolute error between the original features and the filtered features, updating parameters of each gate in the long-short memory network along the gradient descending direction until the set updating times, and storing a model with the minimum error as a filter;
(8) The filtered features are fed as input to a decoder to obtain a reconstructed haptic signal.
The specific way of obtaining the tactile signal characteristics in the step (1) is as follows:
the original force touch signal s with the length of N passes through a first layer encoder, and preliminary characteristics are obtained through a first group of weight matrixes and offset vectors:
f (1) ＝σ(W (1) s+b (1) )
wherein the method comprises the steps ofIs the extracted preliminary feature, M (1) < N represents the length of the preliminary feature, +.>Is a weight matrix for extracting preliminary features, +.>The offset vector corresponding to the offset vector, sigma is an activation function, and ReLU, sigmoid and the like can be selected, and the form is as follows:
wherein x is any random variable, and the obtained preliminary features are further compressed to obtain final features
Where l.epsilon. {2, …, L } is the number of layers of the neural network used to extract the features,f (l-1) representing the features extracted from layer 1, M (l) ＜M (l-1) For the dimension of the first layer, in particular, let M (L) ＝M，f＝f (L) Is characteristic of a force tactile signal.
The specific way of reconstructing the haptic signal in the step (2) of the present invention is:
wherein the method comprises the steps ofWeight matrix required for reconstruction in decoder, < > and method for the reconstruction in decoder>Bias vector in reconstruction, ++>To reconstruct the data obtained using deterministic mapping, at the last layer of the decoder we get a reconstruction of the original signal:
wherein the method comprises the steps ofHaving the same dimensions as the original data.
The specific way of obtaining the average relative error between the original haptic signal and the reconstructed haptic signal in step (3) of the present invention is:
each compression and reconstruction forms a set of weights and bias parameters:
it determines the error between the original data and its reconstruction as follows:
where J represents the number of haptic signals involved in training, |·| represents the absolute value operation performed element by element, and division is also performed element by element.
The step (4) of the invention is to derive the error, update the weight and bias along the direction of the decrease of the derivative, and obtain the encoder and decoder, the specific method is as follows:
Θ (l) the final value Θ of (2) (l)* Determined by the following formula:
t e {0, …, T } represents the T-th update, derivative of the error:
update along L (Θ) (l) ) The direction of derivative decrease is updated until t=tstreams are over, and the weight and bias at which the error is minimal are retained as their final values.
In the step (5), gaussian white noise is mixed in the characteristic signal, and the method is realized by the following steps:
first, the power of the haptic signal characteristics to be mixed with noise is calculated, f i The i-th element representing f:
then randomly generating high white noise n of the same dimension as the haptic signal feature dimension and calculating its power:
in order to maintain a certain signal-to-noise ratio for gaussian white noise to the feature, each data point in the noise needs to be multiplied by a coefficient:
wherein SNR represents the signal-to-noise ratio we set, and adding the noise to the characteristics yields characteristics of the mixed noise
The specific way of filtering the noise-containing characteristic by using the long-short-term memory neural network as a filter in the step (6) is as follows:
sending the tactile signal characteristics with noise into a long-short-term memory network, setting the target output as the original characteristics, filtering by cells in the network, wherein each cell contains three gates, and controlling a forgetting gate F whether the previous memory is forgotten or not t ：
Determining candidate statesInfluence of the current state C t Input gate I of the extent of (2) t ：
Output gate O for controlling how much hidden state can be output at the current time t ：
Together they constitute a cell in a long and short term memory network, which is calculated as follows:
wherein δ (. Cndot.) represents the sigmoid function, +.As vector (I)>Represents the t-th element, t is {1, …, M }, weight matrix ∈ ->Bias vector->The hidden state H is the same as the dimension of the tactile feature, and the filtered feature and the new hidden state can be obtained through the following formula after one cell:
H t ＝O t ⊙tanh(C t )
wherein the method comprises the steps ofIs a vector of length N, ">Is a real number. Since the output of the long and short term memory network depends not only on the current input, but also on the previous state, this allows him to use the potential relationships between elements to obtain better filtering results. The error of the filter is determined by:
wherein Γ is all parameters in the long-short-term memory network, and the value mode is as follows:
deriving the error of the filter:
the parameters are updated in the direction of the decrease in the derivative of L (Γ) until the t=tstreams are over, and the weights and offsets at which the error is minimal are retained as the final parameters of the filter.
The specific way of obtaining the reconstructed haptic signal in the step (8) is as follows:
reconstructing the characteristics of the filtered haptic signal obtained in the previous step:
wherein the method comprises the steps ofFor decodingWeight matrix required for reconstruction in the device, < >>The bias vectors in the reconstruction are parameters in the step (4). />To reconstruct the data using deterministic mapping, at the last layer of the decoder we get:
has the same dimensions as the original data, and is a force haptic signal reconstructed for the filtered features.
The method provided by the invention realizes compression, filtering and reconstruction of the haptic signals through the stack type automatic coding and long-short-time memory network, and has the main advantages that:
the compression ratio of more than 100:1 can be reached, the frequency spectrum characteristic of the signal is not required, and the applicability of the compression method is improved;
the long-short-time memory network is used for filtering noise mixed in the transmission process, so that the whole system has certain noise immunity;
the average relative error between the reconstructed characteristics without mixed noise and the original signal can reach below 5%, and the reconstructed characteristics with mixed noise and the original signal have lower average relative error.
Drawings
FIG. 1 is a block diagram of a codec model generation based on stacked auto-coding and a filter model generation based on long and short memory networks in accordance with the present invention;
FIG. 2 is a data flow diagram of the present invention;
FIG. 3 is a graph of average relative reconstruction error versus compression ratio variation for a haptic signal in accordance with the present invention;
fig. 4 is a graph showing the average relative error between the reconstructed features mixed with white gaussian noise and the original signal as a function of the signal-to-noise ratio.
Detailed Description
Comprises the following steps:
(1): the original tactile signal is subjected to dimension reduction by applying a rectangular weight matrix and a bias vector, and the characteristics of the tactile signal are obtained, wherein the specific way is as follows:
the original force touch signal s with the length of N passes through a first layer encoder, and preliminary characteristics are obtained through a first group of weight matrixes and offset vectors:
f (1) ＝σ(W (1) s+b (1) )
wherein the method comprises the steps ofIs the extracted preliminary feature, M (1) < N represents the length of the preliminary feature, +.>Is a weight matrix for extracting preliminary features, +.>The offset vector corresponding to the offset vector, sigma is an activation function, and ReLU, sigmoid and the like can be selected, and the form is as follows:
wherein x is any random variable, and the obtained preliminary features are further compressed to obtain final features:
f (l) ＝σ(W (l) f (l-1) +b (l) )
where l.epsilon. {2, …, L } is the number of layers of the neural network used to extract the features,f (l -1) representing the features extracted from layer 1, M (l) ＜M (l-1) For the dimension of the first layer, in particular, let M (L) ＝M，f＝f (L) Is characteristic of a force sense signal;
(2) And reconstructing the characteristics by using a rectangular weight matrix and a bias vector, and recovering the characteristics to the same dimension of the input signal to obtain the reconstruction of the touch signal, wherein the specific method comprises the following steps of:
wherein the method comprises the steps ofWeight matrix required for reconstruction in decoder, < > and method for the reconstruction in decoder>Bias vector in reconstruction, ++>In order to use the reconstructed data obtained by deterministic mapping, at the last layer of the decoder, a reconstruction of the original signal is obtained:
wherein the method comprises the steps ofThe same dimension as the original data;
(3) Calculating an average relative error between the original haptic signal and the reconstructed haptic signal:
each compression and reconstruction forms a set of weights and bias parameters:
it determines the error between the original data and its reconstruction as follows:
where J represents the number of haptic signals involved in training, |·| represents the absolute value operation performed element by element, and division is also performed element by element.
(4) Deriving the error updates the weights and offsets in the direction of the derivative decrease to obtain the encoder and decoder:
Θ (l) the final value Θ of (2) (l)* Determined by the following formula:
t e {0, …, T } represents the T-th update, derivative of the error:
update along L (Θ) (l) ) Updating the direction of the derivative decrease until t=Tupdate is finished, and reserving the weight and bias with the minimum error as the final value;
(5) Obtaining the characteristics of compressed data through a model, and mixing Gaussian white noise into the characteristics:
first, the power of the haptic signal characteristics to be mixed with noise is calculated:
then randomly generating the same dimension as the characteristic dimension of the touch signalHigh white noise n i And calculates its power:
in order to maintain a certain signal-to-noise ratio for gaussian white noise to the feature, each data point in the noise needs to be multiplied by a coefficient:
wherein SNR represents the signal-to-noise ratio we set, and adding the noise to the characteristics yields characteristics of the mixed noise
(6) Filtering the noise-containing features by using the long-short-term memory neural network as a filter:
sending the tactile signal characteristics with noise into a long-short-term memory network, setting the target output as the original characteristics, filtering by cells in the network, wherein each cell contains three gates, and controlling a forgetting gate F whether the previous memory is forgotten or not t ：
Determining candidate statesInfluence of the current state C t Input gate I of the extent of (2) t ：
Output gate O for controlling how much hidden state can be output at the current time t ：
Together they constitute a cell in a long and short term memory network, which is calculated as follows:
wherein δ (. Cndot.) represents the sigmoid function, +.As vector (I)>Represents the t-th element, t is {1, …, M }, weight matrix ∈ ->Bias vector->The hidden state H is the same as the dimension of the tactile feature, and the filtered feature and the new hidden state can be obtained through the following formula after one cell:
H t ＝O t ⊙tanh(C t )
wherein the method comprises the steps ofIs a vector of length N, ">Is a real number. Since the output of the long and short term memory network depends not only on the current input but also on the previous state, this allows him to use the potential relationships between elements to obtain better filtering results;
(7) Calculating the average absolute error between the original characteristic and the filtered characteristic, updating parameters of each gate in the long-short-time memory network along the gradient descending direction, and storing a model with the minimum error as a filter until the set updating times, wherein the error of the filter is determined by the following formula:
wherein Γ is all parameters in the long-short-term memory network, and the value mode is as follows:
deriving the error of the filter:
updating along the direction of L (Γ) derivative decrease until t=Tupdating is finished, and preserving the weight and bias at which the error is minimum as the final parameters of the filter;
(8) And sending the filtered features as input to a decoder to obtain a reconstructed haptic signal:
reconstructing the characteristics of the filtered haptic signal obtained in the previous step:
wherein the method comprises the steps ofWeight matrix required for reconstruction in decoder, < > and method for the reconstruction in decoder>The bias vectors in the reconstruction are all parameters in the step (4), and the +.>To reconstruct the data using deterministic mapping, at the last layer of the decoder, we get:
has the same dimensions as the original data, and is a force haptic signal reconstructed for the filtered features.
The applicability of the method for processing the force touch signal based on the stack type automatic coding and long-short-time memory network to the force touch signal is analyzed through simulation experiment data, and software adopted in the simulation experiment is PyCharm software and a Keras frame.
Simulation experiment: the SAE-based haptic signal compression, filtering and reconstruction method was evaluated using the HapTex database. This database contains a tactile signal of the force of a bare finger sliding across the 120 fabric surface, measured 4 times per fabric. Wherein normal and lateral friction signals are employed. The database was divided into training and testing sets using ten fold cross-validation. In each fold, the training set contains 432 sets of data, while the test set contains the rest of the database. The training set is used to train parameters of the stacked automatic codec and the long and short memory network based filter, and the test set is used to evaluate the performance of the station. Fig. 3 shows a graph of average relative reconstruction error of a haptic signal as a function of compression ratio from 32.5 to 260 in accordance with the present invention. Fig. 4 is the error between the reconstructed signal and the original signal after mixing in white gaussian noise with a signal-to-noise ratio of 0 to 20dB in the characteristics and after introducing a filter. According to analysis experimental results, the method provided by the invention can reach a compression ratio of more than 100:1, and can effectively filter noise, so that the whole system has noise immunity; the average relative error between the reconstructed characteristics without mixed noise and the original signal can reach below 5%, and the reconstructed characteristics with mixed noise and the original signal have lower average relative error. The method has no requirement on the frequency spectrum characteristics of the signals, improves the applicability of the compression method, improves the compression ratio and has smaller reconstruction error.
Claims (7)
1. A method for processing a haptic signal based on stack-type automatic coding is characterized by comprising the following steps:
(1) Performing dimension reduction on the original tactile signal by using a rectangular weight matrix and a bias vector to obtain the characteristics of the tactile signal;
(2) Reconstructing the characteristics by using a rectangular weight matrix and a bias vector, and recovering the characteristics to the same dimension of the input signal to obtain the reconstruction of the touch signal; the specific way of reconstructing the haptic signal is:
wherein the method comprises the steps ofWeight matrix required for reconstruction in decoder, < > and method for the reconstruction in decoder>Bias vector in reconstruction, ++>To make use of deterministic mappingThe resulting reconstructed data, at the last layer of the decoder, we have obtained a reconstruction of the original signal:
wherein the method comprises the steps ofThe same dimension as the original data;
(3) Calculating an average relative error between the original haptic signal and the reconstructed haptic signal;
(4) Deriving the error updates weights and offsets in the direction of derivative decrease to obtain encoder and decoder
(5) Mixing Gaussian white noise into the characteristics of the compressed data obtained through the model;
(6) The long-time memory neural network is used as a filter to filter the noise-containing characteristics, and the specific way is as follows:
sending the tactile signal characteristics with noise into a long-short-term memory network, setting the target output as the original characteristics, filtering by cells in the network, wherein each cell contains three gates, and controlling a forgetting gate F whether the previous memory is forgotten or not t ：
Determining candidate statesInfluence of the current state C t Input gate I of the extent of (2) t ：
Input for controlling how much hidden state can be output at the current timeGo out O t ：
Together they constitute a cell in a long and short term memory network, which is calculated as follows:
wherein δ (. Cndot.) represents the sigmoid function, +.As vector (I)>Represents the t-th element, t is {1, …, M }, weight matrix ∈ ->Bias vector->The hidden state H is the same as the dimension of the tactile feature, and the filtered feature and the new hidden state can be obtained through the following formula after one cell:
H t ＝O t ⊙tanh(C t )
wherein the method comprises the steps ofIs a vector of length N, ">As real, the output of the long and short-term memory network depends not only on the current input, but also on the previous state, which makes it possible to use the potential relationship between elements to obtain better filtering results;
(7) Calculating the average absolute error between the original features and the filtered features, updating parameters of each gate in the long-short memory network along the gradient descending direction until the set updating times, and storing a model with the minimum error as a filter;
(8) The filtered features are fed as input to a decoder to obtain a reconstructed haptic signal.
2. The method for processing the haptic signal based on the stack-type automatic coding according to claim 1, wherein: the specific way of obtaining the tactile signal characteristics in the step (1) is as follows:
the original force touch signal s with the length of N passes through a first layer encoder, and preliminary characteristics are obtained through a first group of weight matrixes and offset vectors:
f (1) ＝σ(W (1) s+b (1) )
wherein the method comprises the steps ofIs the extracted preliminary feature, M (1) <N represents the length of the preliminary feature, +.>Is a weight matrix for extracting preliminary features, +.>The offset vector corresponding to the function is selected from ReLU and Sigmoid, and sigma is an activation function, and the form is as follows:
Re LU:
Sigmoid:
wherein x is any random variable, and the obtained preliminary features are further compressed to obtain final features:
f (l) ＝σ(W (l) f (l-1) +b (l) )
where l.epsilon. {2, …, L } is the number of layers of the neural network used to extract the features,f (l-1) representing the features extracted from layer 1, M (l) <M (l-1) For the dimension of the first layer, in particular, let M (L) ＝M，f＝f (L) Is characteristic of a force tactile signal.
3. The method for processing the haptic signal based on the stack-type automatic coding according to claim 1, wherein: the specific way of obtaining the average relative error between the original haptic signal and the reconstructed haptic signal in the step (3) is:
each compression and reconstruction forms a set of weights and bias parameters:
it determines the error between the original data and its reconstruction as follows:
where J represents the number of haptic signals involved in training, |·| represents the absolute value operation performed element by element, and division is also performed element by element.
4. The method for processing the haptic signal based on the stack-type automatic coding according to claim 1, wherein: and (4) deriving the error, updating the weight and the bias along the direction of the decrease of the derivative to obtain the encoder and the decoder, wherein the specific method comprises the following steps of:
Θ (l) the final value Θ of (2) (l)* Determined by the following formula:
t e {0, …, T } represents the T-th update, derivative of the error:
update along L (Θ) (l) ) The direction of derivative decrease is updated until t=tstreams are over, and the weight and bias at which the error is minimal are retained as their final values.
5. The method for processing the haptic signal based on the stack-type automatic coding according to claim 1, wherein: in the step (5), gaussian white noise is mixed into the characteristic signal, and the method is realized by the following steps:
first, the power of the haptic signal characteristics to be mixed with noise is calculated, f i The i-th element representing f:
then randomly generating high white noise n of the same dimension as the haptic signal feature dimension and calculating its power:
in order to maintain a certain signal-to-noise ratio for gaussian white noise to the feature, each data point in the noise needs to be multiplied by a coefficient:
wherein SNR represents the set signal-to-noise ratio, and adding the noise to the characteristics to obtain characteristics of the mixed noise
6. The method for processing the haptic signal based on the stack-type automatic coding according to claim 1, wherein: the error of the filter obtained in the step (7) is determined by the following formula:
wherein Γ is all parameters in the long-short-term memory network, and the value mode is as follows:
deriving the error of the filter:
the parameters are updated in the direction of the decrease in the derivative of L (Γ) until the t=tstreams are over, and the weights and offsets at which the error is minimal are retained as the final parameters of the filter.
7. The method for processing the haptic signal based on the stack-type automatic coding according to claim 1, wherein: the specific way of obtaining the reconstructed haptic signal in the step (8) is as follows:
reconstructing the characteristics of the filtered haptic signal obtained in the previous step:
wherein the method comprises the steps ofWeight matrix required for reconstruction in decoder, < > and method for the reconstruction in decoder>The bias vectors in the reconstruction are all parameters in the step (4), and the +.>To reconstruct the data using deterministic mapping, at the last layer of the decoder, we get:
has the same dimensions as the original data, and is a force haptic signal reconstructed for the filtered features.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110916247.8A CN113627337B (en) | 2021-08-10 | 2021-08-10 | Force touch signal processing method based on stack type automatic coding |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110916247.8A CN113627337B (en) | 2021-08-10 | 2021-08-10 | Force touch signal processing method based on stack type automatic coding |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113627337A CN113627337A (en) | 2021-11-09 |
| CN113627337B true CN113627337B (en) | 2023-12-05 |
Family
ID=78384145
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110916247.8A Active CN113627337B (en) | 2021-08-10 | 2021-08-10 | Force touch signal processing method based on stack type automatic coding |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113627337B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116736975A (en) * | 2023-06-07 | 2023-09-12 | 吉林大学 | Method for generating haptic information based on stack type automatic coding |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109410575A (en) * | 2018-10-29 | 2019-03-01 | 北京航空航天大学 | A kind of road network trend prediction method based on capsule network and the long Memory Neural Networks in short-term of nested type |
| CN110474716A (en) * | 2019-08-14 | 2019-11-19 | 安徽大学 | The method for building up of SCMA codec models based on noise reduction self-encoding encoder |
| CN110866604A (en) * | 2019-10-28 | 2020-03-06 | 国网河北省电力有限公司电力科学研究院 | Cleaning method for power transformer state monitoring data |
| CN111240268A (en) * | 2020-01-14 | 2020-06-05 | 重庆大学 | Axle system thermal error modeling method and thermal error compensation system based on SLSTM neural network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11188813B2 (en) * | 2017-03-01 | 2021-11-30 | Stc.Unm | Hybrid architecture system and method for high-dimensional sequence processing |
-
2021
- 2021-08-10 CN CN202110916247.8A patent/CN113627337B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109410575A (en) * | 2018-10-29 | 2019-03-01 | 北京航空航天大学 | A kind of road network trend prediction method based on capsule network and the long Memory Neural Networks in short-term of nested type |
| CN110474716A (en) * | 2019-08-14 | 2019-11-19 | 安徽大学 | The method for building up of SCMA codec models based on noise reduction self-encoding encoder |
| CN110866604A (en) * | 2019-10-28 | 2020-03-06 | 国网河北省电力有限公司电力科学研究院 | Cleaning method for power transformer state monitoring data |
| CN111240268A (en) * | 2020-01-14 | 2020-06-05 | 重庆大学 | Axle system thermal error modeling method and thermal error compensation system based on SLSTM neural network |
Non-Patent Citations (1)
| Title |
|---|
| 基于长短时记忆网络的股价短期预测研究;成烯;中国优秀硕士学位论文全文数据库经济与管理科学辑(第4期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113627337A (en) | 2021-11-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110619885B (en) | Method for generating confrontation network voice enhancement based on deep complete convolution neural network | |
| CN108877832B (en) | Audio tone quality restoration system based on GAN | |
| Hassen et al. | PVC-SLP: Perceptual vibrotactile-signal compression based-on sparse linear prediction | |
| CN109785847B (en) | Audio compression algorithm based on dynamic residual error network | |
| CN115410583A (en) | Perceptually-based loss functions for audio encoding and decoding based on machine learning | |
| CN110544484B (en) | High-order Ambisonic audio coding and decoding method and device | |
| CN113627337B (en) | Force touch signal processing method based on stack type automatic coding | |
| CN111541900B (en) | Security and protection video compression method, device, equipment and storage medium based on GAN | |
| Lu et al. | Self-supervised audio spatialization with correspondence classifier | |
| Sun et al. | A model compression method with matrix product operators for speech enhancement | |
| CN116233445B (en) | Video encoding and decoding processing method and device, computer equipment and storage medium | |
| Kinsner | Compression and its metrics for multimedia | |
| KR20210003514A (en) | Encoding method and decoding method for high band of audio, and encoder and decoder for performing the method | |
| Masmoudi et al. | A novel bio-inspired static image compression scheme for noisy data transmission over low-bandwidth channels | |
| CN115223244B (en) | Haptic motion simulation method, device, apparatus and storage medium | |
| CN109814716A (en) | A kind of motion intention coding/decoding method based on dynamic surface electromyography signal | |
| Li et al. | Autoencoder for vibrotactile signal compression | |
| CN112564713B (en) | High-efficiency low-time delay kinescope signal coder-decoder and coding-decoding method | |
| Hajizadeh et al. | Predictive compression of animated 3D models by optimized weighted blending of key‐frames | |
| Faridee et al. | Predicting score distribution to improve non-intrusive speech quality estimation | |
| Chen et al. | A two-part predictive coder for multitask signal compression | |
| Guz | A novel image compression method based on classified energy and pattern building blocks | |
| Nockenberg et al. | Mvibcode: Multi-channel vibrotactile codec using hierarchical perceptual clustering | |
| Chithra et al. | Effective lossy and lossless color image compression with Multilayer Perceptron | |
| Wang et al. | Image compression based on wavelet transform and vector quantization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |