WO2018223491A1 - 基于牙齿咬合声音的识别方法及系统 - Google Patents

基于牙齿咬合声音的识别方法及系统 Download PDF

Info

Publication number
WO2018223491A1
WO2018223491A1 PCT/CN2017/092771 CN2017092771W WO2018223491A1 WO 2018223491 A1 WO2018223491 A1 WO 2018223491A1 CN 2017092771 W CN2017092771 W CN 2017092771W WO 2018223491 A1 WO2018223491 A1 WO 2018223491A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
sound
event
identification
threshold
Prior art date
Application number
PCT/CN2017/092771
Other languages
English (en)
French (fr)
Inventor
伍楷舜
赵猛
刘巍峰
邹永攀
Original Assignee
深圳大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Publication of WO2018223491A1 publication Critical patent/WO2018223491A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L17/00—Speaker identification or verification techniques
    • G10L17/26—Recognition of special voice characteristics, e.g. for use in lie detectors; Recognition of animal voices

Definitions

  • the present invention relates to the field of voice recognition technology, and in particular, to a method and system for identifying a tooth based sound.
  • smart watches can provide us with call reminders.
  • EEG recorders we can record the biosignals of each of us EEG all the time, through a series of algorithms analysis and research, to protect the health of each of us, greatly It provides a strong guarantee for the quality of our lives.
  • the privacy of users and the leakage of a series of personal data are particularly prominent. How to protect personal data in the era of big data is particularly important.
  • biometrics with high recognition accuracy.
  • the recognition accuracy is high, and some cost should be paid.
  • fingerprint recognition we all know that on some smart devices and some access control systems, many are equipped with fingerprint recognition technology. When our fingers have sweat, stains, or peeling, fingerprint recognition is easy to identify and must be cleaned up. These things can be re-identified and achieve the effect of identity authentication, which greatly causes great inconvenience to people. At the same time, it is easy for us to leave our fingerprints in many places, so it is easy to be copied by lawless elements. Then achieve the purpose of identification.
  • Vein recognition and iris recognition are also based on each person's physiological characteristics and then take physiological characteristics signals, and then a series of analysis and processing of the signal, and finally achieve the role of recognition, these two biometric technologies can achieve high recognition Accuracy, but also with some drawbacks, veins in the back of the hand in vein recognition may still change with age and physiology, and the permanent has not been confirmed. Because the collection method is limited by its own characteristics, the product is difficult to be miniaturized and cannot be In mass production on smart devices, at the same time, the collection equipment has special requirements, the design is relatively complicated, and the manufacturing cost is high.
  • Face recognition also provides high precision. Face recognition has a lot of loopholes. Only a powerful software for modifying pictures is needed. We can easily imitate the physiological features of people's faces and then perform face recognition. The results of the experiment are also very shocking. It can be correctly identified through a few simple steps, and the protection level is not particularly high.
  • the invention provides a recognition method based on a tooth occlusion sound, comprising the following steps:
  • the identification method further includes:
  • the data processing step includes:
  • the filtering step according to the obtained frequency range, firstly uses Butterworth bandpass filtering, and then the signal after passing through the Butterworth filter is further subjected to third-order adaptive threshold wavelet filtering, and after two noise filtering, effective Remove noise interference;
  • the microphone sensor of the smart device is first detected to collect a sound event, according to the time domain and the frequency domain signal of the sound event. Distribution characteristics, and then calculate the distribution of energy of each sound signal in the frequency range, to obtain the distribution range of frequency accompanying energy, the set frequency range is 100HZ to 10000HZ;
  • the bandpass's Butterworth filter is first used, the bandpass frequency range is 100 Hz to 10000 Hz, and then the third order adaptive threshold is applied to the signal that has passed through the Butterworth filter.
  • Wavelet filtering processing for a noise environment, the set peak threshold is 0.1, and the threshold set at the interval of two tooth occlusion events is within 20,000 sample points;
  • the framed and windowed processing is performed on the twice filtered signal, and each frame of the signal is 10 milliseconds, and the coverage between the frame and the frame is adopted. 5 milliseconds, the added window function is Hamming window, and the event detection is performed on the processed data.
  • the double threshold peak detection method is adopted. Firstly, according to the characteristics of time domain and frequency domain, the minimum of one event is first set.
  • the peak threshold is 0.05, that is, only the signal sample points whose event peak value is greater than or equal to 0.05 are retained, and then the interval between the peaks of the two events is set.
  • the set threshold is a sample point greater than 2000 and less than 25000, and the set double threshold is set. Obtaining the sample point data of the event peak, according to the characteristics of the time domain signal, respectively intercepting the first 100 sample points and the last 500 sample points of each event sample point, and effectively intercepting the event signal;
  • an event signal is acquired, and then the Meer frequency cepstral coefficient of each event signal is extracted, and the 12th-order Mel frequency cepstrum coefficient and the 1st order energy coefficient are extracted as characteristics of each event. .
  • the sound signal of the tooth occlusion is collected by the microphone sensor
  • the one type of model recognition algorithm is One-Class SVM, and the selected parameters are: -n 0.5, -s 2, -t 2, -g 0.0156, for training and identification of the model;
  • the confidence evaluation algorithm refers to calculating the confidence of each sample identification by calculating the maximum distance of each sample point to the hyperplane of the model, and then arranging the confidence from small to large by a small to large sorting algorithm.
  • the invention also provides an identification system based on a tooth occlusion sound, comprising:
  • a data collection module for collecting sound signals of tooth occlusion
  • a data processing module configured to process the received sound signal to extract sound features in the sound signal
  • the identification module is configured to put the sound feature into a type of model recognition algorithm of the machine learning algorithm, and perform data identification and judgment.
  • the identification system further comprises:
  • the data processing module includes:
  • the calculation module is configured to obtain, according to the time domain and the frequency domain characteristics of the signal, the distribution law of the sound signal energy in the frequency range;
  • the filtering module is configured to first use Butterworth bandpass filtering according to the obtained frequency range, and then perform a third-order adaptive threshold wavelet filtering on the signal after passing through the Butterworth filter, and after two noise filtering, effective Remove noise interference;
  • a detection module for framing and windowing the signal, and then using the double threshold peak detection method to detect two events of the toothed sound signal in the processed signal;
  • an extraction module configured to extract a Mel frequency cepstrum coefficient and an energy characteristic of the sound peak in the two events, thereby obtaining a sound feature.
  • the microphone sensor of the smart device is first detected to collect a sound event, and according to the signal distribution characteristics of the time domain and the frequency domain of the sound event, each sound signal is calculated.
  • the distribution pattern of energy in the frequency range, the distribution range of the frequency accompanying energy is obtained, and the set frequency range is 100HZ to 10000HZ;
  • the bandpass's Butterworth filter is first used, the frequency range of the bandpass is 100HZ to 10000HZ, and then the wavelet of the third-order adaptive threshold is applied to the signal that has passed through the Butterworth filter.
  • Filtering processing for a noise environment, the set peak threshold is 0.1, and the threshold set at the interval of two tooth occlusion events is within 20000 sample points;
  • the twice filtered signal is subjected to framing and windowing, and each framing of the signal is 10 milliseconds, and the coverage between the frame and the frame is 5 milliseconds.
  • the added window function is a Hamming window.
  • event detection is performed.
  • the double threshold peak detection method is adopted. First, according to the characteristics of the time domain and the frequency domain, the minimum peak threshold of an event is first set. 0.05, that is, only the signal sample points whose event peak value is greater than or equal to 0.05 are retained, and then the interval between the two event peaks is set.
  • the set threshold is a sample point larger than 2000 and less than 25000, and the event is acquired by setting the double threshold.
  • the peak sample point data according to the characteristics of the time domain signal, respectively intercepts the first 100 sample points and the last 500 sample points of each event sample point, and can effectively intercept the event signal;
  • an event signal is acquired, and then the Meer frequency cepstral coefficient of each event signal is extracted, and the 12th order Mel frequency cepstral coefficient and the 1st order energy coefficient are extracted as characteristics of each event.
  • the data collection module collecting a sound signal of a tooth occlusion through a microphone sensor
  • the one type of model recognition algorithm is One-Class SVM, and the selected parameters are: -n 0.5, -s 2, -t 2, -g 0.0156, for training and identification of the model;
  • the confidence evaluation algorithm refers to calculating the confidence of each sample identification by calculating the maximum distance of each sample point to the hyperplane of the model, and then arranging the confidence from small to large by a small to large sorting algorithm.
  • the invention has the beneficial effects that the invention does not need any hardware cost and is simple in system, and is convenient to use, and can accurately identify legitimate users and illegal users, and has good practicability on current smart devices, and does not affect the user itself at the same time. Good use of smart devices when data privacy is protected.
  • Figure 1 is a schematic diagram of the present invention.
  • Figure 2 is a block diagram of the confidence algorithm of the present invention.
  • FIG. 3 is a diagram of an identification strategy of a model of the present invention.
  • the present invention discloses a method for identifying a sound based on a tooth occlusion, comprising the following steps:
  • the user can bite his or her teeth several times at a suitable distance from the smart device before use, thereby collecting the sound signal of the tooth bite by using the microphone sensor in the smart device.
  • a suitable distance means that for a smart watch, the distance of the smart terminal from the tooth is 15-20 cm. Any bite of the teeth means that there is no limit to the position or strength of the teeth.
  • the data processing step includes:
  • the filtering step according to the obtained frequency range, firstly uses Butterworth bandpass filtering, and then the signal after passing through the Butterworth filter is further subjected to third-order adaptive threshold wavelet filtering, and after two noise filtering, effective
  • smart devices generally use a noise environment where the noise is not continuous.
  • the microphone sensor of the smart device is first detected to collect a sound event, and according to the signal distribution characteristics of the time domain and the frequency domain of the sound event, the energy of each sound signal is calculated under the frequency range.
  • the distribution chart shows the distribution range of the frequency accompanying energy, and the set frequency range is 100HZ to 10000HZ;
  • the bandpass's Butterworth filter is first used, the bandpass frequency range is 100 Hz to 10000 Hz, and then the third order adaptive threshold is applied to the signal that has passed through the Butterworth filter.
  • Wavelet filtering processing for the noise environment, the set peak threshold is 0.1, taking into account the transientity of each person's two teeth occlusion time, basically within 1 second, the threshold set at the interval of two tooth occlusion events is Within 20000 sample points, it can work in an appropriate amount of noise environment;
  • the twice filtered signal is subjected to framing and windowing, and each framing of the signal is 10 milliseconds, in order to avoid the sound feature spanning two frames, the adopted The coverage between the frame and the frame is 5 milliseconds, and the feature is effectively extracted.
  • the added window function is a Hamming window.
  • event detection is performed, and the double threshold peak detection method is adopted. First, according to the characteristics of the time domain and the frequency domain, first set the minimum peak threshold of an event to 0.05, that is, only the signal sample point whose event peak value is greater than or equal to 0.05, and then set the interval between the peaks of the two events.
  • the threshold is greater than 2000 and less than 25000 samples Point, through the set double threshold, obtain the sample point data of the event peak, according to the characteristics of the time domain signal, respectively intercept the first 100 sample points and the last 500 sample points of each event sample point, which can effectively intercept the event signal;
  • an event signal is acquired, and then the Meer frequency cepstral coefficient of each event signal is extracted, and the 12th-order Mel frequency cepstrum coefficient and the 1st order energy coefficient are extracted as characteristics of each event. .
  • the one type of model recognition algorithm is One-Class SVM.
  • the selected parameters are: -n 0.5, -s 2, -t 2, -g 0.0156, and the model is trained and identified.
  • the confidence evaluation algorithm uses the confidence evaluation algorithm to filter the predicted results and put them into the second-class recognition algorithm to improve the accuracy, such as: Two-Class SVM.
  • the invention also discloses an identification system based on tooth occlusion sound, comprising:
  • a data collection module for collecting sound signals of tooth occlusion
  • a data processing module configured to process the received sound signal to extract sound features in the sound signal
  • the identification module is configured to put the sound feature into a type of model recognition algorithm of the machine learning algorithm, and perform data identification and judgment.
  • the identification system further includes: by using a confidence evaluation algorithm, placing some samples with high confidence into the data set for training, thereby improving the accuracy of the recognition.
  • the data processing module includes:
  • the calculation module is configured to obtain, according to the time domain and the frequency domain characteristics of the signal, the distribution law of the sound signal energy in the frequency range;
  • the filtering module is configured to first use Butterworth bandpass filtering according to the obtained frequency range, and then perform a third-order adaptive threshold wavelet filtering on the signal after passing through the Butterworth filter, and after two noise filtering, effective Remove noise interference;
  • a detection module for framing and windowing the signal, and then using the double threshold peak detection method to detect two events of the toothed sound signal in the processed signal;
  • an extraction module configured to extract a Mel frequency cepstrum coefficient and an energy characteristic of the sound peak in the two events, thereby obtaining a sound feature.
  • the computing module first detecting a sound event collected by a microphone sensor of the smart device, According to the signal distribution characteristics of the time domain and the frequency domain of the sound event, the distribution pattern of the energy of each sound signal in the frequency range is calculated, and the distribution range of the frequency accompanying energy is obtained, and the set frequency range is 100HZ to 10000HZ;
  • the bandpass's Butterworth filter is first used, the frequency range of the bandpass is 100HZ to 10000HZ, and then the wavelet of the third-order adaptive threshold is applied to the signal that has passed through the Butterworth filter.
  • Filtering processing for a noise environment, the set peak threshold is 0.1, and the threshold set at the interval of two tooth occlusion events is within 20000 sample points;
  • the twice filtered signal is subjected to framing and windowing, and each framing of the signal is 10 milliseconds, and the coverage between the frame and the frame is 5 milliseconds.
  • the added window function is a Hamming window.
  • event detection is performed.
  • the double threshold peak detection method is adopted. First, according to the characteristics of the time domain and the frequency domain, the minimum peak threshold of an event is first set. 0.05, that is, only the signal sample points whose event peak value is greater than or equal to 0.05 are retained, and then the interval between the two event peaks is set.
  • the set threshold is a sample point larger than 2000 and less than 25000, and the event is acquired by setting the double threshold.
  • the peak sample point data according to the characteristics of the time domain signal, respectively intercepts the first 100 sample points and the last 500 sample points of each event sample point, and can effectively intercept the event signal;
  • an event signal is acquired, and then the Meer frequency cepstral coefficient of each event signal is extracted, and the 12th order Mel frequency cepstral coefficient and the 1st order energy coefficient are extracted as characteristics of each event.
  • the one type of model recognition algorithm is One-Class SVM, and the selected parameters are: -n 0.5, -s 2, -t 2, -g 0.0156, and the model is trained and recognized.
  • the confidence evaluation algorithm refers to calculating the confidence of each sample identification by calculating the maximum distance of each sample point to the hyperplane of the model, and then arranging the confidence from small to large by a small to large sorting algorithm.
  • One-Class SVM is a type of support vector machine.
  • the invention relates to a processing technology of a tooth occlusion sound signal, a sound event detection algorithm and an effective sound feature extraction and a high-precision machine learning algorithm technology, and provides a safe identification and authentication mode for the user to realize data privacy protection of the user smart device.
  • the invention does not need any hardware cost and is simple in system, and is convenient to use, and can accurately identify legal users and illegal users. It has good practicability on the current smart devices, and does not affect the users themselves in the data privacy protection against the smart devices. Good use.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Dental Tools And Instruments Or Auxiliary Dental Instruments (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

一种基于牙齿咬合声音的识别方法及系统，该识别方法包括如下步骤：S1.数据收集步骤，收集牙齿咬合的声音信号；S2.数据处理步骤，对收到的声音信号进行处理，提取声音信号中的声音特征；S3.识别步骤，将声音特征放入机器学习算法的一类模型识别算法中，进行数据的识别和判断。该方法无需任何硬件成本且系统简单，且使用方便，能够准确识别合法用户和非法用户，对于目前智能设备上，有很好的实用性，且同时不影响用户本身在数据隐私保护时对智能设备的良好使用。

Description

基于牙齿咬合声音的识别方法及系统 技术领域
本发明涉及声音识别技术领域，尤其涉及基于牙齿咬合声音的识别方法及系统。
背景技术
随着智能化社会进程的不断推进，新兴智能设备的不断发展，智能设备在日常生活中显得越来越重要了，它能够随时给我们的生活提供便利，比如：智能手表能给我们提供来电提醒，通过智能手表来随时查看信息等，头戴脑电记录仪，能无时无刻的记录我们每个人的脑电的生物体征，通过一系列算法的分析和研究，为我们每个人的健康保驾护航，大大为我们的生活质量提供了强有力的保障。我们在使用智能设备的便利的同时，也伴随着很多严峻的考验，用户的个人隐私和一系列的个人数据的泄露显得尤为突出，那么怎么在大数据时代，保护好个人的数据显得尤为重要。
传统的保护手段有如下几个：在智能手机上，我们常常使用PIN码、密码或者图案和指纹识别保护。还有一些使用指纹识别，静脉识别、虹膜识别、人脸识别等技术。
首先我们比较常用的PIN码、密码和图案保护，在我们可穿戴设备上是不可行的，我们都知道可穿戴设备是穿戴在用户身上的，也就是智能设备的面积是非常的小，这些都是不能够在其上进行安装来达到身份的识别和保护。所以我们常用的保护方式在智能上是行不通的。
然后是一些识别精度比较高的生物识别技术，在识别精度高的同时，也要付出一些应有的代价。最常用的指纹识别，我们都知道在一些智能设备和一些门禁系统上，很多都配置了指纹识别技术，当我们手指有汗液、污渍、或者脱皮等情况下，指纹识别就容易识别失败，必须清理这些东西才能够重新进行识别，达到身份认证的效果，大大给人们造成了很大的不便，与此同时，我们很容易在很多地方留下我们的指纹，那么也就很容易被不法分子仿造，然后达到识别的目的。目前有些厂家在其智能手机上推出了超声波指纹识别，其避免了由于汗液，污渍等不能识别的情况，很高精度的识别的用户，但是其成本昂贵，给智能设备的生产提供了昂贵的代价，这是很多人不愿意接受的。
静脉识别和虹膜识别也都根据每个人的生理特征不一样然后采取生理特征信号，然后对信号进行一系列的分析和处理，最终达到识别的作用，这两种生物识别技术能达到很高的识别精度，但是也伴随着一些弊端，静脉识别里手背的静脉仍有可能随着年龄和生理的变化而变化，永久性尚未得到证实，由于采集方式受自身特点的限制，产品难以小型化，不能够在智能设备上进行量产，在同时，采集设备有特殊要求，设计相对复杂，制造成本较高。
人脸识别也提供了很高的精度，人脸识别存在很大的漏洞，只需要一个强大的修改图片的软件，我们就能够很容易的模仿人脸上的生理特征，然后进行人脸识别，实验结果也是很震惊，通过简单的几步处理，就能够正确的识别，保护级别不是特别高。
发明内容
本发明提供了一种基于牙齿咬合声音的识别方法，包括如下步骤：
S1.数据收集步骤，收集牙齿咬合的声音信号；
S2.数据处理步骤，对收到的声音信号进行处理，提取声音信号中的声音特征；
S3.识别步骤，将声音特征放入机器学习算法的一类模型识别算法中，进行数据的识别和判断。
作为本发明的进一步改进，该识别方法还包括：
S4.通过置信度评估算法，将一些具有很高置信度的样本放入到用来训练的数据集中，从而提高识别的精度。
作为本发明的进一步改进，在所述数据处理步骤中包括：
S21.计算步骤，针对采集到的声音信号，根据信号的时域和频域的特点，得出声音信号能量在频率范围内的分布规律；
S22.滤波步骤，根据得到的频率范围，首先使用巴特沃斯带通滤波，然后把经过巴特沃斯滤波器后的信号再进行三阶自适应阈值小波滤波，经过两次噪声的滤波，有效的去除噪声的干扰；
S23.检测步骤，对信号进行分帧和加窗处理，然后对处理过后的信号采用双阈值峰值检测法来检测咬牙在声音信号中的两个事件；
S24.提取步骤，在所述两个事件中，提取声音峰值的梅尔频率倒谱系数和能量特征，从而得到声音特征。
作为本发明的进一步改进，在所述S21.计算步骤中，先检测智能设备的麦克风传感器采集到声音事件，根据所述声音事件的时域和频域的信号 分布特征，然后计算出每个声音信号的能量在频率范围下的分布走势图，得到频率伴随能量的分布范围，设定的频率范围为100HZ到10000HZ；
在所述S22.滤波步骤中，首先使用的是带通的巴特沃斯滤波器，带通的频率范围是100HZ到10000HZ，然后对已经通过巴特沃斯滤波器后的信号进行三阶自适应阈值的小波滤波处理，针对在噪声环境里，设定的峰值阈值是0.1，在两个牙齿咬合事件的间隔设置的阈值是20000样本点以内；
在所述S23.检测步骤中，对所述经过两次滤波后的信号进行分帧和加窗处理，对信号的每个分帧为10毫秒，所采用的是帧与帧之间的覆盖是5毫秒，所加的窗函数是汉明窗，对处理后的数据，进行事件检测，所采用的是双阈值峰值检测法，首先根据时域和频域的特征，先设定一个事件的最小峰值阈值为0.05，也就是只保留事件峰值大于等于0.05的信号样本点，然后设定两个事件峰值之间的间隔，设定的阈值为大于2000小于25000的样本点，通过设置的双阈值，获取事件峰值的样本点数据，根据时域信号的特点，分别截取每个事件样本点的前100个样本点和后500个样本点，能有效的截取事件信号；
在所述S24.提取步骤中，获取事件信号，然后提取每个事件信号梅尔频率倒谱系数，提取的是12阶的梅尔频率倒谱系数和1阶的能量系数作为每个事件的特征。
作为本发明的进一步改进，
在所述数据收集步骤中，通过麦克风传感器收集牙齿咬合的声音信号；
所述一类模型识别算法为One-Class SVM，所选的参数是：-n 0.5，-s 2，-t 2，-g 0.0156，进行模型的训练和识别；
所述置信度评估算法是指：通过计算每个样本点到模型的超平面的最大距离，来计算每个样本识别的置信度，然后通过一个从小到大的排序算法把置信度从小到大排列，根据拒识率e来选取一部分的样本进入二类识别算法来进行高精度识别，在有序的置信度序列里选取大于e*n+1的样本，拒识率设为e＝0.5％。
本发明还提供了一种基于牙齿咬合声音的识别系统，包括：
数据收集模块，用于收集牙齿咬合的声音信号；
数据处理模块，用于对收到的声音信号进行处理，提取声音信号中的声音特征；
识别模块，用于将声音特征放入机器学习算法的一类模型识别算法中，进行数据的识别和判断。
作为本发明的进一步改进，该识别系统还包括：
通过置信度评估算法，将一些具有很高置信度的样本放入到用来训练的数据集中，从而提高识别的精度。
作为本发明的进一步改进，在所述数据处理模块中包括：
计算模块，用于针对采集到的声音信号，根据信号的时域和频域的特点，得出声音信号能量在频率范围内的分布规律；
滤波模块，用于根据得到的频率范围，首先使用巴特沃斯带通滤波，然后把经过巴特沃斯滤波器后的信号再进行三阶自适应阈值小波滤波，经过两次噪声的滤波，有效的去除噪声的干扰；
检测模块，用于对信号进行分帧和加窗处理，然后对处理过后的信号采用双阈值峰值检测法来检测咬牙在声音信号中的两个事件；
提取模块，用于在所述两个事件中，提取声音峰值的梅尔频率倒谱系数和能量特征，从而得到声音特征。
作为本发明的进一步改进，在所述计算模块中，先检测智能设备的麦克风传感器采集到声音事件，根据所述声音事件的时域和频域的信号分布特征，然后计算出每个声音信号的能量在频率范围下的分布走势图，得到频率伴随能量的分布范围，设定的频率范围为100HZ到10000HZ；
在所述滤波模块中，首先使用的是带通的巴特沃斯滤波器，带通的频率范围是100HZ到10000HZ，然后对已经通过巴特沃斯滤波器后的信号进行三阶自适应阈值的小波滤波处理，针对在噪声环境里，设定的峰值阈值是0.1，在两个牙齿咬合事件的间隔设置的阈值是20000样本点以内；
在所述检测模块中，对所述经过两次滤波后的信号进行分帧和加窗处理，对信号的每个分帧为10毫秒，所采用的是帧与帧之间的覆盖是5毫秒，所加的窗函数是汉明窗，对处理后的数据，进行事件检测，所采用的是双阈值峰值检测法，首先根据时域和频域的特征，先设定一个事件的最小峰值阈值为0.05，也就是只保留事件峰值大于等于0.05的信号样本点，然后设定两个事件峰值之间的间隔，设定的阈值为大于2000小于25000的样本点，通过设置的双阈值，获取事件峰值的样本点数据，根据时域信号的特点，分别截取每个事件样本点的前100个样本点和后500个样本点，能有效的截取事件信号；
在所述提取模块中，获取事件信号，然后提取每个事件信号梅尔频率倒谱系数，提取的是12阶的梅尔频率倒谱系数和1阶的能量系数作为每个事件的特征。
作为本发明的进一步改进，
在所述数据收集模块中，通过麦克风传感器收集牙齿咬合的声音信号；
所述一类模型识别算法为One-Class SVM，所选的参数是：-n 0.5，-s 2，-t 2，-g 0.0156，进行模型的训练和识别；
所述置信度评估算法是指：通过计算每个样本点到模型的超平面的最大距离，来计算每个样本识别的置信度，然后通过一个从小到大的排序算法把置信度从小到大排列，根据拒识率e来选取一部分的样本进入二类识别算法来进行高精度识别，在有序的置信度序列里选取大于e*n+1的样本，拒识率设为e＝0.5％。
本发明的有益效果是：本发明无需任何硬件成本且系统简单，且使用方便，能够准确识别合法用户和非法用户，对于目前智能设备上，有很好的实用性，且同时不影响用户本身在数据隐私保护时对智能设备的良好使用。
附图说明
图1是本发明的原理图。
图2是本发明的置信度算法框架图。
图3是本发明的模型的识别策略图。
具体实施方式
如图1-3所示，本发明公开了一种基于牙齿咬合声音的识别方法，包括如下步骤：
S1.数据收集步骤，收集牙齿咬合的声音信号；
S2.数据处理步骤，对收到的声音信号进行处理，提取声音信号中的声音特征；
S3.识别步骤，将声音特征放入机器学习算法的一类模型识别算法中，进行数据的识别和判断；
S4.由于把用户的数据放入一类识别算法模型中识别的精度可能不是很高，我们采用了模型的置信度评估算法，将一些具有很高置信度的样本放入到用来训练的数据集中，把原来的一类识别模型升级为二类识别模型，从而提高识别的精度。
用户在使用智能设备时，在使用前，距离智能设备合适的距离，任意咬合自己的牙齿若干次，从而利用智能设备中的麦克风传感器收集牙齿咬合的声音信号。合适的距离是指：对于智能手表，智能终端距离牙齿的距离在15-20cm。任意咬合牙齿是指：对咬合牙齿没有位置、力度的限制。
作为本发明的实施例，在所述数据处理步骤中包括：
S21.计算步骤，针对采集到的声音信号，根据信号的时域和频域的特点，得出声音信号能量在频率范围内的分布规律，即：能量主要分布在哪个频率范围内；
S22.滤波步骤，根据得到的频率范围，首先使用巴特沃斯带通滤波，然后把经过巴特沃斯滤波器后的信号再进行三阶自适应阈值小波滤波，经过两次噪声的滤波，有效的去除噪声的干扰，智能设备一般使用场景是噪声不是连续大的噪声环境，我们使用双阈值过滤法，在事件峰值的大小设定一个阈值，然后在两个事件峰值之间设置一个阈值，通过这两个阈值，可以在噪声不是持续较大的环境里，能够有效的得到牙齿咬合事件，就能够使其有效的工作；
S23.检测步骤，对信号进行分帧和加窗处理，然后对处理过后的信号采用双阈值峰值检测法来检测咬牙在声音信号中的两个事件；
S24.提取步骤，在所述两个事件中，提取声音峰值的梅尔频率倒谱系数和能量特征，从而得到声音特征。
作为本发明的优选实施例：
在所述S21.计算步骤中，先检测智能设备的麦克风传感器采集到声音事件，根据所述声音事件的时域和频域的信号分布特征，然后计算出每个声音信号的能量在频率范围下的分布走势图，得到频率伴随能量的分布范围，设定的频率范围为100HZ到10000HZ；
在所述S22.滤波步骤中，首先使用的是带通的巴特沃斯滤波器，带通的频率范围是100HZ到10000HZ，然后对已经通过巴特沃斯滤波器后的信号进行三阶自适应阈值的小波滤波处理，针对在噪声环境里，设定的峰值阈值是0.1，考虑到每个人两次牙齿咬合时间的短暂性，基本在1秒以内，在两个牙齿咬合事件的间隔设置的阈值是20000样本点以内，可以在适量的噪声环境下使其工作；
在所述S23.检测步骤中，对所述经过两次滤波后的信号进行分帧和加窗处理，对信号的每个分帧为10毫秒，为了避免声音特征在跨越两帧，所采用的是帧与帧之间的覆盖是5毫秒，达到特征的有效提取，所加的窗函数是汉明窗(hamming)，对处理后的数据，进行事件检测，所采用的是双阈值峰值检测法，首先根据时域和频域的特征，先设定一个事件的最小峰值阈值为0.05，也就是只保留事件峰值大于等于0.05的信号样本点，然后设定两个事件峰值之间的间隔，设定的阈值为大于2000小于25000的样本 点，通过设置的双阈值，获取事件峰值的样本点数据，根据时域信号的特点，分别截取每个事件样本点的前100个样本点和后500个样本点，能有效的截取事件信号；
在所述S24.提取步骤中，获取事件信号，然后提取每个事件信号梅尔频率倒谱系数，提取的是12阶的梅尔频率倒谱系数和1阶的能量系数作为每个事件的特征。
所述一类模型识别算法为One-Class SVM，在一类识别算法模型中，所选的参数是：-n 0.5，-s 2，-t 2，-g 0.0156，进行模型的训练和识别。
使用置信度评估算法对预测的结果进行筛选，放入二类识别算法中提高精度，比如：Two-Class SVM。我们采用的置信度评估算法是指：通过计算每个样本点到模型的超平面的最大距离，来计算每个样本识别的置信度，然后通过一个从小到大的排序算法把置信度从小到大排列，根据拒识率e来选取一部分的样本进入二类识别算法来进行高精度识别，在有序的置信度序列里选取大于e*n+1的样本，这里我们的拒识率设为e＝0.5％。
本发明还公开了一种基于牙齿咬合声音的识别系统，包括：
数据收集模块，用于收集牙齿咬合的声音信号；
数据处理模块，用于对收到的声音信号进行处理，提取声音信号中的声音特征；
识别模块，用于将声音特征放入机器学习算法的一类模型识别算法中，进行数据的识别和判断。
该识别系统还包括：通过置信度评估算法，将一些具有很高置信度的样本放入到用来训练的数据集中，从而提高识别的精度。
在所述数据处理模块中包括：
计算模块，用于针对采集到的声音信号，根据信号的时域和频域的特点，得出声音信号能量在频率范围内的分布规律；
滤波模块，用于根据得到的频率范围，首先使用巴特沃斯带通滤波，然后把经过巴特沃斯滤波器后的信号再进行三阶自适应阈值小波滤波，经过两次噪声的滤波，有效的去除噪声的干扰；
检测模块，用于对信号进行分帧和加窗处理，然后对处理过后的信号采用双阈值峰值检测法来检测咬牙在声音信号中的两个事件；
提取模块，用于在所述两个事件中，提取声音峰值的梅尔频率倒谱系数和能量特征，从而得到声音特征。
在所述计算模块中，先检测智能设备的麦克风传感器采集到声音事件， 根据所述声音事件的时域和频域的信号分布特征，然后计算出每个声音信号的能量在频率范围下的分布走势图，得到频率伴随能量的分布范围，设定的频率范围为100HZ到10000HZ；
在所述滤波模块中，首先使用的是带通的巴特沃斯滤波器，带通的频率范围是100HZ到10000HZ，然后对已经通过巴特沃斯滤波器后的信号进行三阶自适应阈值的小波滤波处理，针对在噪声环境里，设定的峰值阈值是0.1，在两个牙齿咬合事件的间隔设置的阈值是20000样本点以内；
在所述检测模块中，对所述经过两次滤波后的信号进行分帧和加窗处理，对信号的每个分帧为10毫秒，所采用的是帧与帧之间的覆盖是5毫秒，所加的窗函数是汉明窗，对处理后的数据，进行事件检测，所采用的是双阈值峰值检测法，首先根据时域和频域的特征，先设定一个事件的最小峰值阈值为0.05，也就是只保留事件峰值大于等于0.05的信号样本点，然后设定两个事件峰值之间的间隔，设定的阈值为大于2000小于25000的样本点，通过设置的双阈值，获取事件峰值的样本点数据，根据时域信号的特点，分别截取每个事件样本点的前100个样本点和后500个样本点，能有效的截取事件信号；
在所述提取模块中，获取事件信号，然后提取每个事件信号梅尔频率倒谱系数，提取的是12阶的梅尔频率倒谱系数和1阶的能量系数作为每个事件的特征。
所述一类模型识别算法为One-Class SVM，所选的参数是：-n 0.5，-s 2，-t 2，-g 0.0156，进行模型的训练和识别。
所述置信度评估算法是指：通过计算每个样本点到模型的超平面的最大距离，来计算每个样本识别的置信度，然后通过一个从小到大的排序算法把置信度从小到大排列，根据拒识率e来选取一部分的样本进入二类识别算法来进行高精度识别，在有序的置信度序列里选取大于e*n+1的样本，拒识率设为e＝0.5％。
One-Class SVM是一类支持向量机。
本发明涉及到牙齿咬合声音信号的处理技术，声音事件检测算法和有效声音特征提取以及高精度的机器学习算法技术，给用户提供一个安全的识别认证方式，实现用户智能设备数据隐私保护。
本发明无需任何硬件成本且系统简单，且使用方便，能够准确识别合法用户和非法用户，对于目前智能设备上，有很好的实用性，且同时不影响用户本身在数据隐私保护时对智能设备的良好使用。
以上内容是结合具体的优选实施方式对本发明所作的进一步详细说明，不能认定本发明的具体实施只局限于这些说明。对于本发明所属技术领域的普通技术人员来说，在不脱离本发明构思的前提下，还可以做出若干简单推演或替换，都应当视为属于本发明的保护范围。

Claims (10)

  1. 一种基于牙齿咬合声音的识别方法，其特征在于，包括如下步骤：
    S1.数据收集步骤，收集牙齿咬合的声音信号；
    S2.数据处理步骤，对收到的声音信号进行处理，提取声音信号中的声音特征；
    S3.识别步骤，将声音特征放入机器学习算法的一类模型识别算法中，进行数据的识别和判断。
  2. 根据权利要求1所述的识别方法，其特征在于，该识别方法还包括：
    S4.通过置信度评估算法，将一些具有很高置信度的样本放入到用来训练的数据集中，从而提高识别的精度。
  3. 根据权利要求1所述的识别方法，其特征在于，在所述数据处理步骤中包括：
    S21.计算步骤，针对采集到的声音信号，根据信号的时域和频域的特点，得出声音信号能量在频率范围内的分布规律；
    S22.滤波步骤，根据得到的频率范围，首先使用巴特沃斯带通滤波，然后把经过巴特沃斯滤波器后的信号再进行三阶自适应阈值小波滤波，经过两次噪声的滤波，有效的去除噪声的干扰；
    S23.检测步骤，对信号进行分帧和加窗处理，然后对处理过后的信号采用双阈值峰值检测法来检测咬牙在声音信号中的两个事件；
    S24.提取步骤，在所述两个事件中，提取声音峰值的梅尔频率倒谱系数和能量特征，从而得到声音特征。
  4. 根据权利要求3所述的识别方法，其特征在于，
    在所述S21.计算步骤中，先检测智能设备的麦克风传感器采集到声音事件，根据所述声音事件的时域和频域的信号分布特征，然后计算出每个声音信号的能量在频率范围下的分布走势图，得到频率伴随能量的分布范围，设定的频率范围为100HZ到10000HZ；
    在所述S22.滤波步骤中，首先使用的是带通的巴特沃斯滤波器，带通的频率范围是100HZ到10000HZ，然后对已经通过巴特沃斯滤波器后的信号进行三阶自适应阈值的小波滤波处理，针对在噪声环境里，设定的峰值阈值 是0.1，在两个牙齿咬合事件的间隔设置的阈值是20000样本点以内；
    在所述S23.检测步骤中，对所述经过两次滤波后的信号进行分帧和加窗处理对信号的每个分帧为10毫秒，所采用的是帧与帧之间的覆盖是5毫秒，所加的窗函数是汉明窗，对处理后的数据，进行事件检测，所采用的是双阈值峰值检测法，首先根据时域和频域的特征，先设定一个事件的最小峰值阈值为0.05，也就是只保留事件峰值大于等于0.05的信号样本点，然后设定两个事件峰值之间的间隔，设定的阈值为大于2000小于25000的样本点，通过设置的双阈值，获取事件峰值的样本点数据，根据时域信号的特点，分别截取每个事件样本点的前100个样本点和后500个样本点，能有效的截取事件信号；
    在所述S24.提取步骤中，获取事件信号，然后提取每个事件信号梅尔频率倒谱系数，提取的是12阶的梅尔频率倒谱系数和1阶的能量系数作为每个事件的特征。
  5. 根据权利要求2至4任一项所述的识别方法，其特征在于，
    在所述数据收集步骤中，通过麦克风传感器收集牙齿咬合的声音信号；所述一类模型识别算法为One-Class SVM，所选的参数是：-n 0.5，-s 2，-t 2，-g0.0156，进行模型的训练和识别；
    所述置信度评估算法是指：通过计算每个样本点到模型的超平面的最大距离，来计算每个样本识别的置信度，然后通过一个从小到大的排序算法把置信度从小到大排列，根据拒识率e来选取一部分的样本进入二类识别算法来进行高精度识别，在有序的置信度序列里选取大于e*n+1的样本，拒识率设为e＝0.5％。
  6. 一种基于牙齿咬合声音的识别系统，其特征在于，包括：
    数据收集模块，用于收集牙齿咬合的声音信号；
    数据处理模块，用于对收到的声音信号进行处理，提取声音信号中的声音特征；
    识别模块，用于将声音特征放入机器学习算法的一类模型识别算法中，进行数据的识别和判断。
  7. 根据权利要求6所述的识别系统，其特征在于，该识别系统还包括：
    通过置信度评估算法，将一些具有很高置信度的样本放入到用来训练的数据集中，从而提高识别的精度。
  8. 根据权利要求6所述的识别系统，其特征在于，在所述数据处理模块中包括：
    计算模块，用于针对采集到的声音信号，根据信号的时域和频域的特点，得出声音信号能量在频率范围内的分布规律；
    滤波模块，用于根据得到的频率范围，首先使用巴特沃斯带通滤波，然后把经过巴特沃斯滤波器后的信号再进行三阶自适应阈值小波滤波，经过两次噪声的滤波，有效的去除噪声的干扰；
    检测模块，用于对信号进行分帧和加窗处理，然后对处理过后的信号采用双阈值峰值检测法来检测咬牙在声音信号中的两个事件；
    提取模块，用于在所述两个事件中，提取声音峰值的梅尔频率倒谱系数和能量特征，从而得到声音特征。
  9. 根据权利要求8所述的识别系统，其特征在于，
    在所述计算模块中，先检测智能设备的麦克风传感器采集到声音事件，根据所述声音事件的时域和频域的信号分布特征，然后计算出每个声音信号的能量在频率范围下的分布走势图，得到频率伴随能量的分布范围，设定的频率范围为100HZ到10000HZ；
    在所述滤波模块中，首先使用的是带通的巴特沃斯滤波器，带通的频率范围是100HZ到10000HZ，然后对已经通过巴特沃斯滤波器后的信号进行三阶自适应阈值的小波滤波处理，针对在噪声环境里，设定的峰值阈值是0.1，在两个牙齿咬合事件的间隔设置的阈值是20000样本点以内；
    在所述检测模块中，对所述经过两次滤波后的信号进行分帧和加窗处理，对信号的每个分帧为10毫秒，所采用的是帧与帧之间的覆盖是5毫秒，所加的窗函数是汉明窗，对处理后的数据，进行事件检测，所采用的是双阈值峰值检测法，首先根据时域和频域的特征，先设定一个事件的最小峰值阈值为0.05，也就是只保留事件峰值大于等于0.05的信号样本点，然后设定两个事件峰值之间的间隔，设定的阈值为大于2000小于25000的样本点，通过设置的双阈值，获取事件峰值的样本点数据，根据时域信号的特点，分别截取每个事件样本点的前100个样本点和后500个样本点，能有效的 截取事件信号；
    在所述提取模块中，获取事件信号，然后提取每个事件信号梅尔频率倒谱系数，提取的是12阶的梅尔频率倒谱系数和1阶的能量系数作为每个事件的特征。
  10. 根据权利要求7至9任一项所述的识别系统，其特征在于，
    在所述数据收集模块中，通过麦克风传感器收集牙齿咬合的声音信号；所述一类模型识别算法为One-Class SVM，所选的参数是：-n 0.5，-s 2，-t 2，-g0.0156，进行模型的训练和识别；
    所述置信度评估算法是指：通过计算每个样本点到模型的超平面的最大距离，来计算每个样本识别的置信度，然后通过一个从小到大的排序算法把置信度从小到大排列，根据拒识率e来选取一部分的样本进入二类识别算法来进行高精度识别，在有序的置信度序列里选取大于e*n+1的样本，拒识率设为e＝0.5％。
PCT/CN2017/092771 2017-05-27 2017-07-13 基于牙齿咬合声音的识别方法及系统 WO2018223491A1 (zh)

Applications Claiming Priority (3)

Application Number Priority Date Filing Date Title
CN201710392749 2017-05-27
CN201710433172.1A CN107293301B (zh) 2017-05-27 2017-06-09 基于牙齿咬合声音的识别方法及系统
CN201710433172.1 2017-06-09

Publications (1)

Publication Number Publication Date
WO2018223491A1 true WO2018223491A1 (zh) 2018-12-13

Family

ID=60097489

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/092771 WO2018223491A1 (zh) 2017-05-27 2017-07-13 基于牙齿咬合声音的识别方法及系统

Country Status (2)

Country Link
CN (1) CN107293301B (zh)
WO (1) WO2018223491A1 (zh)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108175436A (zh) * 2017-12-28 2018-06-19 北京航空航天大学 一种肠鸣音智能自动识别方法
CN109740523B (zh) * 2018-12-29 2020-12-29 国网陕西省电力公司电力科学研究院 一种基于声学特征和神经网络的电力变压器故障诊断方法
CN111832332A (zh) * 2019-03-29 2020-10-27 中国石油天然气集团有限公司 泥浆脉冲信号处理方法及装置
CN110321689A (zh) * 2019-07-08 2019-10-11 深圳大学 一种基于响指的身份识别方法和系统
CN110751044B (zh) * 2019-09-19 2022-07-29 杭州电子科技大学 基于深度网络迁移特征与增广自编码的城市噪声识别方法
CN115348049A (zh) * 2022-06-22 2022-11-15 北京理工大学 一种利用耳机内向麦克风的用户身份认证方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5412738A (en) * 1992-08-11 1995-05-02 Istituto Trentino Di Cultura Recognition system, particularly for recognising people
CN104866575A (zh) * 2015-05-25 2015-08-26 深圳创维－Rgb电子有限公司 一种基于智能设备的磨牙检测方法及检测系统
CN105242790A (zh) * 2015-11-04 2016-01-13 钱雄 一种智能设备控制方法及控制装置
CN105988768A (zh) * 2015-02-06 2016-10-05 电信科学技术研究院 智能设备控制方法、信号获取方法及相关设备

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6820056B1 (en) * 2000-11-21 2004-11-16 International Business Machines Corporation Recognizing non-verbal sound commands in an interactive computer controlled speech word recognition display system
CN1512490A (zh) * 2002-12-30 2004-07-14 吕小麟 一种输入控制信号的方法及装置
GB2466242B (en) * 2008-12-15 2013-01-02 Audio Analytic Ltd Sound identification systems
CN103531199B (zh) * 2013-10-11 2016-03-09 福州大学 基于快速稀疏分解和深度学习的生态声音识别方法
WO2015196353A1 (zh) * 2014-06-24 2015-12-30 陈梓铭 一种将牙齿产生的声音信号转换为电信号的控制系统及方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5412738A (en) * 1992-08-11 1995-05-02 Istituto Trentino Di Cultura Recognition system, particularly for recognising people
CN105988768A (zh) * 2015-02-06 2016-10-05 电信科学技术研究院 智能设备控制方法、信号获取方法及相关设备
CN104866575A (zh) * 2015-05-25 2015-08-26 深圳创维－Rgb电子有限公司 一种基于智能设备的磨牙检测方法及检测系统
CN105242790A (zh) * 2015-11-04 2016-01-13 钱雄 一种智能设备控制方法及控制装置

Also Published As

Publication number Publication date
CN107293301A (zh) 2017-10-24
CN107293301B (zh) 2019-04-02

Similar Documents

Publication Publication Date Title
WO2018223491A1 (zh) 基于牙齿咬合声音的识别方法及系统
WO2019192235A1 (zh) 一种基于移动设备的用户身份认证方法及系统
Kang et al. ECG authentication system design based on signal analysis in mobile and wearable devices
CN106446831B (zh) 一种人脸识别方法及装置
Li et al. Generalized face anti-spoofing by detecting pulse from face videos
Gomez-Barrero et al. Biometrics in the era of COVID-19: challenges and opportunities
WO2019109433A1 (zh) 基于步态识别的身份认证方法及装置
Cao et al. PPGPass: Nonintrusive and secure mobile two-factor authentication via wearables
US8232866B2 (en) Systems and methods for remote long standoff biometric identification using microwave cardiac signals
CN110069199B (zh) 一种基于智能手表的皮肤式手指手势识别方法
CN111178155B (zh) 基于惯性传感器的步态特征提取和步态识别方法
Chun et al. ECG based user authentication for wearable devices using short time Fourier transform
WO2008144174A1 (en) Methods for using a biometric parameter in the identification of persons
Liu et al. Recent advances in biometrics-based user authentication for wearable devices: A contemporary survey
Hestbek et al. Biometric gait recognition for mobile devices using wavelet transform and support vector machines
Ohtsuki et al. Biometrie authentication using hand movement information from wrist-worn PPG sensors
El-Bendary et al. HSAS: Heart sound authentication system
CN109858366A (zh) 身份认证方法及装置
Bastos et al. Double authentication model based on ppg and ecg signals
CN112347450A (zh) 一种基于眨眼声音信号的身份验证方法
CN106971203B (zh) 基于走路特征数据的身份识别方法
Ehatisham-ul-Haq et al. Identifying users with wearable sensors based on activity patterns
CN111444489B (zh) 一种基于光电容积脉搏波传感器的双因子认证方法
CN115935314A (zh) 一种基于可穿戴设备运动传感器的用户身份认证方法
CN112949403B (zh) 一种基于下颌骨生物特征的可靠用户认证方法及系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17912498

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 12/03/2020)

122 Ep: pct application non-entry in european phase

Ref document number: 17912498

Country of ref document: EP

Kind code of ref document: A1