WO2013159282A1

WO2013159282A1 - Customized self-learning identification system and method

Info

Publication number: WO2013159282A1
Application number: PCT/CN2012/074584
Authority: WO
Inventors: 陈澎
Original assignee: 北京英福生科技有限公司
Priority date: 2012-04-24
Filing date: 2012-04-24
Publication date: 2013-10-31

Abstract

Provided are a customized self-learning identification system and method, the system comprising: one or more sensors, used to detect the signal of an object to be detected; a storage unit, used to store a sample/model library comprising a model and a training sample set; a processing unit, used to receive the signal of the detected object detected by the one or more sensors and identify the status type of the detected object according to a corresponding model in the sample/model library; an input/output device, used to output an identification result and receive feedback information inputted on the basis of the identification result; the processing unit also uses a corresponding training sample set to train a model and update the sample/model library according to the feedback information. The customized self-learning identification method of the present invention enables a user to establish a customized model, thus greatly improving the identification rate for customized status type.

Description

Personalized self-learning recognition system and method

Technical field

The invention belongs to the field of pattern recognition, and in particular relates to a personalized self-learning recognition system and method.

Background technique

Currently, pattern recognition technology has been used to identify various types of physiological state information of the human body for health monitoring and medical diagnosis.

Specifically, various types of sensors can be used to detect physiological parameters of the human body and the like; for example, motion sensors, electrocardiogram sensors, electromyogram sensors, electroencephalogram sensors, blood oxygen sensors, and the like can be used for various actions and electrocardiograms of the human body. , EMG, EEG, blood oxygen signal, etc. are detected; then the detected signal is preprocessed, and feature extraction and selection are performed, and then the recognition algorithm is used to classify and recognize according to the previously trained model. By using this technique to record the types of physiological states of various human bodies, it is possible to analyze the physiological health status of the human body. A large number of documents document the technical principles of pattern recognition. The background technology part of the Chinese patent (CN200710118156.X) specification details the principle and application of pattern recognition technology; the non-patent literature "Leo" Breina (1996), Bagging Predictor, Marchine Learning, Machine learning (training) principles and algorithms are described in 24(2): 123-140"; Non-patent literature "Upper Limb Motion Recognition System Based on 3D Accelerometer", Journal of Transducer Technology In the sixth issue of 2010, the process of processing and modeling the acceleration data by the acceleration sensor-based upper limb motion recognition system was proposed.

However, the existing pattern recognition technology is based on the classification and recognition of models established by collecting and training physiological state signals of various people, and the models are not collected and trained according to the physiological state signals of the individual users. . Due to individual differences, many existing health monitoring systems and tools do not accurately identify the physiological health status of a particular user.

Summary of the invention

The object of the present invention is to provide a personalized self-learning recognition system and method, which can collect and perform self-learning training on personalized data of a specific object, thereby greatly improving the recognition rate of the state of a specific object.

The technical solution of the present invention is specifically a personalized self-learning recognition system, including

One or more sensors for detecting a signal of the detected object;

a storage unit for storing a sample/model library, including a model and a training sample set;

a processing unit, configured to receive a signal of the detected object detected by the one or more sensors, and identify a state type of the detected object according to a corresponding model in the sample/model library;

Input/output means for outputting a recognition result and receiving feedback information input based on the recognition result;

The processing unit also trains the model using the corresponding training sample set and updates the sample/model library based on the feedback information.

among them,

The set of training samples includes characteristics of signals and/or signals of the detected object.

among them,

The sample/model library includes corresponding models in one or more scenarios.

among them,

The feedback information includes an identification of a model to be established.

among them,

The feedback information includes a scene identifier of a model to be established and a corresponding model identifier.

among them,

When the input identifier of the model to be established is the same as the identifier of the model in the sample/model library, the processing unit adds the corresponding training sample to the training sample set of the corresponding model for training to establish a corresponding to the identifier. model.

among them,

The processing unit is further configured to select to communicate with the one or more sensors according to a selection instruction received by the input/output device.

among them,

The input/output device is also operative to receive an input selected mode of operation.

among them,

The operation mode includes one of a training model mode, a monitoring mode, a recording mode, or a combination thereof;

The processing unit separately performs model training, state recognition, and stores the recognition result and/or the corresponding detection signal into the storage unit according to the input operation mode.

The present invention further provides a personalized self-learning recognition method, including

Using one or more sensors to detect the signal of the detected object;

Receiving the detected signal of the detected object, identifying a state type of the detected object according to a corresponding model stored in the sample/model library, and outputting the recognition result;

The input feedback information is received, and the model is trained based on the feedback information and the sample/model library is updated.

among them,

When the input identifier of the model to be established is the same as the identifier of the model in the sample/model library, the corresponding training sample is added to the training sample set of the corresponding model for training to establish a model corresponding to the identifier.

The invention also provides a personalized self-learning recognition system, comprising

a server for storing a sample/model library, including a training sample set and a model;

The client, the network connection with the server, and further includes,

One or more sensors for detecting a signal of the detected object;

a processing unit, configured to receive a signal of the detected object detected by the sensor, and identify a state type of the detected object according to the sample/model library stored by the server;

among them,

The server is further configured to instruct the processing unit of the client to select to communicate with the one or more sensors.

The invention further provides a personalized self-learning recognition system, including a client and a server:

The server includes a memory for storing a sample/model library;

The client further includes

One or more sensors for detecting a signal of the detected object;

The processing unit further transmits the feedback information and training samples of the model to be established to the server through a network;

The server is further configured to receive the feedback information sent by the client and the training samples of the model to be established, and use the corresponding training sample set to train the model and update the model library.

Other features, objects, and effects of the present invention will become more apparent and understood from

DRAWINGS

1 is a structural diagram of a personalized learning recognition system according to an embodiment of the present invention;

Figure 2 is a structural diagram of a processing unit in the system of Figure 1;

FIG. 3 is a flowchart of a personalized learning and recognition method according to an embodiment of the present invention; FIG.

4 is a schematic structural diagram of a personalized learning recognition system according to another embodiment of the present invention.

In all of the above figures, the same reference numerals are used to indicate the same or similar features or functions.

detailed description

The structure and working principle of the personalized self-learning recognition system 100 of the present invention will be described in detail below with reference to FIG. 1 in conjunction with FIGS. 2 and 3.

The personalized self-learning recognition system 100 shown in FIG. 1 includes one or more sensors 1021-102n, a processing unit 101, a storage unit 103, and an input/output device 104.

The sensors 1021-102n may be signals for detecting the state of the detected object. Wherein, the signal indicating the state of the detected object may include a physiological state signal. For example, it may include a motion sensor (eg, an acceleration sensor, a gyroscope, an angular velocity sensor, etc.), a pulse sensor, an electrocardiogram sensor (ECG) Sensor), EMG sensor, EEG sensor, blood oxygen sensor (SPO2) One or a combination of sensors, audio sensors, video sensors, etc. Further, the sensors 1021-102n may further include an environmental sensor for detecting a state affecting the detected object, for example, one of a temperature sensor, a humidity sensor, or the like combination. According to the usage mode of the type of the sensor, it can be placed on the hand of the object to be detected, for detecting blood oxygen, myoelectricity and pulse of the object to be detected; placing it on the leg of the object to be detected, for Detecting a motion signal of the detected object; placing it on the chest of the detected object for detecting the electrocardiographic signal; placing it on the head of the detected object, for detecting the EEG signal of the detected object, etc.; or placing The position close to the object to be detected and the surroundings of the object to be detected are, for example, used to detect the sound of the object to be detected, the temperature of the video signal and its surroundings, humidity, and the like.

The object to be detected may be the user himself or other person who needs to monitor the user, such as an elderly person, a patient, a child, an athlete in training, or the like.

The storage unit 103 is configured to store a model, such as a physiological state model library, and may further include an action model library, an electrocardiogram model library, a pulse model library, an electromyography model library, an electroencephalogram model library, a blood oxygen model library, and a sound model. Library, video model library, temperature model library, humidity model library, etc. The storage unit 103 may be integrated inside the processing unit 101 or may be disposed outside the processing unit 101.

According to an embodiment of the present invention, when the system 100 is in an initialization state, the storage unit 103 may store an initial reference model, wherein the reference model may be a model trained according to a large sample set, for example, the motion model library may include a walking model, running and jumping Model, sleep behavior model, etc.; myocardial infarction model, arrhythmia model, etc. may be included in the ECG model library; the EEG model library may include a normal model, an epilepsy model, a sleep disorder model, etc.; the pulse model library may include a normal model An abnormal model, etc.; the muscle electrical model library may include a muscle fatigue model and a muscle excitation model; the blood oxygen model library may include a hypoxemia model and a normal oxygen carrying model; the sound model library may include a quiet model, a sleep model, and a working model. The hybrid model; the video model library may include a normal behavior model and an abnormal behavior model; the temperature model library and the humidity model library may include a normal model and a prone disease model.

It can be understood that the initialization time storage unit 103 may not store any reference model, and may continuously establish new various types of models and update the storage unit 103 by training according to the input detected signal of the detected object. Model library.

The processing unit 101 can be connected to the sensors 1021-102n in various manners, for example, through an I2C bus, a UART, an SPI bus, or the like, or can be connected by a wired method such as a USB, a network interface, or the like. Modes such as Bluetooth, Zigbee, Wifi, infrared, etc. are wirelessly connected to communicate with sensors 1021-102n.

As shown in FIG. 2, the processing unit 101 may further include an identification module 101a for identifying a state type of the detected object represented by the detection signal input by the sensors 1021-102n; and a training module 101b for training using the training sample model. Further, the storage unit 103 is further configured to store training samples. The training sample may be a detection signal input by the sensors 1021-102n, or may be a feature after the feature extraction by the identification module 101a.

Referring to FIG. 3, first, the processing unit 101 receives the detection signals of the sensors 1021-102n;

Then, the identification module 101a in the processing unit 101 performs recognition processing in accordance with the corresponding model. The identification processing process further includes a pre-processing step 200, a feature extraction step 201, a classification identification step 202, and the like.

In step 200, the identification module 101a in the processing unit 101 performs corresponding pre-processing on the detection signals from the sensors 1021-102n, which may include removing noise in the signal by using an algorithm such as filtering;

Then, proceeding to step 201, the identification module 101a in the processing unit 101 extracts the pre-processed signal according to the feature. Among them, feature extraction and selection are performed according to different types of sensors and different characteristics required for classification. For example, the features extracted from the motion signals transmitted by the acceleration sensor may include time domain and frequency domain features, wherein the time domain features include, for example, mean, variance, short-term energy, autocorrelation coefficient, and cross-correlation of the amplitude of the motion signal, Signal period, etc.; frequency domain features include cross-correlation coefficients in the frequency domain obtained by FFT (Fast Fourier Transform) of motion signals, MFCC (Meier Cepstral Coefficient), and the like. The feature extracted from the electrocardiographic signal transmitted from the electrocardiographic sensor may include a QT period of the QRS wave, a QRS slope, a ST segment slope, etc.; the feature extracted from the myoelectric signal according to the electromyogram sensor may include, for example, a time domain feature The average absolute value, the absolute value of the average slope, the zero-crossing rate, etc., such as the average power frequency, the median frequency, the peak frequency, etc. in the frequency domain characteristics; the temperature change rate and the like can be extracted according to the temperature signal transmitted from the temperature sensor. Then, the feature is selected by assigning weight values to the extracted features.

Next, proceeding to step 202, the identification module 101a of the processing unit 101 classifies and extracts the extracted features according to corresponding models in the storage unit 103, wherein the classification recognition algorithm may adopt k-nearest neighbor, Gaussian, Bayesian, artificial neural network, etc. .

Then, proceeding to step 203, the processing unit 101 outputs the recognition result to the user via the input/output device 104. The output device may be an audio output device, a liquid crystal display or the like for providing an audio output or a user interface. Wherein, the input device may include a button, a keyboard, a touch screen, an audio or video sensor, and the like.

When the identification module 101a in the processing unit 101 cannot identify the state type of the detected object based on the sample/model library in the storage unit 103, for example, the output recognition result is a rejection type, that is, a model that does not belong to the stored sample/model library. Type; or, when the recognition result output by the processing unit is wrong, for example, the electrocardiographic recognition module outputs the recognition result "normal" through the input/output device 104, but the detected object is abnormal, or the motion recognition module 101a passes through the input/output device 104. The output recognition result is “running”, but when the actual action of the user is “walking”, the processing unit 101 can receive feedback information input by the user through the input/output device 104, for example, the indication system recognizes an error instruction, and judges according to the feedback information. Whether to establish a new model; if yes, proceed to step 204 to receive feedback information of the model identification to be established input by the user through the input/output device 104; if not, receive the next one or more detection signals and proceed to the step 200.

Preferably, the feedback information may further include an identifier of a scene corresponding to each model. For example, if the type of action performed by the detected object is a swing motion in golf and the model needs to be established, the user input model is identified as a "golf scene" and "Swing action." Or, if the detected ECG signal of the detected object, and the current state of the detected object is "angina" and the need to establish the model, the user can enter the model identification as "angina." The model identifier library may be pre-stored in the storage unit 103, and the user may select, by using the input/output device 104, the model identifier to be established from the model identifier library, if the model identifier to be established is not stored in the existing model identifier. When the library is available, the user can enter the model ID.

After the user inputs the model identification, the processing unit 101 takes the corresponding signal/feature as a training sample and trains the model through the training module 101b. The training module 101c in the processing unit 101 can be trained by a machine learning algorithm well known in the art, such as hybrid Gaussian, support vector machine, Bayesian or other well-known algorithms.

Next, the processing unit 101 determines, according to the existing sample/model library, whether the model identifier is identical to a certain model identifier in the sample/model library;

If they are the same, proceed to step 205, the training module 101b in the processing unit 101 adds the extracted and selected features of the current input signal to the existing sample set having the same model identification for training; for example, if the processing unit 101 receives The signal is an electrocardiogram signal and an action signal, and the identifier of the model to be established input by the user is “abnormal”, and the ECG signal and the motion signal can be respectively added to the existing sample/model library and identified as “abnormal”. The corresponding signal training of the model is focused on training and a new "abnormal" model is established.

Then, proceeding to step 206, the existing identically identified model is updated to update the existing sample/model library.

Preferably, the processing unit 101 may also select a detection signal that only communicates with one or more of the sensors or processes only its input, according to an instruction input by the user via the input/output device 104. For example, when the user only needs to monitor their EEG and motion signals, an instruction can be input to instruct the processing unit 101 to communicate only with the EEG and motion sensors, or only the detection signals input by the two sensors.

Further, the storage unit 103 can also store a corresponding sample/model library based on various different scenarios, such as a fitness scene, an office scene, a home scene, and the like. Each scene may also include multiple sub-scenes. For example, the fitness scene may also include a yoga scene, a tennis court view, and the like.

Each scene or sub-scene may include a sample/model library corresponding to each detection signal. For example, the yoga sub-scenario may include a sample/model library corresponding to the motion signal, such as a leg lift, a bent-up sample/model, etc.; the tennis sub-scenario may include a tennis action sample/model library, such as a serve, a swipe sample/model, etc.; The action sample/model library corresponding to the scene may include work, rest, and other action samples/models; the stored action sample/model library of the family scene may include action samples/models such as housework, watching TV, and eating;

In the fitness scene, the office scene, and the family scene, the myoelectric sample/model library corresponding to the myoelectric signal may include a muscle fatigue model and a muscle excitation model; the EEG sample/model library corresponding to the EEG signal may include a mental stress model, Mental relaxation samples/models; ECG samples/model libraries corresponding to ECG signals may include samples/models such as “normal” and “abnormal”, where abnormal samples/model libraries may also include “arrhythmia” and “myocardial infarction” "Sequence samples/models.

The input/output device 104 can output a scene identifier for the user to select, and the user can select a scene or a sub-scene thereof through the input/output 104 device; the processing unit 101 can identify the status type of the detected object according to the model corresponding to each scene. For example, the type of action performed, the type of brain electrical power, the type of electrocardiogram, etc., and the recognition results are respectively output.

Further, the user can also input the scene identifier corresponding to the model to be established through the input/output device 104. For example, if the model to be trained by the detected object is a "swing" action model in a golf scene and a corresponding blood oxygen model, an electroencephalogram model, an electrocardiogram model, a humidity model, a temperature model, etc., the user can input according to the input. The prompt input scene of the output device 104 is identified as "golf scene", and then the input model is identified as "swing"; the processing unit 101 receives the corresponding scene identifier and model identifier input by the user, and marks the current model of the new training as "Swing" and then update the corresponding sample/model library.

Preferably, the processing unit 101 may further perform the training model on the extracted features according to the signals detected by the plurality of sensors 1021-102n. The classification algorithm of the identification module 101a in the processing unit 101 can perform feature extraction according to the maximum feature dimension preset by different sensors, and perform feature selection according to the actually connected sensor; similarly, the training module 101b can be set to be the largest according to the preset. Feature dimension is used to extract features and select features for training.

For example, assume that the sensors that the system 100 can connect are motion sensors, ECG sensors, and EEG sensors, wherein the 3-D feature vector of the mean, variance, and short-time energy of the amplitude is extracted from the motion signal detected by the motion sensor; The signal extracted by the sensor is characterized by the QT period of the QRS wave, the QRS slope, and the ST-segment slope. The two-dimensional vector extracted from the myoelectric sensor is the average absolute value and the average absolute value of the slope. The corresponding classification/training algorithm is to classify/train according to the 8-dimensional feature vector. When the user selects only the myoelectric and motion signals, the weight value of the feature of the electrocardiographic signal in the 8-dimensional feature vector corresponding to the recognition module 101a is set to (0, 0) and classified/trained. By analogy, the detected signals can be separately extracted according to the motion sensor, electrocardiogram, electroencephalogram, blood oxygen, temperature, humidity sensor, etc., and the model can be trained. In this way, the user can monitor the status type of the detected object in different environments.

The sensors 1021-102n continue to collect motion signals for the detected object. Based on the updated state model in the storage unit 103, the processing unit 101 continues to identify the state type of the detected object and trains to continuously update and refine the state samples/models.

It should be noted that due to the insufficient number of samples, when constructing the initial model, for example, the constructed Gaussian model, learning by training samples, calculating mean, variance, etc., initializing Gaussian distribution and prior probability, the recognition rate of the initial model may be Not high.

As the detected signal of the detected object increases, the number of training samples also increases, and the training samples of the same state are combined for training, and adaptive learning is performed on the basis of the existing model, thereby possibly making the training The recognition rate of the model is improved.

Preferably, the storage unit 103 is further configured to store a detection signal and/or a recognition result of the detected object.

Preferably, the input/output device 104 of the system 100 can also provide a user interface for the user to select different modes of operation, for example, a training mode, a monitoring mode, a recording mode, and the like can be provided. The user can select different modes through the input/output device 104 as needed, for example, the training mode, the user can actively perform model training and build a model; in the monitoring mode, the processing unit 101 can be compared according to the recognition algorithm and the sample/model library. The state of the detected object is monitored and identified; in the recording mode, the processing unit 101 can record the signal and/or the recognition result of the detected object.

4 shows a personalized self-learning recognition system 300 of the present invention. Among them, the client 301 and the server 302 are included. Client 301 and server 302 perform data transmission over network 303, such as Wi-Fi, GSM, LAN, USB, Bluetooth, WLAN, etc., as is known in the art.

The server 302 includes a storage unit 3023 for storing a sample/model library;

The components of the client 301 that have the same or similar functions as the system 100 shown in FIG. 1 are not described again.

Wherein, the sensors 30121-3012n are used to detect a signal of the detected object;

The processing unit 3011 is configured to receive the signal of the detected object detected by the sensors 30121-3012n, and identify the current state type of the user according to the sample/model library downloaded from the server 302 and output the recognition result through the input/output device 3013;

When the identified current status type is the rejection type or the error type, the user may input a feedback information through the input/output device 3013, for example, indicating the recognition result error information, and the processing unit 3011 transmits the feedback information to the processing unit 3011 according to the feedback information. ;

The processing unit 3011 uses the corresponding training sample training model, and then sends the trained completed model and the corresponding training samples to the server 302;

The server 302 then updates the sample/model library.

Preferably, the server 302 is further configured to store the detected signal library and/or the recognition result.

Preferably, the input/output device 3013 may also prompt the user to input the model identifier to be trained, and if the model identifier input by the user is the same as the model identifier in the existing sample/model library, then the corresponding model is added. The training samples are focused on training and updating the sample/model library.

Further, the server 302 may further include a processing unit 3021 and an input/output device 3022. The user may send various detected detection signals of the detected object to other remote monitoring terminals through the processing unit 3021 in the server 302. For example, Sending to the remote monitoring center for health analysis; the user can also use the input/output device 3022 in the server 302 to set the signal type of the sensor to be processed by the processing unit 3011 of the client according to the personalized condition of the detected object and its status. . For example, if the detected object is a heart disease patient through the health analysis, the configuration information may be set as the ECG signal and the motion signal to be detected, and sent to the processing unit 3011 of the client, and the processing unit 3011 processes only according to the configuration information. The ECG signal and motion signal of the detected object.

According to another embodiment of the present invention, the system 300 may be further configured to: the processing unit 3011 of the client 301 is configured to receive the signal of the detected object detected by the sensors 30121-3012n, and according to the sample/model library downloaded from the server 302. Identify the user's status type;

According to the identified current state type, if it is a rejection type or an error type, the user can input feedback information including model identification information through the input/output device 3013; the processing unit 3011 receives the feedback information and passes the corresponding training sample through the network. 303 is sent to the server 302;

The processing unit 3021 in the server 302 is configured to receive the feedback information sent by the client 301 and the training samples of the model to be established, and use the corresponding training sample set to train the model and identify the established new model, and then update the storage unit 3023. Sample/model library.

It will be appreciated that the system of the present invention can also be used to monitor the condition of an object. For example, whether the instrument in operation is normal, etc., different types of sensors such as an audio sensor, a vibration sensor, a pressure sensor, etc. can be placed at corresponding positions for detecting sound, vibration, pressure signals, and identifying instruments according to normal or abnormal models. Working condition, the user trains and builds a new model based on the identified results and continuously updates the sample/model library to more accurately monitor the working state of the instrument.

The above description is only a preferred embodiment of the present invention, and it should be noted that those skilled in the art can also make several improvements and retouchings without departing from the principles of the present invention. It is considered as the scope of protection of the present invention.

Claims

A personalized self-learning recognition system comprising:

One or more sensors for detecting a signal of the detected object;

a storage unit for storing a sample/model library, including a model and a training sample set;

a processing unit, configured to receive a signal of the detected object detected by the one or more sensors, and identify a state type of the detected object according to a corresponding model in the sample/model library;

Input/output means for outputting a recognition result and receiving feedback information input based on the recognition result;

The processing unit also trains the model using the corresponding training sample set and updates the sample/model library based on the feedback information.

2. The system of claim 1 wherein:

The set of training samples includes characteristics of signals and/or signals of the detected object.

3. The system of claim 1 wherein:

The sample/model library includes corresponding models in one or more scenarios.

4. The system of claim 1 wherein:

The feedback information includes an identification of a model to be established.

5. The system of claim 3 wherein:

The feedback information includes a scene identifier of a model to be established and a corresponding model identifier.

6. System according to claim 4 or 5, characterized in that:

When the input identifier of the model to be established is the same as the identifier of the model in the sample/model library, the processing unit adds the corresponding training sample to the training sample set of the corresponding model for training to establish a corresponding to the identifier. model.

7. The system of claim 1 wherein:

The processing unit is further configured to select to communicate with the one or more sensors according to a selection instruction received by the input/output device.

8. The system of claim 1 wherein:

The input/output device is also operative to receive an input selected mode of operation.

9. The system of claim 8 wherein:

The operation mode includes one of a training model mode, a monitoring mode, a recording mode, or a combination thereof;

The processing unit separately performs model training, state recognition, and stores the recognition result and/or the corresponding detection signal into the storage unit according to the input operation mode.

10. A personalized self-learning recognition method, including:

Using one or more sensors to detect the signal of the detected object;

Receiving the detected signal of the detected object, identifying a state type of the detected object according to a corresponding model stored in the sample/model library, and outputting the recognition result;

The input feedback information is received, and the model is trained based on the feedback information and the sample/model library is updated.

11. The method of claim 10 wherein:

The feedback information includes an identification of a model to be established.

12. The method of claim 11 wherein:

When the input identifier of the model to be established is the same as the identifier of the model in the sample/model library, the corresponding training sample is added to the training sample set of the corresponding model for training to establish a model corresponding to the identifier.

13. A personalized self-learning recognition system comprising:

a server for storing a sample/model library, including a training sample set and a model;

The client, the network connection with the server, and further includes,

One or more sensors for detecting a signal of the detected object;

a processing unit, configured to receive a signal of the detected object detected by the sensor, and identify a state type of the detected object according to the sample/model library stored by the server;

Input/output means for outputting a recognition result and receiving feedback information input based on the recognition result;

The processing unit also trains the model using the corresponding training sample set and updates the sample/model library based on the feedback information.

14. A system according to claim 13 wherein:

The server is further configured to instruct the processing unit of the client to select to communicate with the one or more sensors.

15. A personalized self-learning recognition system, including client and server:

The server includes a memory for storing a sample/model library;

The client further includes

One or more sensors for detecting a signal of the detected object;

a processing unit, configured to receive a signal of the detected object detected by the sensor, and identify a state type of the detected object according to the sample/model library stored by the server;

Input/output means for outputting a recognition result and receiving feedback information input based on the recognition result;

The processing unit further transmits the feedback information and training samples of the model to be established to the server through a network;

The server is further configured to receive the feedback information sent by the client and the training samples of the model to be established, and use the corresponding training sample set to train the model and update the model library.