TWI682330B - Self-learning data classification system and method - Google Patents

Abstract

A self-learning data classification system includes a data base storing a group of training data with same features of a data signal to be measured; a base generating unit electrically connecting the data base to generate a first group of bases representing principal components of data signals of the training data; an index generating unit used to generate corresponding indices according to the first group of bases and the data signal to be measured; and a index classification computing unit electrically connecting the index generating unit to calculate out classification result of the data signal to be measured according to the generated indices. The base generating unit generates a second group of bases being different from the first group of bases according to the generated indices.

Description

Self-learning data classification system and method

The invention relates to a data classification system and method, in particular to a data classification system and method capable of self-improving classification ability in the process of data classification.

Nowadays, the popularity of computers and the development of the Internet have created a large accumulation of various types of data and the establishment of corresponding databases. In the era of big data, these materials are valuable assets for enterprises and indispensable tools in the process of product development and application. As we all know, the establishment of these data depends on the various processing of the signals that carry these data, and signal classification is the most critical. By accurate signal classification, data of known specific events corresponding to known specific results can be created, and then the data can be used to analyze or predict the results of the events to be tested. Such data classification system products have also been developed . For example, we can classify a large number of physiological signals that have been obtained from human detection, and use the classification results to diagnose human health.

When classifying signals, the processed data signals are often high-dimensional signals and have high data complexity. If these high-dimensional signals are directly used for classification, it is bound to increase the difficulty of the classification system. In order to reduce the difficulty and cost of building a classification system, it is necessary to perform signal pre-processing on these high-dimensional signals and dimension reduction into low-dimensional signals, and then classify these low-dimensional signals. Known signal pre-processing methods are, for example, two-dimensional convolutional neural network analysis, one-dimensional convolutional neural network analysis, or recurrent neural network analysis. However, the amount of data and time required to use these analysis methods is quite large, and it is not cost-effective for a data classification system that emphasizes efficiency and results. Other signal pre-processing methods are, for example, Principal Component Analysis (PCA), which reduces the dimensionality of the data signal to a combination of a specific basis and its corresponding index. The so-called specific basis is a common part of a group of data signals that have the same data attributes as the information signal to be reduced, and the corresponding parameter is the proportion of these bases in the data signal to be reduced. As shown in FIG. 7, by transforming the three common bases of A, B, and C, the five data signals can be expressed as a combination of these common bases and three corresponding parameters, respectively. However, the accuracy of the data classification system built in this way depends on the accuracy of the data signal after dimensionality reduction, and the accuracy of the data signal after dimensionality reduction is closely related to the decision of the base, if the decision of the base is only achieved by human means , It is easy to produce deviations and not easy to get the best base, and the decision of the human way cannot be standardized to produce quantitative indicators, and the accuracy of the classification system cannot be controlled.

Therefore, how to reduce the difficulty and cost of the classification system while reducing the complexity of the data signal and improving the accuracy of the signal classification to ensure that the data classification system product built based on this signal classification can be more accurately analyzed in application Or predicting the result of the event to be measured is a technical problem to be solved by the present invention.

In view of the above problems, the present invention provides a self-learning data classification system and method, which generates a base through a computer automatic learning method, eliminates possible deviations in the selection of artificial methods, improves the accuracy of signal classification, and ensures that based on this signal The data classification system built by the classification method has a high classification recognition rate and an effective classification result (classification result).

In one embodiment, the present invention provides a self-learning data classification system for classifying a data signal to be tested (also called a data signal to be classified). The self-learning data classification system provided includes a first subsystem and a second subsystem. The first subsystem has a database that stores a group of training data (training data), the data signal of the training data has the same attributes as the signal of the data to be tested; a base generation unit, which is electrically connected to the database and used for training The data generates a first group of bases representing the principal component of the data signal of the training data. The number of bases of the first group of bases is at least one; and a return parameter receiving unit electrically connected to the base generating unit . The second subsystem is electrically connected or network-connected with the first subsystem and has a data signal measurement unit for measuring the data signal to be measured; a parameter generation unit is electrically connected to the data signal measurement unit, It is used to generate the index corresponding to the first group of bases according to the first group of bases and the data signal to be tested; an argument return unit is electrically connected to the parameter generating unit to return the generated parameters It is transmitted to the first subsystem; and a parameter classification calculation unit, which is electrically connected to the parameter generation unit, and used to calculate the classification result of the data signal to be tested according to the generated parameters. The return parameter receiving unit receives the parameters returned by the parameter return unit, and the base generation unit generates a second set of bases different from the first set of bases according to the returned parameters and the data signal of the training data The number of substrates in the second group of substrates is at least one.

In one embodiment, the base generation unit has a matrix processing unit, electrically connected to the database, which converts at least part of the data signals in the training data into a matrix; a feature decomposition unit, which is electrically connected to the matrix processing unit Sexual connection, which implements a singular value decomposition (SVD) on the matrix to obtain a singular value and a corresponding singular vector; a low rank approximation processing unit of the matrix , Which is electrically connected to the feature decomposition unit, which calculates the most approximate low-rank matrix of the matrix based on the singular values and singular vectors; and a multi-layer sensing unit, which is electrically connected to the matrix low-rank approximate processing unit, which receives the most approximate low rank The data signal corresponding to the matrix outputs one of the first group of substrates and the second group of substrates.

In one embodiment, the first subsystem further has a loopback parameter evaluation unit, which is electrically connected to the base generation unit and the loopback parameter receiving unit, and is used to evaluate the loopback parameters in order to judge the loopback Whether the argument of is higher than a set threshold.

In one embodiment, the first subsystem further has a substrate output unit, which is electrically connected to the substrate generating unit, and is used to transmit the first group of substrates and the second group of substrates generated by the substrate generating unit to the second subsystem through the network .

In one embodiment, the second subsystem further has a substrate input unit, which is electrically connected to the data signal measurement unit and used to receive the first group of substrates and the second group of substrates transmitted from the substrate output unit.

In one embodiment, the data signal to be tested includes vital signs.

In another embodiment, the present invention provides a self-learning data classification method for classifying a data signal to be tested, including the following steps: generating a first representation of the main component of the data signal representing training data based on a group of training data A group of bases, the data signal of the training data has the same attributes as the data signal to be tested, and the number of bases of the first group of bases is at least one; based on a first group of data signals to be tested and the first group of bases, the corresponding to the first group is generated A first set of parameters of the base; a second set of bases is generated based on the first set of parameters and the data signal of the training data, and the number of bases of the second set of bases is at least one; and the first set of parameters is calculated according to the first set of parameters The classification result of a set of data signals to be tested.

In an embodiment, the steps of generating the first set of bases and the second set of bases are performed on a first subsystem, and the generation of the first set of parameters and the calculation of the classification results of the first set of data signals to be tested are performed in It is executed on a second subsystem, which is far away from and controlled by the first subsystem.

In one embodiment, the provided self-learning data classification method further includes the following steps: determining whether the first set of parameters is higher than a set threshold.

In an embodiment, the provided self-learning data classification method further includes the following steps: generating a second set of parameters corresponding to the second set of bases based on the second set of bases and a second set of data signals to be tested; and based on the first Two sets of parameters calculate the classification results of the second set of data signals to be tested.

In the self-learning data classification system and method proposed by the present invention, since the second subsystem performs signal pre-processing on the measured data signals, the base required by the first subsystem is generated by the first subsystem, thus completely eliminating the artificial The possible deviation of the method selection not only improves the efficiency of signal pre-processing but also improves the accuracy of signal classification, and ensures that the classification results have a high and effective classification recognition rate. In addition, the first subsystem uses a multi-layer perception architecture to generate the base, so the base generation process is completely self-learning, and by using the matrix low-rank approximation method in the inference process of generating the base, high-dimensional signals can be effectively carried out It reduces dimensionality and effectively reduces the number of input layer neurons and hidden layers required by the multi-layer perception architecture, thereby reducing the difficulty and cost of building a classification system. Furthermore, by synchronously transmitting the parameters of the second subsystem used in the classification algorithm to the first subsystem, the inference process of the base generation unit of the first subsystem in generating the base is further strengthened, allowing the first The multi-layer sensing unit of the subsystem can generate the optimal base, thereby enhancing the performance of the second subsystem in signal pre-processing of the measured data signal. In addition, since the classification algorithm of the second subsystem can be performed independently of the signal pre-processing, the adjustment of the calculation mechanism of the classification algorithm is more flexible and flexible.

In order to make the above-mentioned features and advantages of the present invention more obvious and understandable, the embodiments are specifically described below in conjunction with the accompanying drawings for detailed description as follows.

The present invention discloses a self-learning data classification system and method. The basic principles of the data signal pre-processing and the multi-layer perceptron in the neural network-like architecture are already understood by those of ordinary skill in the art, so the following description , No longer complete description. At the same time, the drawings referred to in the following are intended to express the meanings related to the features of the present invention, and they are not and need not be completely drawn according to the actual size, which is stated in advance.

FIG. 1 is a schematic diagram of a system architecture of a self-learning data classification system 10 according to an embodiment of the present invention. In this embodiment, the self-learning data classification system 10 has a first subsystem 100 and a second subsystem 200. The second subsystem 200 is used to measure the data signal to be measured or the data signal to be classified and use a classification algorithm (classification algorithm) to obtain the classification result of the data signal to be tested, and the first subsystem 100 is used to generate the base required by the second subsystem 200 for signal pre-processing of the measured data signal to be tested, so that The second subsystem 200 thereby obtains the corresponding parameters of the base, and uses the parameters for subsequent classification calculations. The first sub-system 100 is usually located far away from the second sub-system 200, but the first sub-system 100 and the second sub-system 200 can be connected to each other through a network to enable the substrate generated by the first sub-system 100 and the second sub-system 200 Transmission and reception of the generated parameters. The first system 100 is, for example, a server system, and the second system 200 is, for example, a client system. In other embodiments, the first sub-system 100 and the second sub-system 200 may be electrically connected instead of communicating through the network, so that the generation of the base and the parameters is implemented under the same hardware architecture.

As shown in FIG. 1, in one embodiment, the first subsystem 100 has a database 101, a substrate generating unit 102, a substrate output unit 103, a loopback parameter receiving unit 104 and a loop electrically connected to each other Passing number evaluation unit 105. The database 101 is a database that stores a group of training data. The storage is implemented by, for example, memory. The data signal of the stored training data has the same attributes as the data signal to be measured measured by the second subsystem 200, for example All are physiological signals with vital signs. During the operation of the first subsystem 100, the base generation unit 102 is used to generate a set of bases representing the main components of the data signals of the training data according to the training data of the database, and the base output unit 103 is updated online The generated substrate is transferred to the second subsystem 200. The number of substrates in each group of substrates is at least one.

2 is a schematic diagram of the functional structure of the base generation unit 102 of the first subsystem 100 of the self-learning data classification system 10 according to an embodiment of the present invention. FIG. 3 is a schematic diagram of the inference process of the base generation unit 102 generating a set of bases according to the known data signal q in the training data. As shown in FIG. 2, in an embodiment, the base generation unit 102 has a matrix processing unit 1021, a feature decomposition unit 1022, a matrix low rank approximation processing unit 1023 and a multi-layer perception unit 1024 electrically connected to each other. The matrix processing unit 1021 is used to convert at least part of the data signal in the training data, for example, the high-dimensional data signal q into the matrix A. The feature decomposition unit 1021 is used to obtain a singular value Σ and a singular vector V ^T after performing a singular value decomposition (SVD) on the matrix A, where A≒UΣV ^T , U and V are orthogonal normalized matrices (orthonormal matrix), namely The vectors contained in the matrix are perpendicular to each other and have a length of 1 in a high-dimensional space. In this way, the matrix A can be regarded as the projection of the data signal q on the V coordinate system, and each coordinate value is multiplied by the value on the diagonal of Σ, and then reorganized on the U coordinate system. As shown in FIGS. 2 and 3, the matrix low-rank approximation processing unit 1023 is used to select the first few singular values Σ′ and the corresponding singular vector V′ ^T arranged in order of magnitude, and calculate the most approximate matrix A according to Low-rank matrix A', and then calculate the dimensionality reduction data signal q'corresponding to the low-rank matrix A'. As shown in FIGS. 2 and 3, the multi-layer perception unit 1024 receives each dimensionality-reduced data signal q′, and outputs a set of data signals q representing these data signals according to the principle of multi-layer perceptrons in a neural network-like architecture. The base of the main component. Since the projected coordinate system already represents the main components of the known data signal q that are linearly independent and sorted according to the size of the feature value, that is, the more relevant information in the original data signal has been clustered on each coordinate axis, it can effectively reduce the multi-layer The number of neurons in the input layer 10241 and the number of hidden layers 10242 required by the sensing unit 1024.

Please refer to FIG. 1 again, the return parameter receiving unit 104 is used to receive the parameters generated and returned by the second subsystem 200, and the return parameter evaluation unit 105 is used to evaluate the returned parameters in stages to determine Whether the returned parameter is higher than a set threshold. The so-called fractional evaluation refers to the evaluation of the return parameters only after a set period of time, such as 10 seconds, instead of evaluating the return parameters every moment. When the judgment result of the returned parameter after fractional evaluation is lower than the set threshold, the continuous evaluation is continued, and when the judgment result of the returned parameter after fractional evaluation is higher than the set threshold, the base generation unit 102 Generate a new set of bases representing the main components of these data signals based on the returned parameters and the data signals of the training data in the database. The number of bases is at least one. During the operation of the first sub-system 100, when the parameters generated by the second sub-system 200 are returned, the first sub-system 100 may modify the parameters used in the operation of the second sub-system 200 in an online update mode as appropriate. The base allows the second subsystem 200 to classify the signals to be measured with a new set of bases.

On the other hand, as shown in FIG. 1, in an embodiment, the second subsystem 200 has a data signal measurement unit 201, a base input unit 202, a parameter generation unit 203, and a parameter return unit 204 that are electrically connected to each other And argument classification calculation unit 205. The data signal measuring unit 201 is used to measure and receive the data signal 210 to be tested input to the second subsystem 200 through the input terminal 211 of the second subsystem 200 and store it accordingly, for example, by a temporary memory, to be tested The data signal 210 has the same attributes as the data signal of the training data stored in the database 101 of the first subsystem 100, such as a physiological signal. The substrate input unit 202 is used to receive the substrate transferred from the substrate output unit 103 of the first subsystem 100. The argument generation unit 203 performs argument calculation on the data signal to be tested 210 received by the data signal measurement unit 201 according to the basis received by the basis input unit 202, and then generates arguments corresponding to these basis. The parameter return unit 204 returns the generated parameter to the first subsystem 100 and is received by the return parameter receiving unit 104. The parameter classification calculation unit 205 is used to calculate the classification result 220 of the data signal to be tested 210 according to the generated parameters through a classification algorithm. The classification algorithm is usually implemented by the execution of a computer program. The classification result 220 is finally output via the output terminal 212 of the second subsystem 200. During the process from the measurement of the data signal to be measured 210 to the classification algorithm, the generated parameters of the second subsystem 200 can be transmitted back to the first subsystem 100, so that the first subsystem 100 can Number strengthens its ability to generate the best substrate.

FIG. 4 is a flow chart of implementation steps of the first subsystem 100 of the self-learning data classification system 10 according to an embodiment of the present invention when performing data classification in the entire self-learning data classification system 10. In this embodiment, the data classification method of the self-learning data classification system 10 has the following steps:

Step 601: Receive the return parameter. As shown in FIG. 1, the return argument receiving unit 104 of the first subsystem 100 of the self-learning data classification system 10 receives the argument returned by the argument return unit 204 of the second subsystem 200.

Step 602: Determine whether there are return parameters. As shown in FIG. 1, the loopback parameter evaluation unit 105 of the first subsystem 100 of the self-learning data classification system 10 determines whether there are loopback parameters from the second subsystem 200. If no, go to step 603; if yes, go to step 604.

Step 603: According to the data signal of the training data in the database, generate a first set of bases representing the main components of the data signal of the training data. As shown in FIGS. 1 to 3, the base generation unit 102 of the first subsystem 100 generates a base of the first component representing the main components of the data signal of the training data according to the training data in the database 101. Then, step 607 is executed.

Step 604: Evaluate the return parameters in stages. As shown in FIG. 1, the feedback parameter evaluation unit 105 of the first subsystem 100 evaluates the returned parameters every time a set period of time, for example, 10 seconds, and then executes step 605.

Step 605: Determine whether the returned parameter is higher than the set threshold. As shown in FIG. 1, the loopback parameter evaluation unit 105 of the first subsystem 100 determines whether the loopback parameters are higher than a set threshold. If not, go back to step 604. If yes, go to step 606.

Step 606: Generate a second set of bases representing the main components of the data signal of the training data based on the returned parameters and the data signal of the training data in the database. The second group of substrates is different from the first group of substrates, and the number of substrates of the second group of substrates is at least one. As shown in FIGS. 1 to 3, the base generation unit 102 of the first subsystem 100 generates a new set of main components representing the data signal of the training data based on the returned parameters and the data signal of the training data in the database 101 Base. Then, step 607 is executed.

Step 607: Output the substrate. As shown in FIG. 1, after the substrate generating unit 102 generates the substrate, the substrate output unit 103 outputs the substrate to the second subsystem 200.

FIG. 5 is a flow chart of implementation steps of the second subsystem 200 of the self-learning data classification system 10 according to an embodiment of the present invention when performing data classification in the entire self-learning data classification system 10. In this embodiment, the data classification method of the self-learning data classification system 10 has the following steps:

Step 701: Measure the data signal to be measured. As shown in FIG. 1, the data signal measurement unit 201 of the second subsystem 200 of the self-learning data classification system 10 measures and receives the data signal to be tested input to the second subsystem 200 through the input terminal 211 of the second subsystem 200 210. Then, step 702 is executed.

Step 702: Generate parameters with the input base. As shown in FIG. 1, the base input unit 202 of the second subsystem 200 of the self-learning data classification system 10 receives the first set of bases and the second set of bases transmitted from the first subsystem 100, and the parameter generating unit 203 According to one of the first group of substrates and the second group of substrates received by the substrate input unit 202, the received data signal to be tested 210 is calculated in parameters, and then the corresponding to the first group of substrates or the second group of substrates is generated Arguments. Then, step 703 and step 705 are executed.

Step 703: Calculate the classification result of the data signal to be tested according to the generated parameters. As shown in FIG. 1, the parameter classification calculation unit 205 calculates the classification result 220 of the data signal to be tested 210 through a classification algorithm according to the generated parameters. Then, step 704 is executed.

Step 704: Output the classification result. As shown in FIG. 1, the classification result 220 is finally output via the output terminal 212 of the second subsystem 200.

Step 705: Return the argument. As shown in FIG. 1, the argument return unit 204 of the second subsystem 200 returns the generated argument to the base generation unit 102 of the first subsystem 100 for reference.

6 is a flowchart of implementation steps of a self-learning data classification method according to another embodiment of the present invention. In this embodiment, the self-learning data classification method has the following steps:

Step 801: Generate a first set of bases representing the principal components of the data signals of the training data according to a group of training data. The data signal of the training data has the same attributes as the data signal to be tested, and the number of bases of the first set of bases is at least For one. Then, step 802 is performed.

Step 802: Generate a first set of parameters corresponding to the first set of bases based on a first set of data signals to be tested and the first set of bases. Next, step 803 and step 806 are performed.

Step 803: Determine whether the first set of parameters is higher than a set threshold. Then, step 804 is performed.

Step 804: When the first set of parameters is higher than a set threshold, generate a second set of bases based on the first set of parameters and the data signals of the training data, and the number of bases in the second set of bases is at least one. Then, step 805 is executed.

Step 805: Generate a second set of parameters corresponding to the second set of bases based on the second set of bases and a second set of data signals to be tested. Then, step 806 is executed.

Step 806: Calculate the classification result of the first set of data signals to be tested according to the first set of parameters or calculate the classification result of the second set of data signals to be tested according to the second set of parameters.

In an embodiment, the above steps 801, 803, and 804 may be performed on a first subsystem 100 as shown in FIG. 1, and the above steps 802, 805, and 806 may be performed on a second as shown in FIG. Executed on the subsystem 200, the second subsystem 200 is far away from and controlled by the first subsystem 100.

In the self-learning data classification system and method proposed by the present invention, since the base required by the second subsystem 200 for signal preprocessing of the measured data signal is generated by the first subsystem 100, it is completely eliminated The possible deviations caused by manual selection not only improve the efficiency of signal pre-processing but also improve the accuracy of signal classification, and ensure that the classification results have a high and effective classification recognition rate. In addition, the first subsystem 100 uses a multi-layer perception architecture to generate the base, so the base generation process is completely self-learning, and the high-dimensional signals are effectively carried out by using the matrix low-rank approximation method during the inference process of generating the base It reduces dimensionality, and effectively reduces the number of input layer neurons and the number of hidden layers required by the multi-layer perception architecture, thereby reducing the difficulty and cost of building a classification system. Furthermore, by synchronously transmitting the parameters of the second subsystem 200 used in the classification algorithm to the first subsystem 100, the inference process of the base generation unit 102 of the first subsystem 100 in generating the base is further strengthened, allowing the first The multi-layer sensing unit 1024 of a sub-system 100 can generate an optimal base, thereby improving the performance of the second sub-system 200 in performing signal pre-processing on the measured data signal. In addition, since the classification algorithm of the second subsystem can be performed independently of the signal pre-processing, the adjustment of the calculation mechanism of the classification algorithm is more flexible and flexible.

In terms of application, when the data signal to be measured measured by the second subsystem 200 of the proposed self-learning data classification system 10 is a physiological signal including vital signs, the data signal measurement unit 201 may be a Vital sign measurement unit is used to measure body temperature, pulse, respiration and blood pressure of the human body. At this time, the second subsystem 200 may be a handheld physiological signal detector, and the first subsystem 100 may be a cloud server that remotely controls the handheld physiological signal measurer, the handheld physiological signal detector and the cloud server The devices are connected by wired or wireless network. Because the data of various physiological signals has its range, the relationship between the base and the parameters obtained after dimensionality reduction can be determined, which is more suitable for applying the self-learning data classification system proposed in this case. In other words, any data signal with a deterministic relationship between the basis and parameters obtained after dimensionality reduction is suitable for classification using the self-learning data classification system and method proposed by the present invention.

Various embodiments of the present invention are disclosed as above, but they are not intended to limit the present invention. Any person with ordinary knowledge in the technical field can make some changes and modifications without departing from the spirit and scope of the present invention. For example, The first subsystem and the second subsystem are integrated together. Therefore, the protection scope of the present invention shall be deemed as defined by the appended patent application scope.

10 self-learning unit 1022 wherein data classification system 100 of the first substrate 102 Database subsystem 101 generates a matrix processing unit 1021 matrix decomposition unit 1023 low rank approximation processing unit 1024 multilayer perceptron output unit 104 back to the base 103 of the hidden layer unit input layer 10242 10241 receiving unit 105 pass arguments return arguments evaluation unit 200 of the second sub-201203 argument information signal measuring unit 202, a base number generation unit 204 input unit primers return arguments classification unit 205 calculating unit 210 data signals to be measured 211 Input terminal 212 Output terminal 220 Classification result 601～607 Steps 701～705 Steps 801～806 Steps

FIG. 1 is a schematic diagram of a system architecture of a self-learning data classification system according to an embodiment of the invention. 2 is a schematic diagram of a functional structure of a base generation unit of a first subsystem of a self-learning data classification system according to an embodiment of the invention. FIG. 3 is a schematic diagram of an inference process of a base generation unit of a first subsystem of a self-learning data classification system according to an embodiment of the present invention to generate a base based on known data signals in training data. FIG. 4 is a flowchart of implementation steps of a first subsystem of a self-learning data classification system according to an embodiment of the present invention when data classification is performed by the entire self-learning data classification system. FIG. 5 is a flowchart of implementation steps of a second subsystem of a self-learning data classification system according to an embodiment of the present invention when data classification is performed by the entire self-learning data classification system. 6 is a flowchart of implementation steps of a self-learning data classification method according to another embodiment of the present invention. FIG. 7 is a schematic diagram of known high-dimensional signal dimensionality reduction into a combination of a specific base and its corresponding parameter.

10 self-learning classification system information database 100 a first subsystem 101 base generation unit 102 output unit 104 of the substrate 103 return arguments receiving unit 105 return arguments evaluation unit 200 second signal measuring subsystem 201 data input unit 202 of the substrate cell number generation unit 203 arguments lead 204 return arguments classification unit 205 calculating unit 210 measured data signals 211 input terminal 220 output the classification result 212

Claims (10)

A self-learning data classification system for classifying a data signal to be tested includes: a first subsystem with: a database storing a group of training data, the data signals of the training data and the test to be tested The data signals have the same attributes; a base generation unit, electrically connected to the database, is used to convert at least part of the data signals of the training data into a matrix and calculate a most approximate low rank matrix of the matrix to generate a first A group of bases representing the main components of the data signals of the training data, the number of bases of the first group of bases is at least one; and a return parameter receiving unit electrically connected to the base generating unit; and a first Two subsystems, electrically connected to the first subsystem or network connection, have: a data signal measuring unit for measuring the data signal to be tested; an argument generating unit and the data signal measuring unit The electrical connection is used to generate a parameter corresponding to the first group of substrates according to the first group of substrates and the data signal to be tested; a parameter return unit is electrically connected to the parameter generation unit to connect The parameter is returned to the first subsystem; and a parameter classification calculation unit electrically connected to the parameter generation unit to calculate the classification result of the data signal to be tested according to the parameter; wherein, the The return parameter receiving unit receives the parameter returned by the parameter return unit, and the base generating unit generates a different one from the first set of bases according to the returned parameter and the data signals of the training data The second group of substrates has at least one substrate. According to the self-learning data classification system described in item 1 of the patent application scope, wherein the base generation unit has: A matrixing processing unit, electrically connected to the database, which converts the partial data signals of the training data into the matrix; a feature decomposition unit, electrically connected to the matrixing processing unit, which connects the matrix After performing a singular value decomposition, a singular value and a corresponding singular vector are obtained; a matrix low-rank approximation processing unit is electrically connected to the feature decomposition unit, and calculates the nearest of the matrix according to the singular value and the singular vector A low-rank-like matrix; and a multi-layer sensing unit electrically connected to the matrix low-rank approximation processing unit, which receives the data signal corresponding to the most approximate low-rank matrix and outputs the first set of bases and the second set of bases one. According to the self-learning data classification system described in item 1 of the patent application scope, wherein the first subsystem further includes: a loopback parameter evaluation unit, which is electrically connected to the base generation unit and the loopback parameter receiving unit, It is used to evaluate the returned parameter in stages to determine whether the returned parameter is higher than a set threshold. According to the self-learning data classification system described in item 3 of the patent application scope, wherein the first sub-system further includes: a substrate output unit electrically connected to the substrate generation unit for the first generation of the substrate generation unit A group of substrates and the second group of substrates are transmitted to the second subsystem through the network. According to the self-learning data classification system described in item 4 of the patent application scope, wherein the second subsystem further includes: a base input unit, electrically connected to the data signal measurement unit, for receiving the base output unit The first group of substrates and the second group of substrates. The self-learning data classification system according to any one of items 1 to 5 of the patent application scope, wherein the data signal to be tested contains vital signs. A self-learning data classification method according to the self-learning data classification system described in item 1 of the patent application scope, used to classify the data signal to be tested, includes the following steps: generating representative training data based on the training data of the group The first set of bases of the main component of the data signal of the training signal, the data signal of the training data has the same attributes as the data signal to be tested, and the number of bases of the first set of bases is at least one; The measured data signal and the first set of bases generate a first set of parameters corresponding to the first set of bases; the second set of bases are generated according to the first set of parameters and the data signals of the training data, the second The number of bases of the group base is at least one; and the classification result of the first set of the data signal to be tested is calculated according to the first set of parameters. According to the self-learning data classification method described in item 7 of the patent application scope, wherein the generating steps of the first set of bases and the second set of bases are performed on the first subsystem, the generation of the first set of arguments and The calculation of the classification result of the first set of data signals to be tested is performed on the second subsystem, which is far away from and controlled by the first subsystem. According to the self-learning data classification method described in item 7 of the patent application scope, the method further includes the following steps: determining whether the first set of parameters is higher than a set threshold. According to the self-learning data classification method described in item 7 of the patent application scope, the method further includes the following steps: generating a second group corresponding to the second group of substrates based on the second group of substrates and a second group of data signals to be tested Parameters; and calculate the classification result of the second set of the data signal to be tested according to the second set of parameters.

Images