TWI761834B - Intelligent method for testing sensed data and system thereof - Google Patents

Abstract

An intelligent method for testing sensed data and a system are provided. The system includes a computer system having a memory for storing instructions and algorithmic programs that are used to perform the method, and a database for storing sensed data obtained from an object to be tested. The sensed data is divided into a training set and a testing set. A deep learning algorithm is performed on the training set for retrieving features of the object so as to establish a deep learning based testing model. The testing set is used to test the testing model, and a machine learning algorithm is performed on the data that fails to pass the test. Historical parameter data is further introduced to train the testing model being established by machine learning, and a model optimization algorithm is applied to optimize the testing model for obtaining optimized control parameters of the object.

Description

Sensing data intelligent detection method and system

The specification discloses a method for detecting sensing data, in particular a method and system for intelligently detecting sensing data by using deep learning and machine learning to establish a detection model from the sensing data.

When detecting whether a product, a system or a field is abnormal, a common method is to use a specific sensor to sense data, such as sensing sound, shooting images, etc., and then use analysis tools to analyze the sensed data. information, thereby judging whether the object to be tested has abnormal conditions.

For example, when it is necessary to determine whether the surface of an object is flawed, a camera can shoot the surface of the object, and the image of the surface of the object can be compared with the sample image to determine whether there is any abnormality. Taking a motor as an example, the general method is to use a sound sensor to record the audio generated when the motor is running. After comparing the sound model, it can be judged whether there is any abnormality in operation, which can be used as a basis for future improvement.

The specification discloses a method and system for intelligent detection of sensing data. The method runs in a computer system. The computer system is provided with a memory, which stores a program set and an algorithm for executing the intelligent detection method of sensing data. The computer system also has a database. The sensing data obtained from an object to be detected is stored therein.

In one embodiment, the sensing data is the data obtained by sensing the object to be detected by the sensor. In the intelligent detection method of the sensing data, the sensing data is firstly divided into training data and test data, and the training data is A deep learning algorithm is executed on the data, the features about the object to be detected are obtained from the training data, a detection model for detecting the object to be detected is established, and then the detection model is tested with the test data therein, and the test result is not passed. The test data is obtained, the machine learning algorithm is executed, the detection model is trained with the historical parameter data, and the detection model is optimized by a model optimization algorithm, so as to obtain the optimized control parameters of the object to be detected.

Further, the verification data is also obtained from the training data. When the detection model is formed from the training data through the deep learning algorithm in the above steps, the verification data can be used to verify the detection model to obtain a parameter model for generating control parameters. . The control parameters are parameters that drive the operation of the object to be detected, and generating optimized control parameters is one of the main purposes of the method.

Further, the model optimization algorithm can be used to optimize the parameter model, and then the optimized control parameters can be generated, and continue to drive the object to be detected, and the above steps are repeated to obtain the sensing data again. Detects the control parameters of the object.

Further, in one embodiment, a K-fold cross-validation method can be used to evaluate multiple models generated from multiple deep learning algorithms and machine learning methods using validation data, and the evaluated factors are, for example, generated from multiple deep learning algorithms. The accuracy, precision and recall rate of each model are obtained from the multiple models, and then the detection model is selected according to the score.

For a further understanding of the features and technical content of the present invention, please refer to the following detailed descriptions and drawings of the present invention. However, the drawings provided are only for reference and description, and are not intended to limit the present invention.

The following are specific embodiments to illustrate the embodiments of the present invention, and those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the concept of the present invention. In addition, the drawings of the present invention are merely schematic illustrations, and are not drawn according to the actual size, and are stated in advance. The following embodiments will further describe the related technical contents of the present invention in detail, but the disclosed contents are not intended to limit the protection scope of the present invention.

It should be understood that although terms such as "first", "second" and "third" may be used herein to describe various elements or signals, these elements or signals should not be limited by these terms. These terms are primarily used to distinguish one element from another element, or a signal from another signal. In addition, the term "or", as used herein, should include any one or a combination of more of the associated listed items, as the case may be.

The disclosure book discloses a method and system for intelligent detection of sensing data. One of the main purposes of the method is to use the deep learning algorithm for initial training and model building, and to cooperate with the machine learning algorithm to perform post-test and training. A detection model for detecting a specific object to be detected is developed. For example, the object to be detected can be a product, a system or a field, and sensors are used to sense specific information of the object to be detected, such as image, sound, vibration, and the obtained sensing data such as photographing the object to be detected The resulting image, the frequency spectrum obtained by recording the sound produced by the object to be detected, or the information of the vibration generated when the object to be detected is detected, these sensing data can be used to detect various objects to be detected. Control parameters that drive the operation of the object to be detected.

For an embodiment of realizing the intelligent detection system for sensing data, reference may be made to the system architecture embodiment diagram shown in FIG. 1 .

A to-be-detected object 10 is shown in the figure. The to-be-detected object is driven by control parameters. The sensing data intelligent detection system can optimize its own control parameters by learning the sensing data, or optimize the system that generates the to-be-detected object. control parameters. For example, the object to be detected 10 detects the object to be detected 10 through one or more sensors (sensor one 101, sensor two 102 and sensor three 103) to obtain sensing data. Taking the image generated by the object 10 to be detected as an example, the first sensor 101, the second sensor 102, and the third sensor 103 can be cameras that shoot from multiple angles; if the sensor is used to sense the environmental conditions of a certain field, the sensor one 101. The second sensor 102 and the third sensor 103 may be sensors located at different positions in the field. The sensing data generated by the first sensor 101 , the second sensor 102 and the third sensor 103 can be stored in the sensing data processing host 12 , and then converted and preliminarily processed into data in the database 145 in the computer system 14 .

The computer system 14 is provided with a processor 141, a memory 143 and a database 145, wherein the memory 143 is used for storing the program sets and algorithms for executing the intelligent detection method of the sensory data, and the database 145 is used for storing one or more sensors Sensing data obtained after measuring the object 10 to be detected, and then executing the intelligent detection method of the sensing data by the processor 141 of the computer system 14 .

Using the above-mentioned system architecture, FIG. 2 shows an embodiment of the main flow of the method for intelligent detection of sensing data executed by the system.

At the beginning of the system, as in step S201, the sensing data obtained from the object to be detected is obtained, and then as in step S203, the sensing data is divided into training data (training data) and testing data (testing data), in a specific embodiment , and the validation data can be obtained from the training data. In step S205, a deep-learning algorithm is executed on the training data, and features about the object to be detected can be obtained from the training data, and these features reflect information about the operation of the object to be detected or the operation of a related system, A detection model for detecting the object to be detected is established by learning from it, which is a detection model based on deep learning.

Then, in step S207, the detection model obtained in the above steps is tested with the test data selected from the sensing data, and the detection result includes the part that passed the test and the part that failed the test. The data passing the test indicates that the detection model is in line with expectations, indicating that the model parameters derived from the current detection model are appropriate and do not need to be adjusted. However, if the detection result includes the part that fails the test, in the intelligent detection method for sensing data, as in step S209, a machine learning algorithm can be executed for the test data that fails, and the above steps can be trained with historical parameter data The obtained detection model can be trained with historical parameter data to obtain a parameter model, and then a model optimization algorithm is used to optimize the detection model, which is a detection model based on machine learning. As in step S211, the optimized detection model can be obtained. Control parameters.

Machine learning methods cover many, such as Support Vector Machine, Random Forest, Naïve Bayes, and Deep Learning algorithms, to name a few. The above-mentioned deep learning algorithm is a kind of machine learning method. Using the computing power of the processor in the computer system, the obtained large amount of sensing data can be passed through linear or non-linear processing in multiple processing layers (layers). Linear or non-linear transformation is a feature extraction step to obtain features in the data that represent the characteristics of the sensing data about the object to be detected.

Common deep learning algorithms use two methods: Convolution Neural Networks (CNN) or Recurrent Neural Network (RNN). Convolutional neural networks can be used in their convolutional layers. The sensory data is filtered by convolution operation, and the features are gradually extracted by the computing power of the computer processor, so as to finally establish a model (such as the above detection model), which can be used to detect the sensory data generated by the system.

For example, the sensing data obtained from the object to be detected, such as image data, sound data, vibration data, etc., in the method, the system can set a part (eg, 90%) of the sensed data as Training data, the training data is used to establish a detection model through deep learning algorithms (such as CNN, RNN), and the remaining part of the sensing data (such as 10%) is set as testing data (testing data) to test the system The established detection model.

FIG. 3 shows a process embodiment of a method for intelligent detection of sensing data in system operation.

At the beginning of the process, sensing data ( 301 ) can be generated from the object to be detected, and the sensing data ( 303 ) for intelligent detection and establishment of the sensing data model, the sensing data or a part thereof (such as training obtained from the sensing data) data) is used to execute a deep learning algorithm (305) to build a detection model based on deep learning, and then test the detection model (307) with partial data (such as test data obtained from sensing data), For the part that does not require subsequent parameter optimization, the process ends ( 309 ); for the part that fails the test, a parameter model is established ( 311 ).

In the process of establishing the parameter model (311), machine learning is performed again on the data that has not passed the test, the historical parameter data is obtained from the historical parameter database (313) in the system, the model is trained on the parameters, the parameter model is established, and the A model optimization algorithm (315) optimizes the model parameters of the detection model, forms a detection model based on machine learning, and finally generates optimized control parameters (317), which can be imported into the object to be detected again and continue to generate sensing data (301).

It is mentioned here that the control parameter is a parameter that drives the operation of the object to be detected, or an operation parameter of a system that generates the object to be detected. In particular, in one embodiment, validation data can also be obtained from the training data, and when the training data is used to form a detection model through a deep learning algorithm, the validation data can be used to verify the detection model to obtain A parameter model for generating control parameters is then used to optimize the parameter model with a model optimization algorithm (315), and the optimized control parameters are output.

In this way, the optimized control parameters ( 317 ) are driven to the object to be detected, and the sensing data ( 301 ) is obtained again, and the control parameters of the object to be detected are continuously optimized by repeating the intelligent detection method of the sensing data until the system presets an expected state.

FIG. 4 then shows another example of the flow of the system operation. This embodiment shows that the continuously generated sensing data ( 303 ) stored in the system will become a historical sensing database ( 400 ), and these historical sensing databases 400 continue to perform deep learning ( 305), using the computing power of the computer system to continuously train the data and optimize the detection data. On the other hand, after testing the detection model, the parameters in the tested data can become part of the historical parameter database (313), and machine learning is performed continuously and repeatedly to optimize the parameter model.

According to an embodiment, when executing the model optimization algorithm, various parameter models can be used, such as Genetic Algorithms, Particle Swarm Optimization, Ant Colony Optimization, One of Simulated Annealing, Tabu Search, Seagull Optimization Algorithm and Bayesian Optimization. The optimized parameter models such as Gradient Boosting Decision Tree, Extreme Gradient Boosting, Categorical boosting, Light GBM, Random Forest Forest), Support Vector Machine, Relevance vector machine, Naïve Bayes, K nearest neighbor, CNN or RNN . In a preferred embodiment, the model optimization algorithm adopted by the system can adopt Bayesian Optimization algorithm, the deep learning algorithm can adopt CNN model, and the parameter model can adopt extreme gradient boosting model (Extreme Gradient Boosting). Gradient Boosting).

According to one of the embodiments, a K-fold cross-validation method may be used in the intelligent detection method of sensing data, and the verification method may be to divide the training data into multiple groups (K) of data, One set of data is used as the data for validating the model, and the other sets of data (K−1) are used as training data. After that, K sets of data are rotated for K times of validation, and finally the parameters of the evaluation model can be obtained. Among them, the K-fold cross-validation method is used to evaluate multiple models generated from multiple deep learning algorithms from the validation data, and then select a detection model from them, and one of the ways to evaluate the multiple models is to obtain an accuracy of each model. factors such as accuracy, precision, and recall, including evaluation equations derived from these factors.

For example, in the K-fold cross-validation method, taking 10-fold cross-validation as an example, during operation, the acquired sensing data is divided into 10 equal parts, and the first score data can be used as a test The test data for the model, the remaining 9 points can be used for training data. Then, in the next round of the process, the second point data will be used as the data for testing the detection model, and the remaining 9 points will also be used for training data. By analogy, a total of 10 rounds of processes can be performed, an accuracy rate of 10 can be obtained, and an average value can be obtained to obtain a more objective accuracy rate.

In one embodiment, the correlation between the predicted result of each detection model is yes or no (YES/NO) and the actual detection result obtained by actually measuring the object to be detected is yes or no (YES/NO). The "Accuracy", "Precision" and "Recall" are obtained as the basis for evaluating the detection model.

For example, the data whose model prediction result is YES (YES) and the actual detection result is also YES (YES) are set to "TP"; the model prediction result is YES (YES) and the actual detection result is NO (NO) (indicating the prediction error) is set to "FP"; data predicted by the model is NO (NO) and the actual detection result is YES (YES) (indicating a prediction error) is set to "FN"; model prediction is NO (NO) and actual Data for which the test is also NO (NO) is set to "TN".

According to the above settings, the "accuracy" indicates the correct proportion of the model's prediction. The mathematical formula is: "Accuracy=(TP+TN)/N _totoal ", that is, the model prediction result is true and the actual detection result is also true. The data ("TP") plus the data for which the model predicted no and the actual test was also no ("TN") divided by the total data (N _total ).

The "accuracy" is one of the main goals for the entire system to optimize the detection model. The mathematical formula is: "Precision=TP/(TP+FP)", that is, the actual detection result is yes and the model prediction is also the number of yes ("TP") accounts for the proportion of the total number (i.e. "TP" + "FP") that the model predicted as yes.

The "recall rate" represents the error rate of the model, and the mathematical formula is: "Recall=TP/(TP+FN)", that is, the data for which the model prediction result is yes and the actual detection result is also yes ("TP") It accounts for the proportion of all data (ie "TP" + "FN") that are actually detected as yes.

Further, other operations (such as F1 Score=2/((1/Precision)+(1/Recall))) can be performed on the calculation results of the above "Accuracy", "Precision" and "Recall" as an evaluation The basis for the pros and cons of the model.

FIG. 5 shows a schematic diagram of an embodiment of the operation of training data, verification data and test data in the method for intelligent detection of sensing data. The following description will be described with reference to the flowchart of the embodiment of the method for intelligent detection of sensing data shown in FIG. 6 .

The database 50 in the system is provided with sensing data and parameter data (step S601 ). The sensing data can be mainly divided into training data 501 and test data 505 , and a part of the training data 501 can be used as the verification data 503 (step S603 ). Based on the training data 501, a deep learning algorithm is used to train the model (52) to obtain a detection model (step S605). The test model ( 56 ) is performed with the test data 505 (step S607 ) to derive a parameter model for generating control parameters. At this time, the control parameters may be generated by the machine learning model (step S609). On the other hand, the verification data 503 is to perform model verification to obtain a parametric model (step S611 ).

According to an embodiment, the parameter model can be optimized by the model optimization algorithm ( 53 ) (step S613 ), and the parameter model can be continuously tested with the test data (step S615 ), and then the detection model ( 54 ) is formed. After repeating the above process, a plurality of detection models can be obtained by introducing different deep learning algorithms and repeatedly optimizing the training data 501 , the verification data 503 and the test data 505 to generate control parameters (step S617 ). After comparing the models 57 , a model suitable for the system is determined according to the requirements ( 58 ), and the intelligent detection process of the sensing data is completed ( 59 ).

For the above-mentioned method of optimizing the model, reference may be made to the exemplary flowchart of optimization in the method for intelligent detection of sensing data shown in FIG. 7 .

According to the above embodiment, in the process of using the model optimization algorithm to optimize the model to define the optimal detection model, the historical parameter data can be imported (step S701 ), and the data can be trained by the machine learning algorithm to establish the parameter model (step S703 ), Next, the control parameters are obtained from this parameter model, and the K-fold cross-validation method is used one by one (step S705 ), and the validation data is used to test multiple models generated from multiple deep learning algorithms (step S707 ), and then evaluate from them (for example, using The above factors such as "accuracy", "precision" and "recall rate") and select the most suitable detection model (step S709). Then, the model parameter optimization algorithm is performed on the selected model (step S711 ) to obtain an optimized detection model (step S713 ), and the detection model can be used to generate optimal control parameters.

For an example, please refer to the flowchart shown in FIG. 8 , which shows an intelligent method for establishing a detection model for detecting a motor.

Acquire the vibration sound of the motor through the sound sensor (step S801 ), convert the audio into a frequency spectrum, create data in the sensing database (step S803 ), and then train the sensing data with a deep learning algorithm (such as the training obtained in the data), extract the sound features related to the motor operation status in the data, obtain the correlation of various audio frequencies related to the abnormal operation of the motor and its control parameters, and establish a detection model for detecting the motor (step S805). Afterwards, the system uses the verification data obtained from the sensing data to verify the above-mentioned training to obtain a detection model (step S807 ), and determines the detection model after verification (step S809 ).

The system uses the detection model to generate control parameters, drive the motor, and record sound through the sensor at the same time to form audio sensing data, and then detect the sensing data (such as an imaged spectrum) (step S811 ). Does the data judge pass the test? (step S813 ), if the test is passed, it means that driving the motor with the control parameters determined by the current detection model meets the system requirements, and the process ends (step S815 ).

If it does not pass the test (No), it means that it is necessary to continuously optimize the parameters with the parameter model (step S817 ), and then use the generated control parameters to drive the motor to operate (step S819 ), and continue to use the sensor to obtain the vibration sound. , according to the detection data of the detection model to determine whether the detection is passed, and the above steps are repeated to continuously optimize the parameter model to obtain the optimal control parameters.

Further, according to another embodiment of the application, the disclosed intelligent detection method of sensing data is applied to the detection screen to display the parameters to drive the screen to display the content. As in the above process, the image sensor can be used to obtain images of a large number of screens first. Data, from which the features in the image data are obtained by deep learning algorithms, the correlation between the image and the screen display parameters can be established, and the detection model of the detection screen can be established, including the verification of the verification data to determine the detection model, and then use this detection model to generate a display parameters, and then generate the image data of the screen for testing, and execute the subsequent machine learning algorithm to optimize the detection model, so as to obtain the optimized display parameters of the screen.

To sum up, according to the method and system for intelligent detection of sensory data described in the above embodiments, in particular, when building a model from training data in the method, in addition to using a part of the sensory data as training data (training data) , and part of it is used as testing data for the model, and part of it is obtained from the training data as validation data to verify the detection model generated by the system. This is an internal optimization step. What's more, when training the model, a variety of deep learning and machine learning algorithms can be used to generate a variety of models. After these models can be optimized, the models can be compared and evaluated, and then one of the models can be selected for practical application.

The contents disclosed above are only preferred feasible embodiments of the present invention, and are not intended to limit the scope of the present invention. Therefore, any equivalent technical changes made by using the contents of the description and drawings of the present invention are included in the application of the present invention. within the scope of the patent.

10: Object to be detected 101: Sensor one 102: Sensor two 103: Sensor three 12: Sensing data processing host 14: Computer System 141: Processor 143: Memory 145:Database 301: Generate sensing data 303: Sensing data 305: Deep Learning Algorithms 307: Test Detection Model 309: End 311: Parametric Models 313: Historical parameter database 315: Model Optimization Algorithms 317: Generate control parameters 400: Historical sensing database 50: Database 501: training data 503: Validate data 505: Test data 52: Train the model 53: Model Optimization Algorithms 54: Detection Model 56: Test the model 57: Compare Models 58: Judgment Model 59: Done Steps S201-S211: the main flow of the intelligent detection method for sensing data Steps S601-S617: the flow of the intelligent detection method for sensing data Steps S701-S713: Optimization process in the intelligent detection method of sensing data Steps S801-S819: flow example of the method for intelligent detection of sensing data

FIG. 1 shows an embodiment diagram of a system architecture of a sensing data intelligent detection system;

FIG. 2 shows a main process embodiment diagram of a method for intelligent detection of sensing data;

FIG. 3 shows one of the process embodiments of the method for intelligent detection of system operation sensing data;

FIG. 4 shows the second embodiment of the process of the method for intelligent detection of system operation sensing data;

5 shows a schematic diagram of an operation embodiment of training data, verification data and test data in the method for intelligent detection of sensing data;

FIG. 6 shows a flowchart of an embodiment of a method for intelligent detection of sensing data;

FIG. 7 shows an exemplary flow chart of optimization in the method of intelligent detection of sensory data; and

FIG. 8 shows an exemplary flow chart of a method for intelligent detection of applied sensing data.

301: Generate sensing data

303: Sensing data

305: Deep Learning Algorithms

307: Test Detection Model

309: End

311: Parametric Models

313: Historical parameter database

315: Model Optimization Algorithms

317: Generate control parameters

Claims (7)

An intelligent detection method for sensing data, comprising: obtaining sensing data of an object to be detected, and distinguishing the sensing data into training data and test data; executing a plurality of deep learning algorithms on the training data, Obtain the characteristics of the object to be detected, obtain verification data from the training data, use the verification data to evaluate a plurality of models generated from the plurality of deep learning algorithms by a K-fold cross-validation method, and then select the Detecting a detection model of the object to be detected; testing the detection model with the test data; and importing historical parameter data for the test data that fails the test result, executing a machine learning algorithm, and training with the machine learning algorithm These data establish a parameter model, then use the parameter model to obtain control parameters, train the detection model with historical parameter data, and optimize the detection model with a model optimization algorithm to obtain the optimized control parameters of the object to be detected ; wherein the object to be detected is driven by the control parameter to operate, and the control parameter of the object to be detected can be optimized by learning the sensing data of the object to be detected. The intelligent detection method for sensing data according to claim 1, wherein the parameter model is further optimized by the model optimization algorithm, and the optimized control parameter is generated. The intelligent detection method for sensing data according to claim 2, wherein the optimized control parameter is then driven to drive the object to be detected to obtain sensing data again, and the intelligent detection method for sensing data is optimized by repeating the intelligent detection method for sensing data. The control parameter of the object to be detected. The intelligent detection method for sensing data as claimed in claim 1, wherein the basis for evaluating the detection model comprises the multiple data generated from the multiple deep learning algorithms. From each model, an accuracy, a precision and a recall rate of each model are obtained. An intelligent detection system for sensing data, comprising: a computer system with a memory, which stores a program set and an algorithm for executing an intelligent detection method for sensing data, the computer system is also provided with a database, which is stored from a to-be-detected Sensing data obtained by an object; wherein a processor of the computer system executes the sensing data intelligent detection method, comprising: obtaining the sensing data, distinguishing the sensing data into training data and test data; Execute a plurality of deep learning algorithms on the data, obtain features about the object to be detected from the training data, obtain verification data from the training data, and use the verification data to evaluate the data from the plurality of deep learning by a K-fold cross-validation method. A detection model for detecting the object to be detected is selected from a plurality of models generated by the algorithm; the detection model is tested with the test data; the historical parameter data is imported for the test data for which the test result fails, and a A machine learning algorithm, using the machine learning algorithm to train the data to establish a parameter model, then using the parameter model to obtain control parameters, training the detection model with historical parameter data, and optimizing the detection model with a model optimization algorithm, to obtain the optimized control parameters of the object to be detected; wherein the object to be detected is driven by the control parameters to operate, and the control parameters of the object to be detected can be optimized by learning the sensing data of the object to be detected. The intelligent detection system for sensing data according to claim 5, wherein the parameter model is further optimized by the model optimization algorithm, the optimized control parameters are generated, and the object to be detected is continuously driven to obtain the sensing data, by repeating The described sensing data intelligent detection method optimizes the control parameter of the object to be detected. The sensing data intelligent detection system as claimed in claim 5 or 6, wherein the basis for evaluating the detection model comprises deriving an accuracy of each model from the multiple models generated by the multiple deep learning algorithms, One precision and one recall.

Images