CN113688853A

CN113688853A - Intelligent detection method and system for sensing data

Info

Publication number: CN113688853A
Application number: CN202010426051.6A
Authority: CN
Inventors: 李荣生; 高于胜; 简嘉宏; 王薇婷
Original assignee: Zhilue Information Integration Co ltd
Current assignee: Data Systems Consulting Co Ltd
Priority date: 2020-05-19
Filing date: 2020-05-19
Publication date: 2021-11-23

Abstract

In the intelligent sensing data detecting method, when the system obtains sensing data, the sensing data is divided into training data and testing data, a deep learning algorithm is executed on the training data to obtain the characteristics of the object to be detected, a detection model based on the deep learning is established, the detection model is tested by the testing data, then a machine learning algorithm is executed aiming at the failed testing data, the detection model is trained by historical parameter data, and the detection model is optimized by a model optimizing algorithm to obtain the optimized control parameters of the object to be detected.

Description

Intelligent detection method and system for sensing data

Technical Field

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

Background

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, capturing images, etc., and then analyze information in the sensed data by an analysis tool, thereby determining whether the object to be detected is abnormal.

For example, when determining whether there is a defect on the surface of an object, the camera can be used to photograph 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 an abnormality. If the motor is taken as an example, the general method is that the sound sensor records the audio signal generated when the motor runs, and whether the operation is abnormal can be judged after comparing the sound sample plate, and the judgment is used as the basis for improvement in the future.

Disclosure of Invention

The description discloses a method and a system for intelligently detecting sensing data, wherein the method is operated in a computer system, the computer system is provided with a memory, a program set and an algorithm for executing the intelligent detection method of the sensing data are stored in the memory, and the computer system is also provided with a database, wherein the database stores the sensing data obtained from an object to be detected.

In one embodiment, the sensed data is data obtained by sensing an object to be detected through a sensor, in the sensed data intelligent detection method, the sensed data is firstly divided into training data and test data, a deep learning algorithm is executed on the training data, characteristics of the object to be detected are obtained from the training data, a detection model for detecting the object to be detected is established, then the detection model is tested through the test data, a machine learning algorithm is executed according to the test data with a non-passing test result, the detection model is trained through historical parameter data, and the detection model is optimized through a model optimization algorithm to obtain control parameters of the optimized object to be detected.

Further, verification data is obtained from the training data, and when the detection model is formed from the training data through a deep learning algorithm in the above step, the detection model can be verified by the verification data to obtain a parameter model for generating the control parameters. The control parameter is a parameter for driving the object to be detected to operate, and the generated optimized control parameter is one of the main purposes of the method.

Further, the parameter model can be optimized through a model optimization algorithm, optimized control parameters are generated, the object to be detected is continuously driven, the steps are repeated, the sensing data are obtained again, and the control parameters of the object to be detected are optimized through an intelligent sensing data repeated detection method.

Further, in one embodiment, the models generated from the deep learning algorithms and the machine learning method may be evaluated using a K-fold cross-validation method using the validation data, for example, by obtaining the accuracy, precision, and recall of each model from the models generated from the deep learning algorithms and selecting the detection model according to the score.

Drawings

FIG. 1 illustrates a diagram of a system architecture embodiment of a sensed data smart detection system;

FIG. 2 is a diagram illustrating a main flow embodiment of a method for intelligent detection of sensed data;

FIG. 3 illustrates one embodiment of a flow of a system operating a method for intelligent detection of sensed data;

FIG. 4 illustrates a second embodiment of a system operation sensing data intelligent detection method;

FIG. 5 is a schematic diagram of an embodiment of operations of training data, verification data and test data in the method for intelligently detecting sensing data;

FIG. 6 illustrates a flow diagram of an embodiment of a sensed data smart detection method;

FIG. 7 illustrates an example flow diagram of optimization in a sensed data smart detection method; and

FIG. 8 illustrates an example flow diagram of an application-sensed data smart detection method.

Detailed Description

The specification discloses an intelligent detection method and system for sensed data, and the method mainly aims to utilize a deep learning algorithm to train and establish a model at the initial stage, perform later-stage test and training by matching with a machine learning algorithm, and train and obtain a detection model for detecting a specific object to be detected from the sensed data. For example, the object to be detected may be a product, a system or a field, the sensor is used to sense specific information of the object to be detected, such as image, sound and vibration, the obtained sensing data is, for example, an image obtained by photographing the object to be detected, a frequency spectrum obtained by recording sound generated by the object to be detected and converting the sound, or information of vibration generated when the object to be detected operates, and the sensing data may be information for detecting various objects to be detected, so as to optimize control parameters for driving the object to be detected to operate.

An embodiment of implementing the intelligent sensing data detection system can refer to a diagram of an embodiment of a system architecture shown in fig. 1.

The figure shows an object to be detected 10, which is driven by control parameters to operate, and the sensed data intelligent detection system can optimize the control parameters of the sensed data or the system generating the object to be detected by learning the sensed data. For example, the object 10 to be detected senses the object 10 to be detected through one or more sensors (the first sensor 101, the second sensor 102 and the third sensor 103) to obtain sensing data, and if the image of the object 10 to be detected is captured to generate an image, the first sensor 101, the second sensor 102 and the third sensor 103 may be cameras for capturing multiple angles; if the environment condition of a certain field is sensed, the first sensor 101, the second sensor 102 and the third sensor 103 may be sensors disposed at different positions in the field. The sensed data generated by the first sensor 101, the second sensor 102 and the third sensor 103 can be stored in the sensed data processing host 12, and then converted and primarily processed into data in the database 145 of the computer system 14.

The computer system 14 includes a processor 141, a memory 143, and a database 145, wherein the memory 143 is used for storing a program set and an algorithm for performing the intelligent sensing method, the database 145 is used for storing the sensing data obtained by measuring the object 10 to be detected with one or more sensors, and the processor 141 of the computer system 14 is used for performing the intelligent sensing method.

With the above system architecture, fig. 2 is a diagram illustrating a main flow embodiment of the intelligent sensing data detection method executed by the system.

The system starts with obtaining the sensing data obtained from the object to be detected in step S201, and then divides the sensing data into training data (training data) and testing data (testing data) in step S203. In step S205, a deep-learning algorithm (deep-learning algorithm) is performed on the training data, such that features of the object to be detected, which reflect the operation of the object to be detected or the operation of related systems, are obtained from the training data, and a detection model for detecting the object to be detected is established by learning from the features, which is a detection model based on deep learning.

Then, in step S207, the detection model obtained in the above step is tested with the selected test data in the sensing data, and the detection result includes a part passing the test and a part failing the test. The data passed the test indicates that the test model is expected, indicating that the model parameters derived from the current test model are appropriate and not adjusted. However, if the detection result includes a part failing the test, in the intelligent sensing data detection method, in step S209, a machine learning algorithm may be performed on the failed test data, and the detection model obtained in the above steps may be trained using historical parameter data, and the parameter model may be trained using historical parameter data, and then the detection model may be optimized using a model optimization algorithm, which is a machine learning-based detection model, in step S211, the control parameters of the object to be detected that are optimized may be obtained.

The machine learning (machine learning) method covers many fields, such as Support vector machine (Support vector machine), Random Forest (Random Forest), simple Bayes (Bayes)

Bayes), and deep learning algorithms, etc. The deep learning algorithm is one of machine learning (machine learning) methods, and a large amount of sensing data obtained by using the computing power of a processor in a computer system may be processed through linear or non-linear transformation (linear or non-linear transformation) in a plurality of processing layers (layers) to obtain features representing characteristics of sensing data about an object to be detected in the data by a feature extraction (feature extraction) step.

The conventional deep learning algorithm adopts two methods, namely, a Convolutional Neural Network (CNN) method and a Recurrent Neural Network (RNN) method, the convolutional Neural network can filter sensing data by using Convolution operation in a convolutional layer (RNN), and features in the convolutional Neural network are gradually extracted by calculation of a computer processor so as to finally establish a model (such as the detection model) which can be used for detecting the sensing data generated by the system.

For example, in the method, the system may set a part (e.g., 90%) of the sensed data as training data (training data), the training data establishes a detection model through a deep learning algorithm (e.g., CNN, RNN), and the remaining part (e.g., 10%) of the sensed data is set as test data (testing data) for testing the detection model established by the system.

FIG. 3 illustrates a flow embodiment of a system operation sensed data smart detection method.

At the beginning of the process, sensing data (301) can be generated from the object to be detected, and the sensing data (303) of the sensing data smart detection modeling is formed, the sensing data or a part thereof (such as training data obtained from the sensing data) is used for executing a deep learning algorithm (305) to build a detection model based on deep learning, and then the detection model is tested (307) with a part of data (such as test data obtained from the sensing data), and the process is ended (309) for the part passing the test, i.e. the process without subsequent parameter optimization; for the part failing the test, a parametric model is built (311).

In the process of establishing a parameter model (311), machine learning is carried out on the data which do not pass the test again, historical parameter data are obtained from a historical parameter database (313) in the system, model training is carried out on the parameters, the parameter model is established, model parameters of the detection model are optimized through a model optimization algorithm (315), the detection model based on the machine learning is formed, optimized control parameters (317) are finally generated, the optimized control parameters can be remitted into the object to be detected again, and sensing data (301) are continuously generated.

It is noted that the control parameter is a parameter for driving the object to be detected to operate, or an operation parameter of a system for generating the object to be detected. Specifically, in one embodiment, validation data (validation data) may be obtained from the training data, and the validation data may be used to validate the detection model when the training data is used to form the detection model by the deep learning algorithm to obtain a parameter model for generating the control parameters, and then the model optimization algorithm (315) is used to optimize the parameter model before outputting the optimized control parameters.

In this way, the optimized control parameters (317) are used for driving the object to be detected, the sensing data (301) is obtained again, and the control parameters of the object to be detected are continuously optimized by repeating the sensing data intelligent detection method until a system defaults to a desired state.

FIG. 4 next illustrates another embodiment of the system operation flow, which shows that the system stores the continuously generated sensing data (303) as historical sensing databases (400), and these historical sensing databases 400 continue to perform deep learning (305) to optimize the detection data by continuously training the data with the computer system's computing power. On the other hand, after the test detection model is tested, parameters in the test data can become a part of a historical parameter database (313), and machine learning is continuously and repeatedly carried out to optimize the parameter model.

According to an embodiment, in executing the model Optimization Algorithm, a variety of parametric models may be employed, such as one of Genetic Algorithms (Genetic Algorithms), Particle Swarm Optimization (Particle Swarm Optimization), Ant Colony Optimization (Ant Colony Optimization), Simulated Annealing (Simulated Annealing), restricted Search (scratch Search), gull Optimization (Seagull Optimization Algorithm), and Bayesian Optimization (Bayesian Optimization). And the optimized parameter models include Gradient Boosting Decision Tree (Gradient Boosting Decision Tree), Extreme Gradient Boosting (Extreme Gradient Boosting), class Boosting (category Boosting), Light GBM (Light GBM), Random Forest (Random Forest), and Support Vector (SVM)Model (Support Vector Machine), relational Vector Machine (Relevance Vector Machine), and simple Bayesian classification model (A/D) ((A/D))

Bayes), K nearest neighbor algorithm model (K nearest neighbor), CNN model, or RNN model. In a preferred embodiment, the model Optimization algorithm used by the system may be a Bayesian Optimization (Bayesian Optimization), the deep learning algorithm may be a CNN model, and the parametric model may be an Extreme Gradient Boosting (Extreme Gradient Boosting).

According to one embodiment, the intelligent sensing data detection method may adopt a K-fold cross validation method (K-fold cross validation), and the validation method may divide training data into a plurality of groups (K) of data, where one group of data is used as data of the validation model, and other groups of data (K-1) are used as training data, and then K times of validation are performed for K groups of data in turn, so as to finally obtain parameters of the evaluation model. The K-fold cross-validation method is used for evaluating a plurality of models generated from a plurality of deep learning algorithms from validation data and selecting a detection model from the plurality of models, wherein one of the modes of evaluating the plurality of models is to obtain factors such as accuracy (accuracy), precision (precision) and recall (recall) of each model, and the factors comprise evaluation equations obtained by using the factors.

For example, the K-fold cross validation method, taking 10-fold (10-fold) cross validation as an example, divides the acquired sensing data into 10 equal parts during operation, wherein the 1 st part of data can be used as the test data for the test model, and the remaining 9 parts can be used for the training data. Then, the next round of the procedure was performed, and the data of the 2 nd part was used as data for the test detection model, and the remaining 9 parts were also used for training data. By analogy, 10 cycles of processes are executed in total, 10 accuracy rates can be obtained, and an average value is obtained, so that objective accuracy rates are obtained.

In one embodiment, the "accuracy", "precision" and "recall" can be derived from the correlation between YES/NO (YES/NO) of the predicted result of each test model and YES/NO (YES/NO) of the actual test result obtained by actually measuring the object to be tested, as the basis for evaluating the test models.

For example, data with a model prediction result of YES (YES) and an actual detection result of YES (YES) is set as "TP"; data with a YES model prediction result and NO actual detection result (NO) (indicating a prediction error) is set as "FP"; data in which the model prediction result is NO (NO) and the actual detection result is YES (YES) (indicating a prediction error) is set to "FN"; the data for model prediction NO (NO) and actual test NO (NO) is set to "TN".

According to the above setting, "accuracy" represents the correct ratio of model prediction, and the mathematical formula is: "Accuracy ═ TP + TN)/N_total", that is: data ("TP") with a model prediction result of yes and an actual test result of yes plus data ("TN") with a model prediction of no and an actual test of no divided by the total data (N)_total)。

The "accuracy" is one of the main objectives of the whole system to optimize the detection model, and the mathematical formula is: "Precision ═ TP/(TP + FP)", i.e., the number of actual detections that are yes and the model predictions that are also yes ("TP") is a proportion of the total number of model predictions that are yes (i.e., "TP" + "FP").

The recall rate represents the error rate of the model, and the mathematical expression is as follows: "Recall ═ TP/(TP + FN)", that is: the data ("TP") that the model predicts is and the actual test is also true is a proportion of the total data actually tested is (i.e., "TP" + "FN").

Further, the calculation results of the above "accuracy", "Precision" and "Recall" may be further calculated (e.g., F1 Score ═ 2/((1/Precision) + (1/Recall))) as the basis for evaluating the model.

Fig. 5 is a schematic diagram illustrating an embodiment of operations of training data, verification data and test data in the intelligent sensing data detection method, and the following description is made in conjunction with a flowchart of an embodiment of the intelligent sensing data detection method illustrated 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, the model is trained with a deep learning algorithm (52), resulting in a detection model (step S605). The test model (56) is performed according to the test data 505 (step S607), so as to obtain a parametric model for generating the control parameters. At this time, the control parameters may be generated with a machine learning model (step S609). On the other hand, the verification data 503 performs 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 test data can be continued to test the parameter model (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 the control parameters (step S617), and then the comparison model 57 can be used to determine the model (58) suitable for the system according to the requirement, and the sensed data intelligent detection process (59) is completed.

The method for optimizing the model may refer to an exemplary flowchart of the optimization in the intelligent sensing method of fig. 7.

In the process of optimizing the model by using the model optimization algorithm to define the optimal detection model according to the above embodiment, historical parameter data may be imported (step S701), the data are trained by the machine learning algorithm to build the parameter model (step S703), the control parameters are obtained by the parameter model, the multiple models generated by the multiple deep learning algorithms are tested by the verification data (step S707) one by one in the K-fold cross-validation method (step S705), and then the most suitable detection model is evaluated (for example, by using the above-mentioned factors such as "accuracy", "precision", and "recall ratio") and selected (step S709). A model parameter optimization algorithm is then performed on the selected model (step S711) to derive an optimized detection model (step S713), which may be used to generate optimal control parameters.

For an example, reference may be made to the flow chart of FIG. 8, which illustrates an intelligent method of establishing a test model for a test motor.

The method comprises the steps of obtaining vibration sounds of a motor through a sound sensor (step S801), converting audio signals into frequency spectrums, establishing data in a sensing database (step S803), training the sensing data (such as training data obtained from the sensing data) through a deep learning algorithm, extracting sound characteristics related to the operation condition of the motor from the data, obtaining whether various audio frequencies are abnormal or not related to the operation of the motor and the relevance of control parameters of the audio frequencies, and establishing a detection model for detecting the motor (step S805). Then, the system verifies the detection model obtained by the training in the above step using the verification data obtained from the sensing data (step S807), and determines the detection model after verification (step S809).

The system generates control parameters by using the detection model, drives the motor, records sound through the sensor to form audio sensing data, and then detects the sensing data (such as imaged spectrum) (step S811), and determines from the sensing data whether the test is passed? (step S813), if the test result shows that the driving of the motor with the control parameters determined by the current detection model is satisfactory to the system, the process is ended (step S815).

If the test is not passed (no), it indicates that the parameter model needs to be continuously optimized (step S817), the generated control parameter is used to drive the motor to operate (step S819), the sensor is continuously used to obtain the vibration sound, after the frequency spectrum is converted, the detection model detection data is used to judge whether the detection is passed, and the above steps are repeated to continuously optimize the parameter model to obtain the optimal control parameter.

Further, according to another embodiment of the present invention, the disclosed method for intelligently detecting sensing data is applied to a detection screen to show the parameter driving screen to display the content, and as described above, the image sensor can first obtain a large amount of image data of the screen, and then obtain the features in the image data by a deep learning algorithm, and establish the correlation between the image and the screen display parameters, so as to establish a detection model of the detection screen, including determining the detection model by verifying the data, then generating the display parameters by using the detection model, and then generating the image data of the screen, performing a test, and performing a subsequent machine learning algorithm to optimize the detection model, so as to obtain the optimized display parameters of the screen.

In summary, according to the method and system for intelligently detecting sensing data described in the above embodiments, particularly, when the training data is modeled in the method, in addition to a part of the sensing data as training data (training data), a part of the sensing data also serves as data (testing data) of the test model, and a part of the testing data is obtained from the training data as verification data (verification data) to verify the detection model generated by the system, which is an internal optimization step. Furthermore, when the models are trained, various deep learning and machine learning algorithms can be adopted to generate various models, and after the models are optimized, one of the models is selected to be put into practical application after model comparison and evaluation are carried out finally.

It should be understood that the above description is only a preferred embodiment of the present invention, and not intended to limit the scope of the present invention, so that equivalent structural changes made by using the description and drawings of the present invention are included in the scope of the present invention.

Claims

1. A method for intelligently detecting sensed data, the method comprising:

acquiring sensing data of an object to be detected, and dividing the sensing data into training data and testing data;

executing a deep learning algorithm on the training data, acquiring the characteristics of the object to be detected from the training data, and establishing a detection model for detecting the object to be detected;

testing the test model with the test data; and

and aiming at the test data with the test result of failure, executing a machine learning algorithm, training the detection model by using historical parameter data, and optimizing the detection model by using a model optimization algorithm to obtain the optimized control parameters of the object to be detected.

2. The method as claimed in claim 1, wherein verification data is obtained from the training data, and the detection model is verified by the verification data when the training data is processed by the deep learning algorithm to form the detection model, so as to obtain a parameter model for generating the control parameter.

3. The intelligent sensed data detection method as claimed in claim 2, wherein the model optimization algorithm is further used to optimize the parametric model, and the optimized control parameters are further generated.

4. The intelligent sensing data detecting method as claimed in claim 3, wherein the optimized control parameters are used to drive the object to be detected, and the sensing data is obtained again, and the control parameters of the object to be detected are optimized by repeating the intelligent sensing data detecting method.

5. The intelligent sensing data detection method as claimed in any one of claims 1 to 4, wherein a plurality of models generated from a plurality of deep learning algorithms are evaluated by using the validation data in a K-fold cross validation method, and the detection model is selected therefrom.

6. The method of claim 5, wherein evaluating the basis for the detection model comprises deriving an accuracy, a precision, and a recall for each model from the plurality of models generated by the plurality of deep learning algorithms.

7. An intelligent sensed data detection system, the system comprising:

the computer system is provided with a memory, a database and a data processing module, wherein the memory stores a program set and an algorithm for executing the intelligent detection method of the sensing data;

wherein, a processor of the computer system executes the intelligent detection method of the sensing data, which comprises the following steps:

acquiring the sensing data, and dividing the sensing data into training data and testing data;

testing the test model with the test data;

8. The system of claim 7, wherein verification data is derived from the training data, and the detection model is verified by the verification data when the training data is processed by the deep learning algorithm to form the detection model, so as to obtain a parameter model for generating the control parameter.

9. The system of claim 8, wherein the model optimization algorithm is further used to optimize the parametric model, generate the optimized control parameters, and continue to drive the object to be detected, and obtain the sensing data again, and the control parameters of the object to be detected are optimized by repeating the sensing data smart detection method.

10. The system according to claim 8 or 9, wherein the verification data is used to evaluate a plurality of models generated from a plurality of deep learning algorithms by a K-fold cross-verification method, and the detection model is selected therefrom; wherein evaluating the basis for the detection model comprises deriving an accuracy, a precision, and a recall for each model from the plurality of models generated by the plurality of deep learning algorithms.