CN108898157A - The classification method of the radar chart representation of numeric type data based on convolutional neural networks - Google Patents
The classification method of the radar chart representation of numeric type data based on convolutional neural networks Download PDFInfo
- Publication number
- CN108898157A CN108898157A CN201810525151.7A CN201810525151A CN108898157A CN 108898157 A CN108898157 A CN 108898157A CN 201810525151 A CN201810525151 A CN 201810525151A CN 108898157 A CN108898157 A CN 108898157A
- Authority
- CN
- China
- Prior art keywords
- numerical data
- data
- convolutional neural
- neural network
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 78
- 238000000034 method Methods 0.000 title claims abstract description 72
- 238000012549 training Methods 0.000 claims abstract description 75
- 238000012545 processing Methods 0.000 claims description 19
- 238000011176 pooling Methods 0.000 claims description 7
- 238000007781 pre-processing Methods 0.000 claims description 7
- 238000010606 normalization Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000003062 neural network model Methods 0.000 claims description 2
- 238000000605 extraction Methods 0.000 abstract description 6
- 238000012360 testing method Methods 0.000 description 36
- 230000006870 function Effects 0.000 description 18
- 230000008569 process Effects 0.000 description 11
- 238000001514 detection method Methods 0.000 description 8
- 238000004422 calculation algorithm Methods 0.000 description 4
- 238000013101 initial test Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000011478 gradient descent method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 239000007788 liquid Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000004886 process control Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of classification method of the radar chart representation of numeric type data based on convolutional neural networks, and steps are as follows：Numeric type data carries out feature ordering；Optimal feature is selected using the data after feature ordering to combine；Optimal characteristics combined value type data are expressed as image data using radar map；Construct convolutional neural networks basic structure；Training convolutional neural networks model.The present invention passes through radar chart representation method, numeric type data is converted into image data, retain the information between data as far as possible again simultaneously, to make full use of the powerful ability in feature extraction of convolutional neural networks, by in image data between numeric type data topological features and deeper time feature extraction come out and be used to classify, compared with traditional data-driven classification method, the classification accuracy of numeric type data is effectively increased.
Description
Technical Field
The invention belongs to the technical field of pattern recognition, relates to a convolutional neural network classification technology, and particularly relates to a classification method for radar chart representation of numerical data based on a convolutional neural network.
Background
As an image recognition method, a convolutional neural network has attracted much attention in recent years. It is a deep feedforward neural network whose essence is to construct a large number of filters to extract features of the input data. As networks get deeper and deeper, the extracted features become more abstract. Finally, the input data is represented as a series of combinations of abstract features that have been translated, rotated, and scaled multiple times. In addition, the convolutional neural network has the characteristics of sparse connection, weight sharing and space or time sub-sampling. Sparse connections establish a spatial relationship of incomplete connections and reduce the number of training parameters by topology. Weight sharing can make the algorithm effectively avoid overfitting. The secondary sampling fully utilizes the local characteristics of the data, reduces the data dimension, optimizes the network structure and ensures that the displacement to a certain degree cannot cause deformation.
Based on the structure and the characteristics, the convolutional neural network achieves huge achievement in the aspect of computer vision. In 2012, the most impressive achievement was achieved in the ImageNet LSVRC competition. Models with convolutional neural network depth achieve an unprecedented low error rate compared to other algorithms. In the field of target detection, from R-CNN, R-CNN to YOLO, the structures of the target detection are becoming simpler and the detection capability is becoming stronger. In addition to this, convolutional neural networks have also been introduced into the field of speech recognition, where the original frequency domain features are extracted by it to express some variability speech.
In view of their superior performance, some researchers have considered the introduction of convolutional neural networks into the field of industrial data processing. In the field of industrial fault diagnosis. Its application can be generalized to two aspects. In a first aspect, a convolutional neural network and its improved algorithm are used for feature extraction and recognition. A second aspect is the use of a convolutional neural network as a classifier. Based on current research in the industry, it can be concluded that convolutional neural networks are an effective recognition or classification tool, but they typically process image data. If we want to use a convolutional neural network to process numerical data, we must first find a suitable way to represent the numerical data as image data. The image data can express hidden characteristics among original numerical data and can also enable the correlation among data characteristics to be expressed.
Te (tennessee eastman) is a benchmark process control case for an actual industrial process based on the research of the eastman chemical company. Up to now, the detection of faults in the TE process 21 has mostly focused on traditional data-driven methods like SVM, ELM, etc. The traditional data driving method is difficult to extract the characteristics of certain faults, so that the faults are difficult to detect, and the overall fault detection rate is low.
In view of the above, there is a need for improvements in the prior art.
Disclosure of Invention
The invention aims to solve the technical problem of overcoming the defect that the convolutional neural network is difficult to carry out feature extraction on numerical data in the prior art by utilizing the strong feature extraction capability of the convolutional neural network on image data, and provides a classification method for radar chart representation of the numerical data based on the convolutional neural network.
In order to solve the technical problem, the invention provides a classification method for radar chart representation of numerical data based on a convolutional neural network, which comprises the following steps:
s1, collecting and preprocessing numerical data, and establishing a training set:
each numerical data includes its characteristics and label;
s2, representing the numerical data in the data set obtained in the step S1 into image data, and establishing a picture numerical data set train _ x by using the obtained image data; the method comprises the following steps:
2.1, sequentially carrying out normalization processing on the characteristics of the numerical data in the training set;
wherein,a j-th dimension characteristic representing the i-th numerical type data,a j-dimension feature representing the normalized i-th numerical data; x is the number ofjFeatures of dimension j, max x, representing all numerical data in the training setjDenotes xjMaximum value of (1), minxjDenotes xjMinimum value of (1);
2.2, obtained according to step 2.1Utilizing a polar function to draw a radar map corresponding to numerical data, and utilizing the obtained radar map to establish a picture numerical data set train _ x;
s3, constructing a basic structure of the convolutional neural network;
s4, training by using the picture numerical data set train _ x obtained in the step S2 and the label thereof to obtain a convolutional neural network model;
and S5, representing the numerical data to be classified into image data and inputting the image data into the convolutional neural network model for classification.
The classification method of the numerical data radar chart representation based on the convolutional neural network is improved as follows:
the convolutional neural network comprises an input layer, two convolutional layers, two pooling layers, a full-connection layer and an output layer;
the activation function of the convolutional layer selects a sigmoid function, and the output layer adopts an RBF neural network for probability classification;
and selecting a root mean square error function as a final cost function of the convolutional neural network, and adjusting the weight W and the bias b by adopting back propagation of errors.
The classification method of the numerical data radar chart representation based on the convolutional neural network is further improved as follows:
the method for collecting and preprocessing numerical data in step S1 includes:
1.1, collecting numerical data, and performing characteristic sorting on the numerical data:
and 1.2, selecting an optimal feature combination by using the data after feature sorting, and establishing a training set according to numerical data corresponding to the optimal feature combination.
The classification method of the numerical data radar chart representation based on the convolutional neural network is further improved as follows:
the step 1.1 is to collect numerical data and perform characteristic sorting on the numerical data as follows:
(1a) collecting numerical data and establishing an initial training set S;
(1b) and (2) carrying out standardization processing on the numerical data in the initial training set S established in the step (1a), wherein the processing formula is as follows:
S＝{(x1，y1)，(x2，y2)，…，(xN，yN)}
wherein: 1,2, N, j 1,2, d; x is the number ofiDenotes the i-th numerical data, yiA tag indicating ith numerical type data;representing j dimension characteristic of i dimension training data; mu.sjMean value, σ, of j-th dimension characteristicjThe variance of the j-th dimension of the feature is represented,denotes xiThe j-th dimension feature of (2) is subjected to standardization processing to obtain a new feature;
normalized initial training setComprises the following steps:
whereinRepresenting the ith digitized data after the normalization processing;
(1c) respectively calculating the normalized initial training set obtained in the step (1b)Euclidean distance between the middle feature and the tag:
(1d) the Euclidean distance between each feature obtained according to the step (1c) and the labelThe feature combinations are obtained according to the sequence from big to small
WhereinIndicating Euclidean distance d between each feature obtained in the step (1c) and the labeljThe smallest one-dimensional feature data.
The classification method of the numerical data radar chart representation based on the convolutional neural network is further improved as follows:
the method for selecting the optimal feature combination by using the data after feature sorting in the step 1.2 comprises the following steps:
(2a) establishing an empty characteristic subset F;
(2b) combining the characteristics obtained in step (1d)Sequentially combining the characteristics of the characteristic data of each dimension in the sequence of the characteristic data of each dimension and adding the characteristics into a characteristic subset F; feature subsets
I.e. combining featuresIn(djThe largest one-dimensional feature data) is placed as a combination of features into the feature subset F, which is then placed into the feature subset FPut into F as a combination of features, and so on until willAll features are put into the feature subset F as a feature combination, resulting in d feature combinations.
Calculating the Euclidean distance between the features and the features in each feature combination in the obtained feature subset F and the labels, and then calculating the average value of the Euclidean distance;
selecting the feature combination with the maximum average value and at least 4 features as the optimal feature combination;
note: the above-described calculation method of the euclidean distance may refer to step (1 c).
When the number of the features in the feature combination with the maximum average value is less than 4, taking the feature combination obtained in the step (1d)The first 4 characteristics form the optimal characteristic combination;
the method for establishing the training set according to the numerical data corresponding to the optimal feature combination in the step 1.2 comprises the following steps:
and (3) according to the optimal feature combination selected in the step (2b), extracting the features corresponding to the optimal feature combination from the initial training set established in the step (1a) to establish a training set.
The classification method of the numerical data radar chart representation based on the convolutional neural network is further improved as follows:
the method for drawing the radar chart corresponding to the numerical data by using the polar function in the step 2.2 comprises the following steps:
and determining the number of the radiuses of the radar map according to the number of the characteristics in the optimal characteristic combination, wherein each radius represents the corresponding characteristic, and the length of each radius is the value of the normalized characteristic.
The classification method of the numerical data radar chart representation based on the convolutional neural network is further improved as follows:
the method for classifying the input convolution neural network model after expressing the numerical data to be classified as the image data by the S5 is as follows:
drawing a radar chart of numerical data to be classified by using a polar function, and inputting the radar chart into a convolutional neural network model for classification;
the method for drawing the radar map of the numerical data to be classified by the polar function comprises the following two ways:
the method comprises the following steps of 1, drawing a corresponding radar map by using a polar function according to the characteristics of numerical data to be classified;
and 2, preprocessing the numerical data to be classified according to the step S1, obtaining the optimal feature combination, and then drawing the radar map according to the features corresponding to the optimal feature combination by using the polar function.
Note: the specific steps of preprocessing the numerical data to be classified and drawing the radar map by using the polar function refer to the processing method for the numerical data.
On one hand, the numerical data are vividly represented by a radar map method, the correlation among the numerical data is reserved, the information loss among the numerical data is enabled to be as small as possible, the advantage of extracting features of a convolutional neural network is fully utilized, and the topological structure features among the radar maps are extracted. On the other hand, compared with the traditional data driving method, the method provided by the invention also provides an image thought of face thinking for other technical personnel, so that massive industrial numerical data can be identified and classified.
Compared with the prior art, the invention has the following technical advantages:
1. the invention effectively represents the numerical data in a graphical mode from the aspect of surface thinking by a radar map representation method, and fully utilizes the relation among all dimensions of the numerical data.
2. The invention extracts the topological structure characteristics and deep level characteristics of the radar map for classification through the convolutional neural network, and compared with the traditional data-driven classification method, the classification accuracy of numerical data is improved.
Drawings
The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.
FIG. 1 is a flow chart of a classification method of radar-graphic representation of numerical data based on a convolutional neural network of the present invention;
fig. 2 is a schematic diagram of the basic structure of the convolutional neural network of the present invention.
Detailed Description
The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto.
The TE process mainly comprises five operation units of a reactor, a condenser, a steam/liquid separator, a circulating compressor and a stripping tower and eight components: a, B, C, D, E, F, G and H.
Embodiment 1, a classification method for radar chart representation of numerical data based on a convolutional neural network, as shown in fig. 1-2, the method includes the steps of firstly performing feature sorting on the numerical data, selecting an optimal feature combination by using the data after the feature sorting, and then representing the optimal feature combination numerical data into image data by using a radar chart, so that a basic structure of the convolutional neural network is constructed, and a convolutional neural network model is obtained through training. And then, the numerical data is effectively represented in a graphical mode from the aspect of surface thinking by a radar chart representation method, and the relationship among all dimensions of the numerical data is fully utilized. And finally, extracting the topological structure characteristics and deep level characteristics of the radar map for classification through the convolutional neural network model, and the classification accuracy is high.
The present embodiment incorporates numerical fault data generated by the TE process. The TE process has 52 variables and 22 working conditions (1 normal working condition and 21 fault working conditions), and d00 and 21 fault condition data sets are obtained by collecting data every 180s in the invention.
d00 is a training data set under normal operating conditions, comprising 500 sets of data (training data).
Each fault condition data set includes a training data set da and a test data set da _ test for the corresponding fault operating condition, where a ∈ {01, 02, …, 21 }. The training data set includes 480 sets of training data; the test data set includes 960 sets of test data, the first 160 sets of test data being normal data. Each set of data (i.e., training data and test data) is 53-dimensional, the first 52-dimensional represents the values (i.e., features) of 52 variables, the last dimension is a label, the label of normal working condition is 0, and the labels of the remaining 21 faults are 1,2, …, 21 in sequence.
The classification method of the radar chart representation of the numerical data based on the convolutional neural network specifically comprises the following steps:
(1) the method specifically comprises the following steps of performing characteristic sorting on numerical fault data generated in the TE process:
(1a) and establishing an initial training set S and an initial test set.
The initial training set S consists of d00 and da. The initial test set consists of da _ test.
Note: and establishing an initial training set and a test set corresponding to the fault working conditions by sequentially adopting a training data set da and a test data set da _ test in the 21 fault condition data sets.
Since each fault operation condition detection method is the same, only fault 1 is specifically described in this embodiment, that is, the initial training set in this embodiment is composed of d00 and d 01. The initial test set consists of d01_ test.
(1b) And carrying out standardization processing on the training data in the initial training set S, wherein the processing formula is as follows:
S＝{(x1，y1)，(x2，y2)，…，(xN，yN)}
wherein: 1,2, N, j 1,2, 980, 52. S represents the initial training set, x, established in step (1a)iDenotes the ith training data, yiA label representing the ith training data.A j-th dimension feature representing the i-th training data, e.g.,denotes xiThe first feature of (1). Mu.sjMean value, σ, of j-th dimension characteristicjThe variance of the j-th dimension of the feature is represented,denotes xiThe j-th dimension feature of (2) is normalized.
The feature of the training data is a 52-dimensional variable affecting the fault condition, the label is 0 under the normal condition, and the embodiment adopts the fault condition data set of the first fault condition, and the corresponding label is 1.
Note: in industry, the characteristics refer to the relevant factors that affect the output.
Normalized initial training setComprises the following steps:
whereinRepresenting the ith digitized data after the normalization processing;
(1c) sequentially calculating the normalized initial training set obtained in the step (1b)Euclidean distance between the middle feature and the tag:
wherein: 1,2, N, j 1,2, 980, 52. djRepresenting j-th dimension feature and its corresponding label yiThe euclidean distance between.Denotes xiThe j-th dimension feature of (2) is subjected to standardization processing to obtain a new feature; y isiThe label representing the ith training data, i.e., the output, may be the actual output value or an artificially given output value.
(1d) Sorting the Euclidean distances between the features and the labels obtained in the step (1c) from large to small, and recording the obtained feature combinations as
WhereinRepresenting the normalized training data obtained in step (1c)In djThe smallest one-dimensional feature data.
(2) Training data sorted by using the features obtained in the step (1d)And selecting the optimal feature combination.
The optimal feature combination is selected according to the Euclidean distance average value between the features and the labels, and the selection method comprises the following steps:
(2a) an empty feature subset F is created.
(2b) The feature combinations reordered according to step (1d)In the order of (1) sequentiallyThe features in (1) are combined and added into a feature subset F;
feature subsetsI.e. combining featuresIn(djThe largest one-dimensional feature data) is placed as a combination of features into the feature subset F, which is then placed into the feature subset FPut into F as a combination of features, and so on until willAll features are placed as one feature combination in the feature subset F.
The euclidean distance between the features and the features in each feature combination in the obtained feature subset F, and between the features and the tags is calculated (refer to the manner in which the euclidean distance is calculated in step (1c) above), and the average value thereof is calculated. The feature combination with the largest average value and at least 4 features is the optimal feature combination.
If the number of features in the feature combination with the largest average is less than 4, then the training data reordered from step (1d)The first 4 features are selected to form the optimal feature combination.
Since the methods of calculating features and features, euclidean distances between features and tags, and calculating their average values are prior art, they are not described in detail herein.
Such as: the combination of characteristics isThe algorithm for the average is as follows:
y refers to the label to which the feature corresponds.
The invention adopts the characteristics and the characteristics, and the Euclidean distance between the characteristics and the labels to combine and sort the characteristics, so that the new characteristic combination can fully reflect the relation with the output.
(3) And representing the optimal feature combination into image data by using the radar map. Using a MATLAB tool, a 32 x 32 size closed radar map image was drawn.
The process of representing numerical data into image data using radar mapping includes:
(3a) and (3) according to the optimal feature combination selected in the step (2b), extracting the features corresponding to the optimal feature combination from the initial training set (namely, the training set which is not subjected to standardization processing) established in the step (1a), thereby establishing a training set, namely, training data in the training set only comprises the features corresponding to the optimal feature combination.
(3b) And (3) carrying out normalization processing on the features selected in the step (3 a):
wherein x isjJ-dimension features representing all training data in the training set established in step (3a),representing the j-th dimension characteristic of the normalized i-th data. max xjDenotes xjMaximum value of (1), min xjDenotes xjMinimum value of (1).
(3c) Normalizing the data after step (3b)And drawing a closed radar map, namely a radar map corresponding to training data one by one, by using a polar function in MATLAB.
In this embodiment, the number of features in the optimal feature combination selected according to the above steps is 7, the number of radii of the radar map is 7, each radius represents a corresponding feature, and the length of the radius is a value obtained by normalizing the feature. These radii are connected to form a closed figure. Each training data corresponds to a radar map.
Note: and (4) correspondingly processing each test data in the initial test set and each actual numerical data to be classified according to the steps to obtain a radar map corresponding to the test data.
(3d) And (4) removing coordinate axes in the radar map obtained in the step (3c), and storing the radar map into a 32 × 32 gray scale picture in a PNG format.
The drawn radar map is stored as a picture training data set train _ x and a picture test data set test _ x, where the picture training data set train _ x includes 980 pictures and the picture test data set test _ x includes 960 pictures.
In the invention, numerical fault data in the TE process is represented by a radar map method, so that information among the data is converted, and the correlation among the data is kept. The conversion between the numerical type data and the image type data is completed.
In actual work, the labels corresponding to the picture training data set train _ x and the picture testing data set test _ x, that is, the output values of the convolutional neural network model, can be determined according to the needs of the user.
In this embodiment, a one-hot code (one-hot code) is used to determine a picture training data set tag train _ y and a picture testing data set tag test _ y. Since the present embodiment includes the normal operating condition and the fault operating condition, the code of the normal operating condition is (1,0), and the code of the fault operating condition is (0, 1). Therefore, train _ y is 980 x 2, the first 500 groups of data are labeled (1,0), and the 501-980 groups of data are labeled (0, 1). test _ y is 960 x 2, the first 160 groups of data are labeled (1,0), and the 161-960 groups of data are labeled (0, 1).
Through one-hot coding, the position corresponding to the maximum value of the predicted output and the position corresponding to the maximum value of the actual output can be conveniently compared, and if the positions are the same, the positions are considered to be the same type.
In conclusion, the invention selects the optimal features according to the features and the labels, draws the radar map with the optimal features, forms a new image training set, and endows the labels with one-hot coding again. In this embodiment, the radar map is composed of a trace _ x and a test _ x, and the trace _ y and the test _ y are composed by one-hot encoding, that is, the data finally input to the convolutional neural network are the trace _ x, the test _ x, the trace _ y and the test _ y.
(4) And constructing a basic structure of the convolutional neural network: the overall structure of the basic convolutional neural network is determined, as well as the initialized parameters.
The basic structure of the convolutional neural network constructed in this embodiment is shown in fig. 2, and includes an input layer, a convolutional layer 1, a pooling layer 1, a convolutional layer 2, a pooling layer 2, a full-link layer, and an output layer.
The convolution kernel size of convolution layers 1 and 2 in the figure can be selected to be 5 x 5 or 3 x 3; the weight W is randomly selected from (0-1), the initial value of the bias b is 0, and sigmoid functions are selected for the activation functions.
The pooling modes of the pooling layers 1 and 2 are both average pooling.
And the output layer adopts an RBF neural network to carry out probability classification.
The cost function of the convolutional neural network constructed in the embodiment selects a root mean square error function. And the error adopts back propagation to adjust the weight W and the bias b, and the adjustment mode is a gradient descent method.
Note: the sigmoid function and the root mean square error function are all the prior art, and the back propagation and gradient descent method is also the prior common technical means, so the specific work content is omitted in the specification.
In this embodiment, the convolutional neural network parameters for each fault condition are selected as follows:
table 1 convolutional neural network parameter selection under different conditions.
Note: the number of feature maps in table 1 refers to (the number of convolution kernels).
(5) And putting the picture training data set train _ x and the picture testing data set test _ x, and the picture training data set label train _ y and the picture testing data set label test _ y (namely, outputting) into a convolutional neural network for training to obtain training classification accuracy and testing classification accuracy and recording the sizes.
The method is a conventional technical means for constructing a convolutional neural network model by training a convolutional neural network by using a training set and a test set, and in the implementation, the convolutional neural network is trained by respectively changing the size of a convolutional kernel, the moving step length, the learning rate, the size of each batch, the iteration times and the like.
Finally, a group of convolutional neural network parameters is selected, so that the TE test classification accuracy is maximized on one hand, and the training classification accuracy and the test classification accuracy are close to each other on the other hand, and therefore the phenomenon of under-fitting or over-fitting is prevented. The classification accuracy of each fault is finally obtained as shown in table 2.
Table 2: training and test classification accuracy for different faults
And sequentially establishing a convolutional neural network model corresponding to each fault working condition according to the steps. And (4) bringing the test data of each fault into a respective model to detect whether the fault occurs.
In the actual use process, if redundant or useless features exist among the features of the numerical data to be classified, feature selection is carried out according to the steps, the optimal feature combination is selected, a radar map is drawn, and classification is carried out by utilizing the established convolutional neural network model. If the characteristics of the numerical data are all useful information and no useful or useless characteristics exist, the characteristics are not required to be selected, the radar chart can be drawn by directly utilizing the original data, and the classification can be carried out by utilizing the established convolutional neural network.
Compared with the prior art, the classification method based on the radar chart representation of the numerical data of the convolutional neural network is adopted, the numerical data are effectively represented into the image data through the radar chart representation method, the information of the numerical data and the correlation among the data are reserved, the topological structure characteristics of the image data are extracted through the convolutional neural network, the similarity and the rotation non-deformation between the image data are well utilized, the classification accuracy is improved, and the fault detection capability and the average detection level are improved.
The invention can convert numerical fault data generated in the TE process into image data, and simultaneously, the information among the data is kept as much as possible. The powerful feature extraction capability of the convolutional neural network is fully utilized, and topological structure features and deeper features among numerical data in the image data are extracted for classification. Compared with the traditional data driving method such as ELM, the average classification precision of the common method can only reach 65%, the method can reach 86.79, and the method fully shows that after numerical data are visually represented by a radar chart, the relationship among the data can be fully mined, and the classification precision is improved.
In summary, the present invention utilizes the radar map method to convert the industrial numerical data into the image data, and further can use the advantage of the convolutional neural network in image data processing. The present invention can compensate for this disadvantage.
Finally, it is also noted that the above-mentioned lists merely illustrate a few specific embodiments of the invention. It is obvious that the invention is not limited to the above embodiments, but that many variations are possible. All modifications which can be derived or suggested by a person skilled in the art from the disclosure of the present invention are to be considered within the scope of the invention.
Claims (7)
1. The classification method of radar map representation of numerical data based on the convolutional neural network is characterized by comprising the following steps of:
s1, collecting and preprocessing numerical data, and establishing a training set:
each numerical data includes its characteristics and label;
s2, representing the numerical data in the data set obtained in the step S1 into image data, and establishing a picture numerical data set train _ x by using the obtained image data; the method comprises the following steps:
2.1, sequentially carrying out normalization processing on the characteristics of the numerical data in the training set;
wherein,a j-th dimension characteristic representing the i-th numerical type data,setting j dimension characteristics of the normalized i-th numerical data; x is the number ofjFeatures of dimension j, max x, representing all numerical data in the training setjDenotes xjMaximum value of (1), min xjDenotes xjMinimum value of (1);
2.2, obtained according to step 2.1Utilizing a polar function to draw a radar map corresponding to numerical data, and utilizing the obtained radar map to establish a picture numerical data set train _ x;
s3, constructing a basic structure of the convolutional neural network;
s4, training by using the picture numerical data set train _ x obtained in the step S2 and the label thereof to obtain a convolutional neural network model;
and S5, representing the numerical data to be classified into image data and inputting the image data into the convolutional neural network model for classification.
2. The method of classifying radar map representations of convolutional neural network-based numerical data according to claim 1, wherein:
the convolutional neural network comprises an input layer, two convolutional layers, two pooling layers, a full-connection layer and an output layer;
the activation function of the convolutional layer selects a sigmoid function, and the output layer adopts an RBF neural network for probability classification;
and selecting a root mean square error function as a final cost function of the convolutional neural network, and adjusting the weight W and the bias b by adopting back propagation of errors.
3. The method of classifying radar map representations of convolutional neural network-based numerical data according to claim 1 or 2, wherein:
the method for collecting and preprocessing numerical data in step S1 includes:
1.1, collecting numerical data, and performing characteristic sorting on the numerical data:
and 1.2, selecting an optimal feature combination by using the data after feature sorting, and establishing a training set according to numerical data corresponding to the optimal feature combination.
4. The method of classifying radar map representations of convolutional neural network-based numerical data according to claim 3,
the step 1.1 is to collect numerical data and perform characteristic sorting on the numerical data as follows:
(1a) collecting numerical data and establishing an initial training set S;
(1b) and (2) carrying out standardization processing on the numerical data in the initial training set S established in the step (1a), wherein the processing formula is as follows:
S＝{(x1，y1)，(x2，y2)，…，(xN，yN)}
wherein: 1,2, N, j 1,2, d; x is the number ofiDenotes the i-th numerical data, yiA tag indicating ith numerical type data;representing j dimension characteristic of i dimension training data; mu.sjMean value, σ, of j-th dimension characteristicjThe variance of the j-th dimension of the feature is represented,denotes xiThe ith dimension characteristic of (1) is subjected to standardization processing to obtain a new characteristic;
normalized initial training setComprises the following steps:
whereinRepresenting the ith digitized data after the normalization processing;
(1c) respectively calculating the normalized initial training set obtained in the step (1b)Euclidean distance between the middle feature and the tag:
(1d) the Euclidean distance d between each feature obtained according to the step (1c) and the labeljThe feature combinations are obtained according to the sequence from big to small
WhereinRepresenting the Euclidean distance d between each feature obtained in the step (1c) and the labeljThe smallest one-dimensional feature data.
5. The method of classifying radar map representations of convolutional neural network-based numerical data according to claim 4, wherein:
the method for selecting the optimal feature combination by using the data after feature sorting in the step 1.2 comprises the following steps:
(2a) establishing an empty characteristic subset F;
(2b) sequentially combining the characteristics of the characteristic data of each dimension obtained in the step (1d) and adding the characteristics into a characteristic subset F; feature subsets
Calculating the Euclidean distance between the features and the features in each feature combination in the obtained feature subset F and the labels, and then calculating the average value of the Euclidean distance;
selecting the feature combination with the maximum average value and at least 4 features as the optimal feature combination;
when the number of the features in the feature combination with the maximum average value is less than 4, taking the feature combination obtained in the step (1d)The first 4 characteristics form the optimal characteristic combination;
the method for establishing the training set according to the numerical data corresponding to the optimal feature combination in the step 1.2 comprises the following steps:
and (3) according to the optimal feature combination selected in the step (2b), extracting the features corresponding to the optimal feature combination from the initial training set established in the step (1a) to establish a training set.
6. The method of classifying radar map representations of convolutional neural network-based numerical data according to claim 5, wherein:
the method for drawing the radar chart corresponding to the numerical data by using the polar function in the step 2.2 comprises the following steps:
and determining the number of the radiuses of the radar map according to the number of the characteristics in the optimal characteristic combination, wherein each radius represents the corresponding characteristic, and the length of each radius is the value of the normalized characteristic.
7. The method of classifying radar map representations of convolutional neural network-based numerical data according to claim 6, wherein:
the method for classifying the input convolution neural network model after expressing the numerical data to be classified as the image data by the S5 is as follows:
drawing a radar chart of numerical data to be classified by using a polar function, and inputting the radar chart into a convolutional neural network model for classification;
the method for drawing the radar map of the numerical data to be classified by the polar function comprises the following two ways:
the method comprises the following steps of 1, drawing a corresponding radar map by using a polar function according to the characteristics of numerical data to be classified;
and 2, preprocessing the numerical data to be classified according to the step S1, obtaining the optimal feature combination, and then drawing the radar map according to the features corresponding to the optimal feature combination by using the polar function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810525151.7A CN108898157B (en) | 2018-05-28 | 2018-05-28 | Classification method for radar chart representation of numerical data based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810525151.7A CN108898157B (en) | 2018-05-28 | 2018-05-28 | Classification method for radar chart representation of numerical data based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108898157A true CN108898157A (en) | 2018-11-27 |
| CN108898157B CN108898157B (en) | 2021-12-24 |
Family
ID=64343388
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810525151.7A Expired - Fee Related CN108898157B (en) | 2018-05-28 | 2018-05-28 | Classification method for radar chart representation of numerical data based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108898157B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113156391A (en) * | 2021-04-25 | 2021-07-23 | 电子科技大学 | Radar signal multi-dimensional feature intelligent sorting method |
| CN113688931A (en) * | 2021-09-01 | 2021-11-23 | 什维新智医疗科技(上海)有限公司 | Ultrasonic image screening method and device based on deep learning |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103971129A (en) * | 2014-05-27 | 2014-08-06 | 浙江大学 | Classification method and device based on learning image content recognition in cross-data field subspace |
| CN105354223A (en) * | 2015-10-05 | 2016-02-24 | 北京工商大学 | Visualization method aiming at discontinuous hierarchical data and application |
| CN105389569A (en) * | 2015-11-17 | 2016-03-09 | 北京工业大学 | Human body posture estimation method |
| CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
| CN107844755A (en) * | 2017-10-23 | 2018-03-27 | 重庆邮电大学 | A kind of combination DAE and CNN EEG feature extraction and sorting technique |
| CN107944412A (en) * | 2017-12-04 | 2018-04-20 | 国网山东省电力公司电力科学研究院 | Transmission line of electricity automatic recognition system and method based on multilayer convolutional neural networks |
-
2018
- 2018-05-28 CN CN201810525151.7A patent/CN108898157B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103971129A (en) * | 2014-05-27 | 2014-08-06 | 浙江大学 | Classification method and device based on learning image content recognition in cross-data field subspace |
| CN105354223A (en) * | 2015-10-05 | 2016-02-24 | 北京工商大学 | Visualization method aiming at discontinuous hierarchical data and application |
| CN105389569A (en) * | 2015-11-17 | 2016-03-09 | 北京工业大学 | Human body posture estimation method |
| CN106874957A (en) * | 2017-02-27 | 2017-06-20 | 苏州大学 | A kind of Fault Diagnosis of Roller Bearings |
| CN107844755A (en) * | 2017-10-23 | 2018-03-27 | 重庆邮电大学 | A kind of combination DAE and CNN EEG feature extraction and sorting technique |
| CN107944412A (en) * | 2017-12-04 | 2018-04-20 | 国网山东省电力公司电力科学研究院 | Transmission line of electricity automatic recognition system and method based on multilayer convolutional neural networks |
Non-Patent Citations (5)
| Title |
|---|
| FRANCISCO, AT EL.: ""Cascade Feature Selection and ELM for automatic fault diagnosis of the Tennessee Eastman process"", 《SCIENCE DIRECT》 * |
| JIHEN AMARA,AT EL.: ""CNN based Chart Image Classification"", 《25TH INTERNATIONAL CONFERENCE IN CENTRAL EUROPE ON COMPUTER GRAPHICS, VISUALIZATION AND COMPUTER VISION 2017》 * |
| 刘文远等: ""基于多维数据雷达图表示的图形分类器研究"", 《计算机工程与应用》 * |
| 王金甲等: ""雷达图图形特征提取中的特征排序"", 《燕山大学学报》 * |
| 陆汝华等: ""基于雷达图重心特征提取的轴承故障诊断方法"", 《轴承》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113156391A (en) * | 2021-04-25 | 2021-07-23 | 电子科技大学 | Radar signal multi-dimensional feature intelligent sorting method |
| CN113688931A (en) * | 2021-09-01 | 2021-11-23 | 什维新智医疗科技(上海)有限公司 | Ultrasonic image screening method and device based on deep learning |
| CN113688931B (en) * | 2021-09-01 | 2024-03-29 | 什维新智医疗科技(上海)有限公司 | Deep learning-based ultrasonic image screening method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108898157B (en) | 2021-12-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111753189A (en) | Common characterization learning method for few-sample cross-modal Hash retrieval | |
| CN112766229B (en) | Human face point cloud image intelligent identification system and method based on attention mechanism | |
| CN111368885A (en) | Aero-engine gas circuit fault diagnosis method based on deep learning and information fusion | |
| CN110827260B (en) | Cloth defect classification method based on LBP characteristics and convolutional neural network | |
| CN112732921B (en) | False user comment detection method and system | |
| CN112964469A (en) | Online fault diagnosis method for rolling bearing under variable load of transfer learning | |
| CN103714148B (en) | SAR image search method based on sparse coding classification | |
| CN112200121A (en) | Hyperspectral unknown target detection method based on EVM and deep learning | |
| CN105930792A (en) | Human action classification method based on video local feature dictionary | |
| CN113806580B (en) | Cross-modal hash retrieval method based on hierarchical semantic structure | |
| CN110111365B (en) | Training method and device based on deep learning and target tracking method and device | |
| CN111739017A (en) | Cell identification method and system of microscopic image under sample unbalance condition | |
| CN111401156A (en) | Image identification method based on Gabor convolution neural network | |
| CN115393631A (en) | Hyperspectral image classification method based on Bayesian layer graph convolution neural network | |
| CN108898157B (en) | Classification method for radar chart representation of numerical data based on convolutional neural network | |
| CN114420151B (en) | Speech emotion recognition method based on parallel tensor decomposition convolutional neural network | |
| CN114187506B (en) | Remote sensing image scene classification method of viewpoint-aware dynamic routing capsule network | |
| CN116152678A (en) | Marine disaster-bearing body identification method based on twin neural network under small sample condition | |
| CN113469084B (en) | Hyperspectral image classification method based on contrast generation countermeasure network | |
| CN109033304A (en) | Multi-modal retrieval method based on online deep layer topic model | |
| CN113935413A (en) | Distribution network wave recording file waveform identification method based on convolutional neural network | |
| Madokoro et al. | Adaptive Category Mapping Networks for all-mode topological feature learning used for mobile robot vision | |
| CN113590867B (en) | Cross-modal information retrieval method based on hierarchical measurement learning | |
| CN117154256A (en) | Electrochemical repair method for lithium battery | |
| CN109697257A (en) | It is a kind of based on the network information retrieval method presorted with feature learning anti-noise |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20181127 Assignee: Zhejiang Zhongjing Bearing Co.,Ltd. Assignor: ZHEJIANG SCI-TECH University Contract record no.: X2022330000080 Denomination of invention: Classification method of radar chart representation of numerical data based on convolutional neural network Granted publication date: 20211224 License type: Common License Record date: 20220506 Application publication date: 20181127 Assignee: CHANGSHAN XINLONG BEARING Co.,Ltd. Assignor: ZHEJIANG SCI-TECH University Contract record no.: X2022330000079 Denomination of invention: Classification method of radar chart representation of numerical data based on convolutional neural network Granted publication date: 20211224 License type: Common License Record date: 20220506 |
|
| EE01 | Entry into force of recordation of patent licensing contract | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20211224 |